{"id": "2407.11933", "title": "Fairness-Aware Multi-Group Target Detection in Online Discussion", "abstract": "Target-group detection is the task of detecting which group(s) a piece of content is ``directed at or about''. Applications include targeted marketing, content recommendation, and group-specific content assessment. Key challenges include: 1) that a single post may target multiple groups; and 2) ensuring consistent detection accuracy across groups for fairness. In this work, we investigate fairness implications of target-group detection in the context of toxicity detection, where the perceived harm of a social media post often depends on which group(s) it targets. Because toxicity is highly contextual, language that appears benign in general can be harmful when targeting specific demographic groups. We show our {\\em fairness-aware multi-group target detection} approach both reduces bias across groups and shows strong predictive performance, surpassing existing fairness-aware baselines. To enable reproducibility and spur future work, we share our code online.", "date": "2024-07-16", "categories": []}
{"id": "2412.20298", "title": "An Experimental Study on Fairness-aware Machine Learning for Credit Scoring Problems", "abstract": "The digitalization of credit scoring has become essential for financial institutions and commercial banks, especially in the era of digital transformation. Machine learning techniques are commonly used to evaluate customers' creditworthiness. However, the predicted outcomes of machine learning models can be biased toward protected attributes, such as race or gender. Numerous fairness-aware machine learning models and fairness measures have been proposed. Nevertheless, their performance in the context of credit scoring has not been thoroughly investigated. In this paper, we present a comprehensive experimental study of fairness-aware machine learning in credit scoring. The study explores key aspects of credit scoring, including financial datasets, predictive models, and fairness measures. We also provide a detailed evaluation of fairness-aware predictive models and fairness measures on widely used financial datasets. The experimental results show that fairness-aware models achieve a better balance between predictive accuracy and fairness compared to traditional classification models.", "date": "2024-12-28", "categories": []}
{"id": "2510.19893", "title": "EQPO: Equitable Group Relative Policy Optimization for Clinical Reasoning", "abstract": "Medical AI systems demonstrated impressive diagnostic performance, yet they routinely show uneven accuracy across demographic groups, disadvantaging underrepresented populations. Although multimodal reasoning foundation models have pushed clinical diagnosis forward, reinforcement learning-based post-training tends to absorb and magnify the biases present in majority-dominated training corpora. We propose Equitable Group Relative Policy Optimization (EQPO), a hierarchical reinforcement learning method that encourages balanced learning across heterogeneous clinical populations by adaptively reweighting samples according to subgroup representation, task difficulty, and data source. As demographic annotations are frequently missing in real-world clinical data, EQPO additionally applies unsupervised clustering to recover latent subpopulations when they are unavailable. On 7 diagnostic benchmarks covering 5 modalities (X-ray, CT, dermoscopy, mammography, ultrasound), EQPO reduces F1 standard deviation by 43.9% and the maximum cross-group F1 gap by 42.7% on QoQ-Med3-8B over vanilla GRPO, and narrows predictive parity gaps by 27.2% on MedGemma-4B over bias-mitigated RL baselines while raising F1 by 12.5% even without any demographic labels. Examining the training trajectory shows that EQPO steadily improves fairness over the course of optimization, in contrast to baseline methods whose fairness degrades as training proceeds, and the discovered implicit groups remain stable and align with masked demographic attributes. We further release EquiMedGemma-4B and EquiQoQ-Med3-8B, equitability-aware clinical VLLMs that attain state-of-the-art accuracy with markedly smaller demographic gaps.", "date": "2025-10-22", "categories": []}
{"id": "2602.04306", "title": "DeFrame: Debiasing Large Language Models Against Framing Effects", "abstract": "As large language models (LLMs) are increasingly deployed in real-world applications, ensuring their fair responses across demographics has become crucial. Despite many efforts, an ongoing challenge is hidden bias: LLMs appear fair under standard evaluations, but can produce biased responses outside those evaluation settings. In this paper, we identify framing -- differences in how semantically equivalent prompts are expressed (e.g., \"A is better than B\" vs. \"B is worse than A\") -- as an underexplored contributor to this gap. We first introduce the concept of \"framing disparity\" to quantify the impact of framing on fairness evaluation. By augmenting fairness evaluation benchmarks with alternative framings, we find that (1) fairness scores vary significantly with framing and (2) existing debiasing methods improve overall (i.e., frame-averaged) fairness, but often fail to reduce framing-induced disparities. To address this, we propose a framing-aware debiasing method that encourages LLMs to be more consistent across framings. Experiments demonstrate that our approach reduces overall bias and improves robustness against framing disparities, enabling LLMs to produce fairer and more consistent responses.", "date": "2026-02-04", "categories": []}
{"id": "2606.18649", "title": "Gender Bias in LLM Hiring Decisions: Evidence from a Japanese Context and Evaluation of Mitigation Strategies", "abstract": "Large language models (LLMs) are increasingly deployed in hiring workflows, yet most research on gender bias in LLM hiring decisions has focused on English-language, Western-format resumes. This study examines whether pro-female gender bias extends to a Japanese corporate context and evaluates two practical mitigation strategies. Using a counterfactual resume design with 60 Japanese rirekisho-format resumes, 12 name pairs selected on linguistically grounded gender-signal criteria, and five state-of-the-art LLMs (Claude Sonnet 4.6, GPT-4o, DeepSeek-V3, Gemini 2.5 Flash, Llama 3.3 70B), we conducted 43,200 API calls across baseline, prompt instruction, and privacy filter conditions. A crossed random-effects linear mixed model confirms a significant pro-female bias across all five models, replicating Western findings in a non-Western context. A prompt-level gender-neutrality instruction produces no meaningful reduction in bias. A name-reliance analysis formally identifies the candidate name as the primary gender channel: removing the name from the prompt reduces the female effect by nearly its full magnitude. An unexpected incompatibility between the privacy filter and GPT-4o's content safety filter, resulting in a 42% refusal rate, highlights a practical deployment challenge for name anonymization in LLM-assisted recruitment pipelines.", "date": "2026-06-17", "categories": []}
{"id": "2606.18679", "title": "Fair Online Resource Allocation", "abstract": "We study the problem of fair online resource allocation, motivated by applications such as refugee resettlement and airline scheduling, where agents arrive sequentially and must be assigned to facilities with limited capacities. We introduce a model that maximizes the overall welfare subject to resource constraints and a Lipschitz fairness requirement, which ensures that similar agents arriving in the same batch receive similar expected outcomes. We first analyze the offline problem, proving that the value of the optimal fair allocation is at least an $\\Omega(1/\\gamma)$ fraction of the optimal unfair allocation, where $\\gamma$ is the fairness coefficient, thereby bounding the price of fairness. For the online setting, we propose an algorithm based on dual mirror descent that enforces fairness constraints within batches while estimating optimal dual variables. We prove that this algorithm achieves sublinear regret relative to the optimal offline fluid benchmark. Finally, we validate our theoretical results using real-world data from the Refugee Economies Programme, demonstrating the algorithm's performance and examining the trade-offs between welfare maximization and fairness enforcement.", "date": "2026-06-17", "categories": []}
{"id": "2606.20241", "title": "BAFIS: Dataset + Framework to assess occupational Bias and Human Preference in modern Text-to-image Models", "abstract": "Generative artificial intelligence has the potential to improve productivity and transform the production of creative content. However, existing research indicates that image generation models are significantly influenced by biases. This work investigates the inherent biases and language-induced biases present in text-to-image models within the context of occupation-related image generation, complementing established metrics with human preference feedback. We present a comprehensive evaluation of five current text-to-image models: Midjourney v6.1, Stable Diffusion 3 Medium, DALL-E 3, Playground v2.5, and FLUX.1-dev , focusing on gender and ethnicity bias, image quality, and prompt alignment. To facilitate this evaluation, we developed the \"Battle-Arena for Fair Image Synthesis\" (BAFIS), a platform designed to collect human feedback on bias in generated images. Furthermore, we created a dataset comprising 21,140 synthetic images generated using multilingual prompts, which serves as a basis for our analysis. We further place our results within a broader social context by comparing them to official statistics from the German Federal Employment Agency. Our findings reveal systematic biases in text-to-image models, with established evaluation metrics in partial correlation with subjective user ratings. Thus, our research emphasizes the need for including human preferences to develop fairer and more inclusive text-to-image models.", "date": "2026-06-18", "categories": []}
{"id": "2606.20461", "title": "Data Bias Mitigation under Coverage Constraints & The Price of Fairness", "abstract": "Machine learning models have been shown to exhibit discriminatory outcomes or degraded performance for individuals at the intersection of multiple sensitive attributes, such as race and gender. This stems in part from two interrelated challenges: the lack of principled measures for quantifying bias (potentially intersectional), and insufficient representation of intersectional subgroups in training data. We extend a recent bias mitigation framework to incorporate coverage constraints that enforce sufficient representation across groups, including intersectional subgroups. Since achieving exactly zero bias for all groups may not be data efficient (meaning it may require large amounts of data), our solution trades small approximation errors in bias for greater data efficiency while satisfying coverage constraints. We also formulate bias mitigation as an integer linear program that optimizes over all mitigation strategies, and characterize the price of fairness, the minimum data modification cost, as a function of fairness tolerance. This is essential both for legal compliance, where regulations may mandate specific fairness thresholds, and for data governance, enabling practitioners to make informed trade-offs between bias reduction and data modification (particularly, data purchasing) costs. We evaluate our techniques on publicly available datasets, demonstrating that bias mitigation via our framework preserves predictive accuracy across multiple classifiers, and that coverage constraints, while motivated by statistical considerations, are essential for preserving downstream ML performance.", "date": "2026-06-18", "categories": []}
{"id": "2606.20527", "title": "StylisticBias: A Few Human Visual Cues Drive Most Social Biases in MLLMs", "abstract": "Multimodal large language models (MLLMs) are increasingly deployed in personally and societally consequential settings, yet the visual cues that shape how these models judge people remain poorly understood. Prior work often compares different (groups of) individuals, making it difficult to separate appearance effects from identity differences. We introduce StylisticBias, a controlled benchmark for evaluating attribute-level social bias in MLLMs. We generate 500 photorealistic base faces and create about 50 single-attribute variations per face, producing about 25K images. This design keeps identity fixed and changes one visual attribute at a time. It lets us measure how specific cues shift model judgments. We evaluate six MLLMs across 25 binary social judgment scenarios. We find that age and body type dominate identity-level effects, while fashion style and other visual cues drive the largest attribute-level shifts. We further find that about 15 attributes account for nearly 80\\% of the total variation, showing that bias is concentrated in a small set of visual cues. Sensitivity is strongest in judgments that are semantically aligned with appearance, especially socioeconomic and style-related judgments. We release StylisticBias as a benchmark for fine-grained bias evaluation in multimodal models. Code and dataset: https://github.com/timo-cavelius/StylisticBias and https://hf.co/datasets/shaghayegh/stylistic-bias-dataset.", "date": "2026-06-18", "categories": []}
{"id": "2606.20557", "title": "Optimal Deterministic Multicalibration and Omniprediction", "abstract": "A model is multicalibrated on a collection of group weights $G$ if it is calibrated -- i.e. unbiased even conditional on its prediction -- not just overall, but also after reweighting contexts by each $g \\in G$. It is a useful property for many downstream applications and is a basic desideratum of trustworthy machine learning. Before this work, all predictors known to attain the minimax-optimal $\\widetilde O(\\varepsilon^{-3})$ sample complexity rate for $\\varepsilon$-multicalibration were randomized, while deterministic predictors were known only with substantially worse sample complexity. Whether randomization is necessary for optimal sample complexity in multicalibration was explicitly asked by [CLNR26] and implicitly in several prior works.\n  We resolve this open problem by giving a minimax-optimal multicalibration algorithm that outputs a deterministic predictor. We then generalize the algorithm to produce optimal deterministic predictors that satisfy outcome indistinguishability (OI) with respect to finite or finitely covered collections of tests. As an application, this also gives deterministic omnipredictors and panpredictors with optimal sample complexity, resolving open problems posed by [OKK25] and [BHHLZ25].", "date": "2026-06-18", "categories": []}
{"id": "2508.03483", "title": "When Cars Have Stereotypes: Auditing Demographic Bias in Objects from Text-to-Image Models", "abstract": "While prior research on text-to-image generation has predominantly focused on biases in human depictions, demographic bias in generated objects remains relatively underexplored. We introduce SODA (Stereotyped Object Diagnostic Audit), a novel framework for systematically measuring these biases through automated attribute discovery and three standardized metrics: Base vs. Demographic Divergence (BDS), Cross-Demographic Disparity (CDS), and Visual Attribute Concentration (VAC). Applying SODA to 8,000 images across five state-of-the-art models and eight object categories (e.g., cars), we find that \"neutral\" prompts produce outputs most visually similar to middle-aged and White people, suggesting these groups are implicitly over-represented in model defaults. Furthermore, demographic cues trigger highly skewed stereotypical outputs: 26.6% of object-model-demographic combinations produce results where all 20 generated images share the exact same attribute value (e.g., rose gold laptops for women). Finally, prompt-level debiasing reduces inter-group disparity but paradoxically collapses within-group diversity, replacing one stereotype with another. SODA offers a practical pipeline for making these implicit associations measurable, serving as a step toward more responsible AI development.", "date": "2025-08-05", "categories": []}
{"id": "2606.18571", "title": "Fair Cognitive Impairment Detection Through Unlearning", "abstract": "Mild Cognitive Impairment (MCI) is a medical condition characterized by a noticeable decline in memory, language, or thinking abilities. MCI detection from spontaneous speech is promising for scalable screening. However, learned models often exploit demographic cues correlated with labels, resulting in a large performance gap across subgroups. We present a multimodal framework that combines (i) cross-model fusion between modalities (speech, text, and image), and (ii) unlearning using gradient reversal that discourages the shared embedding from encoding task-irrelevant demographic attributes. Evaluated on the multilingual benchmarks TAUKADIAL and PREPARE, our method outperforms the state-of-the-art multilingual and multimodal baseline in MCI classification while substantially reducing the performance gap across patient subgroups (sex and language). We further analyze transfer across datasets, showing that demographic unlearning helps learn more robust representations for MCI detection.", "date": "2026-06-17", "categories": []}
{"id": "2501.15351", "title": "Fairness in LLM-Generated Surveys", "abstract": "Large Language Models (LLMs) excel in text generation and understanding, especially in simulating socio-political and economic patterns, serving as an alternative to traditional surveys. However, their global applicability remains questionable due to unexplored biases across socio-demographic and geographic contexts. This study examines how LLMs perform across diverse populations by analyzing public surveys from Chile and the United States, focusing on predictive accuracy and fairness metrics. The results show performance disparities, with LLM consistently outperforming on U.S. datasets. This bias originates from the U.S.-centric training data, remaining evident after accounting for socio-demographic differences. In the U.S., political identity and race significantly influence prediction accuracy, while in Chile, gender, education, and religious affiliation play more pronounced roles. Our study presents a novel framework for measuring socio-demographic biases in LLMs, offering a path toward ensuring fairer and more equitable model performance across diverse socio-cultural contexts.", "date": "2025-01-25", "categories": []}
{"id": "2507.20708", "title": "Exposing the Illusion of Fairness: Auditing Vulnerabilities to Distributional Manipulation Attacks", "abstract": "The rapid deployment of AI systems in high-stakes domains, including those classified as high-risk under the The EU AI Act (Regulation (EU) 2024/1689), has intensified the need for reliable compliance auditing. For binary classifiers, regulatory risk assessment often relies on global fairness metrics such as the Disparate Impact ratio, widely used to evaluate potential discrimination. In typical auditing settings, the auditee provides a subset of its dataset to an auditor, while a supervisory authority may verify whether this subset is representative of the full underlying distribution. In this work, we investigate to what extent a malicious auditee can construct a fairness-compliant yet representative-looking sample from a non-compliant original distribution, thereby creating an illusion of fairness. We formalize this problem as a constrained distributional projection task and introduce mathematically grounded manipulation strategies based on entropic and optimal transport projections. These constructions characterize the minimal distributional shift required to satisfy fairness constraints. To counter such attacks, we formalize representativeness through distributional distance based statistical tests and systematically evaluate their ability to detect manipulated samples. Our analysis highlights the conditions under which fairness manipulation can remain statistically undetected and provides practical guidelines for strengthening supervisory verification. We validate our theoretical findings through experiments on standard tabular datasets for bias detection. Code is publicly available at https://github.com/ValentinLafargue/Inspection.", "date": "2025-07-28", "categories": []}
{"id": "2512.15792", "title": "A Multifaceted Analysis of Social Biases in Large Language Models", "abstract": "Large language models (LLMs) have rapidly become indispensable tools for acquiring information and supporting human decision-making. However, ensuring that these models uphold fairness across varied contexts is critical to their safe and responsible deployment. In this study, we undertake a comprehensive examination of four widely adopted LLMs, probing their underlying biases and inclinations across the dimensions of politics, ideology, alliance, language, and gender. Through a series of carefully designed experiments, we investigate their political neutrality using news summarization, ideological biases through news stance classification, tendencies toward specific geopolitical alliances via United Nations voting patterns, language bias in the context of multilingual story completion, and gender-related affinities as revealed by responses to the World Values Survey. Results indicate that while the LLMs are aligned to be neutral and impartial, they still show biases and affinities of different types.", "date": "2025-12-16", "categories": []}
{"id": "2602.04155", "title": "Maximin Relative Improvement: Fair Learning as a Bargaining Problem", "abstract": "When deploying a single predictor across multiple subpopulations, we propose a fundamentally different approach: interpreting group fairness as a bargaining problem among subpopulations. This game-theoretic perspective reveals that existing robust optimization methods such as minimizing worst-group loss or regret correspond to classical bargaining solutions and embody different fairness principles. We propose relative improvement, the ratio of actual risk reduction to potential reduction from a baseline predictor, which recovers the Kalai-Smorodinsky solution. Unlike absolute-scale methods that may not be comparable when groups have different potential predictability, relative improvement provides axiomatic justification including scale invariance and individual monotonicity. We establish finite-sample convergence guarantees under mild conditions.", "date": "2026-02-04", "categories": []}
{"id": "2604.16450", "title": "Evaluating Intersectional Fairness across Clinical Machine Learning Use Cases using Fairlogue and the All of Us Research Program", "abstract": "Intersectional biases in healthcare data can produce compound disparities in clinical machine learning models, yet most fairness evaluations assess demographic attributes independently. FairLogue, a toolkit for intersectional fairness auditing, was applied across multiple clinical prediction tasks to evaluate disparities across combined demographic groups. Using the All of Us dataset, two published models were selected for replication and evaluation: (A) prediction of selective serotonin reuptake inhibitor associated bleeding events and (B) two-year stroke risk in patients with atrial fibrillation. Observational fairness metrics were computed across race, gender, and intersectional subgroups, followed by counterfactual analysis to evaluate whether disparities were attributable to group membership. Intersectional evaluation revealed larger disparities than single-axis analyses; however, counterfactual diagnostics indicated that most observed disparities were comparable to those expected under randomized group membership. These results highlight the importance of intersectional fairness auditing and demonstrate how FairLogue provides deeper insight into bias in clinical machine learning systems.", "date": "2026-04-07", "categories": []}
{"id": "2606.17684", "title": "Geometrical fairness in graph neural networks", "abstract": "Graph-based learning methods have become increasingly prominent due to their strong performance across diverse applications. Among these, recent frameworks grounded in diffusion processes provide a unifying perspective that extends traditional graph neural network formulations while addressing limitations of standard message-passing mechanisms. Despite these advances, concerns remain regarding the fairness of such models, as they may propagate or amplify biases present in the data. In this work, we introduce a fairness-aware adaptation of graph-based diffusion by modifying the underlying Laplacian operator. Our approach incorporates multiple complementary transformations, including subspace projections, spectral adjustments, and frequency-based filtering, to mitigate bias-related components. Leveraging the intrinsic smoothing properties of graph diffusion, we provide a principled analysis of the resulting behavior and establish theoretical insights into fairness properties. We evaluate the proposed framework on both synthetic and real-world datasets, demonstrating that it achieves competitive performance while improving fairness metrics with limited additional computational cost.", "date": "2026-06-16", "categories": []}
{"id": "2606.17810", "title": "No-Free-Fairness: Fundamental Limits and Trade-offs in Learning Systems", "abstract": "In this paper, we establish a set of theoretical impossibility results, termed the No-Free-Fairness theorems, that identify three fundamental sources of disparity in learning systems. First, we show that when a task exhibits irreducible cost on a subgroup, any decision rule must trade off overall performance with disparity, yielding an inherent fairness--cost frontier. Second, we prove that even in ideal, noise-free settings where a perfectly fair and accurate solution exists, finite-sample learning alone induces nontrivial subgroup disparity, ruling out distribution-free fairness guarantees. More seriously, enforcing strict relative fairness creates a statistical bottleneck: achieving low cost may require exponentially many samples. Third, we show that limitations of the model class can independently induce disparity: if the model cannot represent accurate solutions for a subgroup, fairness remains unattainable regardless of data or training procedure. Overall, these results demonstrate that unfairness is not solely a consequence of biased data or suboptimal optimization, but arises from the intrinsic structure of decision problems, the constraints of finite data, and the expressivity of models. Our framework applies broadly beyond standard supervised learning, and suggests that achieving fairness requires explicit trade-offs and should be treated as a core design consideration.", "date": "2026-06-16", "categories": []}
{"id": "2409.06708", "title": "Ensuring Fairness with Transparent Auditing of Quantitative Bias in AI Systems", "abstract": "With the rapid advancement of AI, there is a growing trend to integrate AI into decision-making processes. However, AI systems may exhibit biases that lead decision-makers to draw unfair conclusions. Notably, the COMPAS system used in the American justice system to evaluate recidivism was found to favor racial majority groups; specifically, it violates a fairness standard called equalized odds. Various measures have been proposed to assess AI fairness. We present a framework for auditing AI fairness, involving third-party auditors and AI system providers, and we have created a tool to facilitate systematic examination of AI systems. The tool is open-sourced and publicly available. Unlike traditional AI systems, we advocate a transparent white-box and statistics-based approach. It can be utilized by third-party auditors, AI developers, or the general public for reference when judging the fairness criterion of AI systems.", "date": "2024-08-24", "categories": []}
{"id": "2501.19337", "title": "Token-Level Entropy Reveals Demographic Disparities in Language Models", "abstract": "We ask whether demographic identity, signaled by a name alone, systematically reshapes the generative distribution of a language model. Measuring full-vocabulary Shannon entropy at temperature zero across six open-weight base models and 5,760 implicit sentence-completion prompts (e.g., \"Tanisha walked into the office on a Monday morning and\"), we find that Black-associated names produce higher first-token entropy than White-associated names across all six architectures - opposite to the output-level homogeneity bias documented under explicit demographic prompting (Lee et al., 2024) - and Black-associated names always produce greater entropy above identity-neutral baselines than White-associated names ($\\Delta\\Delta > 0$ in all six models). Women-associated names co-occur with lower first-token entropy (DL-pooled $\\hat\\beta = -0.041, p = .019$) and more homogeneous outputs ($\\hat\\alpha = +0.024, p < .001$) than men-associated names - a pattern convergent with homogeneity bias; race and gender effects are additive. Instruction tuning does not attenuate the race gap (matched-format DL-pooled $\\hat{\\beta}=+0.153$). Running the same templates with explicit group labels instead of names yields null race effects in 10 of 12 models where implicit probing is significant - establishing that probing methodology is a primary determinant of which distributional structure is recovered.", "date": "2025-01-31", "categories": []}
{"id": "2504.11775", "title": "Discrimination-free Insurance Pricing with Privatized Sensitive Attributes", "abstract": "Fairness has become an important concern in insurance pricing as insurers increasingly rely on machine learning models to predict expected losses. At the same time, regulatory and privacy constraints often restrict insurers' ability to access or use sensitive attributes such as gender or race. Recent actuarial research addresses fairness in this context through the concept of the discrimination-free premium, which removes both the direct and indirect effects of sensitive attributes while preserving actuarial consistency. However, implementing this approach typically requires access to the sensitive attributes themselves, which may not be available in practice. This paper studies the estimation of discrimination-free insurance premiums when sensitive attributes are observed only in privatized or noise-perturbed form. We consider a multi-party data setting in which insurers observe non-sensitive attributes and outcomes, while a trusted third party holds privatized sensitive attributes generated through a privacy mechanism. Within this framework, we develop statistical methods for estimating discrimination-free premiums using only the privatized attributes. We study two settings of practical relevance: when the privacy mechanism is known and when its noise level is unknown. For both cases, we establish theoretical guarantees for the proposed estimators. Numerical experiments and empirical applications demonstrate that the proposed approach enables fair insurance pricing while respecting privacy and regulatory constraints.", "date": "2025-04-16", "categories": []}
{"id": "2605.23234", "title": "Assessing Predictive Models for Fairness Based on Movement Patterns", "abstract": "Assessing the spatial fairness of predictive models involves establishing whether they are statistically penalizing (favoring) individuals associated with certain geographical locations. Literature on this topic makes the fundamental assumption that each individual is assigned to a single geographical location (e.g., place of residence). However, fairness with respect to the set of locations where one has been, i.e., their movement patterns over different regions, also matters when fairness is considered. Consequently, we argue that it is necessary to generalize the notion of spatial fairness to also include movement patterns, leading to the novel problem of assessing predictive models for fairness relative to the movements of individuals. To deal with this problem, we propose an approach that first associates the movements of individuals to certain geographic regions, considering multiple spatial partitions with different resolutions and alignments, and then employs a suitable spatial scan statistic to assess whether a predictive model is fair based on movement patterns. In the experimental evaluation, we study the performance of our approach over thousands of synthetic unfair datasets, showing that it is effective at detecting this new type of unfairness and at retrieving the set of objects treated unfairly, while localization performance exhibits a consistent multi-resolution trade-off.", "date": "2026-05-22", "categories": []}
{"id": "2606.14977", "title": "Identification and Inference for Algorithmic Frontiers with Selective Labels", "abstract": "This paper provides identification results to characterize a fairness-accuracy (FA) frontier, and statistical inference tools to test hypotheses and build a confidence set for the FA-frontier, when outcomes are observed only for selected individuals. When the selection process is unrestricted but loss is measured in specific ways, we provide a characterization of the sharp identification region of the FA-frontier. Under an assumption of unconfoundedness conditional on observables (and unrestricted loss functions), we obtain point identification and propose a debiased machine learning estimator, derive its asymptotic distribution, and show how this can be used to carry out inference for the FA-frontier. In work in progress, we extend the partial identification results to a broader class of loss functions.", "date": "2026-06-12", "categories": []}
{"id": "2606.15914", "title": "Contaminated Collaboration: Measuring Gender Bias Transfer in LLM-Assisted Student Writing", "abstract": "Gender bias in LLMs has been studied extensively in model outputs, with biased prompts shown to amplify stereotyped generations. Whether such bias propagates into text produced by humans who use these systems, however, remains underexplored. We investigate whether gender bias in an LLM writing assistant transfers into career plan essays written by students. We first verify that a gender-biased prompt induces gender-differentiated language in LLM-generated essays, while a neutral prompt does not. We then recruited participants (N = 123) in a controlled environment to write career plan essays for paired biographical profiles differing only in gender under three conditions: no AI assistance, neutral LLM assistance, or gender-biased LLM assistance. Students in the biased condition produced essays with a significantly larger agentic gap and more gender-stereotypic occupation suggestions than those in the control and neutral conditions. Our results also reveal that this bias transfer is asymmetric: agency is suppressed in female-target essays while male-target writing remains largely unaffected. Our findings highlight the risk of bias propagation in AI-assisted writing, calling for fairness-aware design in educational AI tools.", "date": "2026-06-14", "categories": []}
{"id": "2606.16562", "title": "MIRAGE: Auditing Anti-Muslim Bias in Frontier LLMs Across Reasoning, Agentic, and Time-Coupled Conditions", "abstract": "Five years after the discovery of persistent anti-Muslim bias in large language models, most evaluations remain confined to single-turn prompt completion, a setting that no longer reflects how frontier LLMs are deployed. We introduce \\textbf{MIRAGE} (Muslim-Identity Reasoning and Agentic Generation Evaluation), a benchmark of 1{,}200 prompts spanning three deployment-realistic conditions: direct completion, chain-of-thought reasoning, and simulated agentic decision-making across content moderation, lending triage, refugee claim summarization, and hiring screens. Across six frontier models, we find that (i) chain-of-thought reasoning \\emph{amplifies} rather than suppresses Muslim-violence associations by 12--34\\% relative to direct completion, (ii) agentic decisions exhibit a 9--22 percentage-point asymmetry between Muslim and matched non-Muslim cases on identical evidence, and (iii) bias is sharply time-coupled to retrieved news context, increasing 18--27\\% under recent-conflict retrieval. Existing prompt-based mitigations transfer poorly across our three conditions, suppressing direct-completion bias while leaving agentic asymmetry largely intact. We release MIRAGE and an open evaluation harness to support targeted mitigation research.", "date": "2026-06-15", "categories": []}
{"id": "2606.16723", "title": "AgentFairBench: Do LLM Agents Discriminate When They Act?", "abstract": "Large language model (LLM) agents increasingly take actions (screening applicants, recommending credit, triaging patients), yet fairness for LLMs is still measured by grading answers. We introduce AgentFairBench, a cheap, reproducible, multi-domain benchmark for demographic disparity in the actions of LLM agents. Grounded in a companion framework, the Bias Conduction Framework (BCF, restated here), it spans three regulator-anchored domains: hiring, lending, and medical triage. Synthetic, demographic-neutral profiles are evaluated in counterfactual matched sets that vary only a name-coded race x gender signal (in the Bertrand Mullainathan tradition), under four agent scaffolds of increasing agency (direct, chain-of-thought, multi-agent deliberation, tool-augmented). A NumPy-only harness computes counterfactual flip rate, mean absolute score difference (MASD), action-rate disparity, and tool-invocation disparity, with bootstrap confidence intervals, paired tests, and false-discovery-rate control, for single-digit dollars per model. A live leaderboard with a held-out private split and a contamination canary admits external models by submission. Our pilot (864 decisions plus a test-retest replication) carries a methodological lesson: comparing a six-group score spread against a two-run noise difference overstates disparity by ~ 2.4X through statistic arity alone. Against an arity matched noise floor and an omnibus group test, claude haiku 4 5 shows no demographic effect above sampling noise (0 of 120 pairwise and 0 of 9 omnibus contrasts survive correction); a planted-bias test confirms the instrument detects disparity when present. The contribution is a sound, sensitive, adoption-ready instrument, the arity matched null methodology, and open artifacts to scale it. Code, data, and harness are released under open licenses, with an anonymized review artifact.", "date": "2026-06-15", "categories": []}
{"id": "2305.07609", "title": "Is ChatGPT Fair for Recommendation? Evaluating Fairness in Large Language Model Recommendation", "abstract": "The remarkable achievements of Large Language Models (LLMs) have led to the emergence of a novel recommendation paradigm -- Recommendation via LLM (RecLLM). Nevertheless, it is important to note that LLMs may contain social prejudices, and therefore, the fairness of recommendations made by RecLLM requires further investigation. To avoid the potential risks of RecLLM, it is imperative to evaluate the fairness of RecLLM with respect to various sensitive attributes on the user side. Due to the differences between the RecLLM paradigm and the traditional recommendation paradigm, it is problematic to directly use the fairness benchmark of traditional recommendation. To address the dilemma, we propose a novel benchmark called Fairness of Recommendation via LLM (FaiRLLM). This benchmark comprises carefully crafted metrics and a dataset that accounts for eight sensitive attributes1 in two recommendation scenarios: music and movies. By utilizing our FaiRLLM benchmark, we conducted an evaluation of ChatGPT and discovered that it still exhibits unfairness to some sensitive attributes when generating recommendations. Our code and dataset can be found at https://github.com/jizhi-zhang/FaiRLLM.", "date": "2023-05-12", "categories": []}
{"id": "2606.14068", "title": "Harsher on Male? Evaluating LLMs on Gender-Asymmetric Moral Framing Across Diverse Conflict Scenarios", "abstract": "Existing studies on gender bias in LLMs have largely focused on stereotypes, occupational associations, or explicit harmful outputs. In this work, we ask whether LLMs apply consistent response standards to the same negative behavior under matched male-actor and female-actor conditions. We introduce GAMA-Bench, a gender-mirrored benchmark of 1,298 scenarios covering intimate relationship and public social conflicts. It constructs gender-neutral misconduct templates through controlled grids and cross-model review, then compiles them into paired first-person prompts with matched actor-gender and role-reference variations. We further design a structured response-framing protocol to measure how models allocate punishment, empathy, escalation, instruction, and blame. Experiments on 10 representative LLMs reveal a consistent male-disadvantaging asymmetry: male actors receive more punitive, escalatory, and blame-centered framing, whereas female actors receive more therapeutic and empathy-oriented framing for the same misconduct. Further analyses show that this pattern persists across model families, scenario tracks, model scale, and explicit thinking-style reasoning. The official code is available at https://github.com/xufeiqiong/GAMA-Bench.", "date": "2026-06-12", "categories": []}
{"id": "2506.23033", "title": "How Reliable are Fairness Audits with Unreliable Data?", "abstract": "Fairness audits are a key component of responsible machine-learning deployment. Yet, audit-recommendation reliability under incomplete protected-label access is still poorly understood. In this work, we focused on protected-label missingness in fairness mitigation audits. We introduced a seed-calibrated stress test to separate missingness effects from seed-to-seed movement already present under complete labels. Across ACS/Folktables tasks, missingness settings that retain some protected labels usually do not move selected mitigation methods beyond a complete-label seed-to-seed baseline. At $0%$ protected-label access, candidates collapse to an empirical-risk-minimization baseline and deterministic tie-breaking rather than revealing a broad missingness effect. We also found that threshold optimization can turn fairness gains on a single protected axis into intersectional harm above a seed baseline, and this threshold-optimizer finding persists under random-forest validation. Overall, our results highlight that protected-label missingness should be reported with seed-null calibration, candidate-set context, and intersectional consequences before it is treated as evidence of audit fragility.", "date": "2025-06-28", "categories": []}
{"id": "2606.12435", "title": "Auditing Discriminatory Patterns in Mortgage Lending Through Association Rules and Fair Binning", "abstract": "Mortgage lending in the United States exhibits persistent racial and gender disparities. We investigate whether standard data preprocessing steps, specifically attribute binning, amplify these disparities in downstream pattern mining. Using 103,481 cleaned mortgage applications from the HMDA 2023 dataset (Chicago metropolitan area), we build a three-stage pipeline: (1) a PySpark data cleaning and binning pipeline that implements both standard equal-frequency binning and the epsilon-biased fair binning algorithm from Asudeh et al. [1], (2) FP-Growth association rule mining that compares denial patterns under both binning regimes, and (3) K-Means clustering with a per-cluster disparate impact audit. Our standard binning shows 9.63% racial bias in income discretization, consistent with the 8-10% reported in prior work. Fair binning with seven race groups is infeasible at epsilon=0.03 and only succeeds at epsilon=0.08 with a Price of Fairness of 29.4%. FP-Growth reveals that high debt-to-income ratio is the dominant denial predictor (67.2% confidence, 2.81 lift), while racial bias does not appear as explicit high-support rules. However, K-Means clustering followed by a disparate impact audit flags 10 out of 45 cluster-group pairs, showing that Black applicants face significantly higher denial rates than White applicants even among financially similar groups.", "date": "2026-05-16", "categories": []}
{"id": "2606.12615", "title": "Towards Provably Fair Machine Learning: Bayesian Approaches For Consistent and Transparent Predictions", "abstract": "ML classifiers deployed in high-stakes domains produce predictions whose quality varies systematically across subgroups. For granular subgroups defined by intersections of multiple features, predictions are often inconsistent with the observed data: the model's outputs contradict the evidence available for that subgroup. This problem is exacerbated by regularisation, which improves aggregate performance by collapsing small subgroups into larger groups, disproportionately affecting demographic minorities. We define two requirements for consistent prediction: determinism (identical individuals receive identical predictions) and statistical consistency (we cannot reject, at significance level alpha, the hypothesis that the predictions for a subgroup were drawn from the Bayesian optimal target distribution inferred for that subgroup). From these requirements we derive the Fair Bayesian classifier, which enforces both across every group and subgroup simultaneously and abstains whenever no consistent deterministic prediction is possible. On three benchmark datasets (Adult, COMPAS, and Bank Marketing), standard classifiers produce statistically inconsistent predictions for a substantial proportion of subgroups. Our classifier achieves zero consistency error by construction while exceeding baseline accuracy and multicalibration on every dataset tested. Statistical consistency provides a principled foundation for prediction quality with direct implications for algorithmic fairness. Minority demographics are disproportionately concentrated in small subgroups, precisely where frequentist inference is least reliable; addressing this inference problem is therefore a necessary step toward fair ML. By enforcing Bayesian consistency at the finest resolution the data supports, the our classifier demonstrates that exhaustive subgroup fairness with principled abstention is achievable in practice.", "date": "2026-06-10", "categories": []}
{"id": "2606.13105", "title": "Disparate Impact in Synthetic Data Generation", "abstract": "We revisit the fairness notion of disparate impact for synthetic data generation (SDG), that assesses whether the utility of generated records is the same across sensitive groups. Our approach departs from existing work on fair SDG, that address the problem of correcting for undue biases in the observed distribution, hence redefining SDG as learning a distribution that is not that of the real data. By contrast, non-disparate impact is notably achieved when the synthetic and real distributions are the same. We expose reasons why SDG may fail to reach that solution and discuss why approximation and estimation errors occur and can be disparate across groups. We notably look into the expressive power of SDG methods relative to distribution complexity, sampling errors due to group proportions, and estimation errors induced by differential privacy mechanisms. We illustrate cases of disparate impact on both artificial and real-world data, focusing on SDG methods that rely on probabilistic graphical models. We also introduce a strategy of learning group-wise SDG models and illustrate how it can improve both the overall utility and its parity in many settings.", "date": "2026-06-11", "categories": []}
{"id": "2606.11214", "title": "From Awareness to Action: Understanding and Overcoming the Research-Practice Gap in Algorithmic Fairness for Public Health", "abstract": "Algorithmic fairness is essential for responsible ML-driven public health research, yet its practical implementation remains limited. To investigate this awareness-action gap, we conducted a sequential mixed-methods study comprising expert interviews, an online survey, and systematic mapping. The expert interviews informed the design of the survey, which in turn revealed fragmented definitions of fairness, limited training and guidance, reliance on external sources, and rare use of formal assessment, mitigation, or monitoring. These findings were subsequently mapped onto three established research-practice gap lenses: the Knowledge-Practice Gap, the Knowledge-to-Action Cycle, and the Knowing-Doing Gap, each offering complementary perspectives. Building on this synthesis, we introduce the Fairness-to-Action framework, which integrates methodological, organizational, and systemic dimensions to identify where translation of algorithmic fairness knowledge stalls. Our analysis shows that fairness remains weakly institutionalized, translation mechanisms are externally driven, and system-level priorities continue to emphasize accuracy over fairness. These insights suggest critical leverage points for advancing safe, fair, and ethical ML-driven public health research practice.", "date": "2026-05-02", "categories": []}
{"id": "2606.12088", "title": "Debiasing Without Protected Attributes: Latent Concept Erasure from Textual Profiles", "abstract": "Most fairness research in NLP assumes direct access to protected attributes such as gender, race, or nationality. In practice, however, such information is often unavailable due to privacy constraints, missing metadata, or legal restrictions, even though models may infer it from indirect textual cues. This raises a key question: can debiasing succeed without direct access to sensitive attributes? We propose H-SAL, which performs post-hoc concept and attribute erasure using self-description text as an implicit debiasing signal. To support this setting, we introduce a multi-domain Stack Exchange-based fairness benchmark for helpfulness prediction that includes both explicit and implicit signals, enabling comparison between standard debiasing with protected labels and debiasing without access to sensitive information. Across encoder and decoder-only language models, we find that implicit self-description often matches or outperforms explicit-label-based debiasing. Our results broaden representation-level fairness research and provide a new benchmark for studying debiasing under realistic data constraints.", "date": "2026-06-10", "categories": []}
{"id": "2603.04689", "title": "Generalizing Fair Top-$k$ Selection: An Integrative Approach", "abstract": "Fair top-$k$ selection, which ensures appropriate proportional representation of members from minority or historically disadvantaged groups among the top-$k$ selected candidates, has drawn significant attention. We study the problem of finding a fair (linear) scoring function with multiple protected groups while also minimizing the disparity from a reference scoring function. This generalizes the prior setup, which was restricted to the single-group setting without disparity minimization. Previous studies imply that the number of protected groups may have a limited impact on the runtime efficiency. However, driven by the need for experimental exploration, we find that this implication overlooks a critical issue that may affect the fairness of the outcome. Once this issue is properly considered, our hardness analysis shows that the problem may become computationally intractable even for a two-dimensional dataset and small values of $k$. However, our analysis also reveals a gap in the hardness barrier, enabling us to recover the efficiency for the case of small $k$ when the number of protected groups is sufficiently small. Furthermore, beyond measuring disparity as the \"distance\" between the fair and the reference scoring functions, we introduce an alternative disparity measure$\\unicode{x2014}$utility loss$\\unicode{x2014}$that may yield a more stable scoring function under small weight perturbations. Through careful engineering trade-offs that balance implementation complexity, robustness, and performance, our augmented two-pronged solution demonstrates strong empirical performance on real-world datasets, with experimental observations also informing algorithm design and implementation decisions.", "date": "2026-03-05", "categories": []}
{"id": "2606.10126", "title": "Pareto-Guided Teacher Alignment for Fair Personalized Text Generation", "abstract": "Personalized persuasive text generation can improve relevance and engagement, but demographic conditioning may also introduce unequal framing across groups. We study fairness mitigation in personalized generation as a constrained multi-objective alignment problem: reduce demographic disparities while preserving personalization fidelity. We propose a Pareto-guided teacher alignment framework that combines revision-based candidate generation, pair-aware feasibility gating, Pareto-style candidate selection, and optional preference optimization through supervised fine-tuning and direct preference optimization. We evaluate the framework on climate change and vaccination persuasion tasks using a controlled context-rich demographic grid with matched gender and age pairs and a unified five-audit evaluation suite spanning persuasion bias, formality disparity, emotional framing disparity, lexical association disparity, and personalization fidelity. Across both domains and cross-family transfer settings, no single alignment strategy dominates all objectives simultaneously. Instead, methods occupy different regions of a fairness-personalization Pareto frontier: some achieve stronger disparity reductions, while others better preserve personalization or demographic stability. Our results show that fairness mitigation effects are objective-dependent and transfer inconsistently across domains and model families, motivating bounded-regression, multi-audit model selection over single-metric optimization for fairness-sensitive personalized generation.", "date": "2026-06-08", "categories": []}
{"id": "2606.10632", "title": "Is Fairness Truly Fair? Towards Reliable Lipschitz Fairness in Multi-Task Learning via Fixed-\\texorpdfstring{$\\delta$}{delta} Alignment", "abstract": "Lipschitz-style individual fairness formalizes the idea that semantically similar examples should receive similar predictions, but its evaluation in multi-task learning (MTL) can be confounded by method-induced representation scales. This paper identifies threshold confounding: when the auditing tolerance is derived from each model's own representation distances, different algorithms are compared under different semantic thresholds. A threshold-drift analysis further shows how Bias rankings can change and identifies sufficient conditions for ranking preservation.\n  We propose \\textbf{ReLiF}, a reliability-aware framework that separates evaluation-time fixed-$\\delta$ auditing from training-time controlled regularization. ReLiF uses a shared reference tolerance for comparable auditing and a violation-rate feedback controller to keep the Lipschitz surrogate active without letting it dominate stochastic training. This work also develops supporting analysis for threshold drift, reference-tolerance selection, and the relationship between the huberized training surrogate and its unsmoothed positive-margin counterpart.\n  Experiments on clinical time-series benchmarks and NYUv2 (NYU Depth V2) dense prediction show that fixed-$\\delta$ auditing exposes utility--fairness trade-offs that method-dependent thresholds can obscure. On NYUv2 with a ResNet50 backbone, ReLiF achieves competitive utility while substantially reducing aligned bias under shared fixed thresholds. On clinical benchmarks, ReLiF yields controlled fairness-regularized trade-offs, while fixed-$\\delta$ auditing reveals that task-balancing baselines can sometimes achieve lower bias and that genuine utility--fairness trade-offs persist. These results support fixed-$\\delta$ auditing as a semantically consistent protocol for evaluating Lipschitz fairness in MTL.", "date": "2026-06-09", "categories": []}
{"id": "2601.23221", "title": "Optimal Fair Aggregation of Crowdsourced Noisy Labels using Demographic Parity Constraints", "abstract": "As acquiring reliable ground-truth labels is usually costly, or infeasible, crowdsourcing and aggregation of noisy human annotations is the typical resort. Aggregating subjective labels, though, may amplify individual biases, particularly regarding sensitive features, raising fairness concerns. Nonetheless, fairness in crowdsourced aggregation remains largely unexplored, with no existing convergence guarantees and only limited post-processing approaches for enforcing $\\varepsilon$-fairness under demographic parity. We address this gap by analyzing the fairness s of crowdsourced aggregation methods within the $\\varepsilon$-fairness framework, for Majority Vote and Optimal Bayesian aggregation. In the small-crowd regime, we derive an upper bound on the fairness gap of Majority Vote in terms of the fairness gaps of the individual annotators. We further show that the fairness gap of the aggregated consensus converges exponentially fast to that of the ground-truth under interpretable conditions. Since ground-truth itself may still be unfair, we generalize a state-of-the-art multiclass fairness post-processing algorithm from the continuous to the discrete setting, which enforces strict demographic parity constraints to any aggregation rule. Experiments on synthetic and real datasets demonstrate the effectiveness of our approach and corroborate the theoretical insights.", "date": "2026-01-30", "categories": []}
{"id": "2606.00827", "title": "Beyond Independent Manipulation: Individual Fairness-aware Strategic Classification with Peer Imitation", "abstract": "Strategic classification (SC) investigates scenarios where agents manipulate their features to obtain favorable decisions from predictive models. Existing fairness-aware SC approaches primarily focus on group fairness and typically assume that agents respond independently. However, when individual fairness is required, ensuring similar individuals receive similar outcomes, agents' manipulation becomes interdependent: an agent's preferred manipulation depends on the neighborhoods' outcomes. This induces a mismatch between classical SC formulations and fairness-aware decision settings, where independent models no longer accurately characterize strategic manipulations. To address this issue, we introduce individual fairness-aware strategic classification (IFSC), a framework that models peer-driven manipulation arising from individual fairness, where agents imitate nearby positively decided peers to obtain favorable outcomes. IFSC characterizes strategic manipulation as similarity-based imitation toward visible accepted peers and learns classifiers under the resulting post-manipulation distributions. To account for uncertainty in peer observability, IFSC employs a robust learning process that introduces stochastic perturbations during manipulation simulation. Experiments on synthetic and real-world datasets demonstrate that IFSC improves individual-fairness consistency and mitigates imitation-induced distortions.", "date": "2026-05-30", "categories": []}
{"id": "2606.07964", "title": "What Does Debiasing Really Remove? A Geometric Study of PCA-Based Gender Debiasing in Word Embeddings", "abstract": "Debiasing methods based on principal component analysis (PCA) are broadly used to reduce gender bias in word embeddings used in LLMs, yet it remains unclear what aspects of bias they actually remove and how destructive this process is. These methods are based on the understanding that bias resides in a low-dimensional subspace, with the assumption that most of it can be captured by a few principal components. In this work, we conduct a systematic geometric analysis of PCA-based gender debiasing and investigate what is actually removed from the embedding space. Our experiments across multiple embeddings show that direct gender bias is primarily concentrated in the first principal component, supporting the low-rank bias hypothesis. However, associative bias measured by WEAT does not align with these principal directions and is instead spread across multiple embedding dimensions. Furthermore, as expected, we demonstrate that removing an increasing number of principal components leads to a consistent degradation of the embedding geometry, affecting semantic structure and vector relationships. These results reveal that PCA-based debiasing operates as a trade-off: while it effectively reduces certain forms of direct bias, it fails to eliminate distributed associations and introduces geometric distortion. Moreover, there is no universal optimal level of debiasing, as the balance between bias reduction and semantic preservation depends on the chosen metric and embedding. Overall, our findings suggest that bias in word embeddings is not purely low-rank and that simple subspace removal methods may be insufficient for comprehensive debiasing.", "date": "2026-06-06", "categories": []}
{"id": "2606.07969", "title": "Neutrality Bites: Gender Representation in AI-Generated Animal Stories", "abstract": "Gender bias in AI-generated stories is a well-documented problem. While much attention has been paid to reducing or mitigating this bias, it is not always clear whether interventions produce genuinely fairer results. To investigate this issue, we examine how large language models (LLMs) handle gender assignment in a narrative context that is popular, highly ambiguous, and also known to closely reproduce human stereotypes: stories about talking animals. We prompt six leading LLMs to complete an English-language story about seven different anthropomorphic animal characters whose gender is unstated. We additionally iterate with four different narrative settings and a range of model temperatures. Across the 23.8K stories, we find that models frequently avoid gendering the animal character in the story (19% on average) or use gender-neutral language like \"it\" or \"its\" (38.2% on average). However, when gender is assigned, there is a significant masculine bias. Feminine animal characters are virtually absent, present in just 2.2% of stories vs. 40.6% that feature masculine characters. Our findings point to a broader argument: neutrality bites. In other words, models that prioritize neutrality to address social bias may actually contribute to the erasure of marginalized perspectives and identities. We suggest that alternative strategies beyond neutrality need to be pursued, such as ones that more equally distribute social possibilities across imagined subjects.", "date": "2026-06-06", "categories": []}
{"id": "2606.07988", "title": "PAFO: Pareto Fairness Optimization for Personalized Reward Modeling", "abstract": "Large language models (LLMs) increasingly rely on reward models to align their outputs with diverse user preferences. While personalized reward models aim to capture such heterogeneity, they are often trained on imbalanced user preference data and may therefore favor users whose preferences are more common in the training population. In this paper, we identify this failure mode as personalized reward bias, where reward modeling quality varies systematically with preference support rate. We formulate its mitigation as a Pareto fairness problem over group utilities, aiming to improve under-served users without degrading other user groups. To this end, we propose PAFO, a Pareto fairness optimization framework for personalized reward modeling. PAFO first trains group-specialized reward models for majority and minority preference groups, then constructs conditional margin-level supervision to distill their heterogeneous preference boundaries into a single unified model. The resulting model uses group information only during training and requires no explicit group labels at inference time. Experiments on Personal-LLM and DSP show that PAFO improves both minority-group and majority-group accuracy while reducing user-level unfairness across multiple metrics, demonstrating its effectiveness for fairer LLM personalization.", "date": "2026-06-06", "categories": []}
{"id": "2604.27011", "title": "Automatic Causal Fairness Analysis with LLM-Generated Reporting", "abstract": "AutoML, intended as the process of automating the application of machine learning to real-world problems, is a key step for AI popularisation. Most AutoML frameworks are not accounting for the potential lack of fairness in the training data and in the corresponding predictions. We introduce \\textsc{FairMind}, a software prototype aiming to automatise fairness analysis at the dataset level. We achieve that by resorting to the assumptions of the \\emph{standard fairness model}, recently proposed by Ple\\v{c}ko and Bareinboim. This allows for a sound fairness evaluation in terms of causal effects, based on \\emph{counterfactual} queries involving the target, possibly confounders and mediators, and the different values of an input feature we regard as \\emph{protected}. After the necessary data preprocessing, the tool implements a closed-form computation of the effects. LLMs are consequently exploited to generate accurate reports on the fairness levels detected in the training dataset. We achieve that in a zero-shot setup and show by examples the expected advantages with respect to a direct analysis performed by the LLM. To favour applications, extensions to ordinal protected variable and continuous targets and novel decomposition results are also discussed.", "date": "2026-04-29", "categories": []}
{"id": "2606.06514", "title": "Detecting and Mitigating Bias by Treating Fairness as a Symmetry Operation", "abstract": "Machine learning systems deployed in high stakes socioeconomic settings routinely display bias. We formalize bias as a symmetry breaking operation: a classifier is fair if its outputs remain invariant under the counterfactual operation of switching a sensitive attribute, with merit features held fixed. We implement loss based regularization as a symmetry restoring mechanism and evaluate the framework on four synthetic datasets with varying levels of noise, correlation, and bias. The framework achieves upwards of 90\\% violation reduction, with accuracy costs around 5\\%. This framework does not require causal graph knowledge, is computationally lightweight, and generalizes to any sensitive attribute definable as a bit-flip, making it suitable for contexts where local sources of discrimination remain absent from mainstream benchmarks.", "date": "2026-06-02", "categories": []}
{"id": "2606.06694", "title": "The Geography of Algorithmic Judgment: LLM Intermediaries, Place Identity, and Racial Steering in Housing Search", "abstract": "Large language models (LLMs) are rapidly assuming an intermediary role in housing search through the integration of listing platforms within conversational interfaces, mediating access to information, search, and recommendations within urban settings. We expand on prior work on racial steering in LLMs by conducting a behavioral audit of seven open-weight and closed-source LLMs across four U.S. cities, testing location recommendations across three iterative prompting conditions that progressively add lifestyle preference context and reflect fair housing paired-testing methodologies. We find that steering is an emergent behavior of the model's interpretive license rather than primarily a static property. Steering results from the interaction of a user's identity, preference articulation, and the spatial logic that a model has internalized about learned representations of place, preference, and opportunity in a given city, and how different types of users relate to it. While steering was present, it was not uniform in direction or magnitude across evaluated conditions. Preference-conditioned testing often increased or reconfigured the number of models that exhibited steering behaviors relative to baseline conditions, suggesting that LLMs may interpret what the same housing preference means differently depending on the racial identity of the user. Our findings also demonstrate that the city is not a neutral testing unit for LLM evaluation in place-based sectors, and results from one local market cannot be assumed to generalize to another. Local and domain expertise will be required in the housing sector to ensure that legal and institutional commitments to fair housing are not undermined while adopting AI tools that mediate spatial access.", "date": "2026-06-04", "categories": []}
{"id": "2606.06830", "title": "Learning Fair Demand Models", "abstract": "Data-driven pricing is increasingly prevalent in sectors such as airlines, lending, insurance, and retail. By learning demand models from customer features and setting prices accordingly, these systems may generate discriminatory outcomes that raise fairness concerns. This leads to fundamental questions - how and where should systems incorporate fairness considerations in the pricing pipeline, and how does it ultimately affect societal outcomes? To answer these, we study a stylized model where a seller has a two-stage decision pipeline comprising linear demand model estimation followed by price optimization. The seller considers fairness notions in training loss, price, and demand, under both parity-wise and Rawlsian perspectives.\n  We show that equalizing training loss across consumer groups leads to multiple solutions, which in turn can result in undesirable outcomes despite being a standard approach in fair machine learning. Focusing instead on fairness applied directly to prices or demand, we compare two strategies that enforce fairness in either the demand estimation stage or the price optimization stage. For parity-wise fairness, we characterize when each strategy yields higher social welfare under small fairness levels. We show that when market sizes and prices in the dataset are similar, imposing price fairness in the estimation stage is more beneficial to consumers, whereas imposing demand fairness in the optimization stage yields better consumer outcomes. For Rawlsian fairness, the two strategies coincide exactly. Lastly, we extend our model to alternate demand functions and conduct a case study using real-world vaccine pricing data.", "date": "2026-06-05", "categories": []}
{"id": "2602.16149", "title": "Toward Trustworthy Portrait Editing: Evaluation of Demographic Misrepresentation in I2I Models", "abstract": "Instruction-guided image-to-image (I2I) editors are increasingly used in consumer and professional visual workflows, where trustworthiness depends not only on prompt compliance but also on equitable preservation of identity-relevant attributes. We formalize two failure modes: Soft Erasure, where requested edits are weakly realized or silently suppressed, and Stereotype Replacement, where edits introduce unrequested, stereotype-consistent demographic attributes. Using a controlled benchmark of 5,040 edited portraits, we evaluate these failures across three recent open-weight editors with vision-language model scoring and human evaluation. Our results show that identity-preservation failures are pervasive and demographically uneven. In particular, 62--71% of outputs exhibit skin lightening, with Indian and Black source portraits affected at 72--75%, compared with 44% for White source portraits, indicating output-level drift toward lighter or more White-presenting appearances when identity constraints are underspecified. In a mitigation case study, prompt-level appearance constraints reduce race-change scores for non-White source portraits by up to 1.48 points, while leaving White source portraits largely unchanged, without modifying model weights. These findings show that identity preservation is not a uniform property of I2I portrait editing systems, but an unevenly distributed trustworthiness failure with direct social consequences. At deployment scale, such silent distortions can shape AI-mediated self-representation and reinforce representational disparities. We introduce a controlled audit protocol for fairness-aware evaluation and governance of generative editing systems. Project page: https://seochan99.github.io/i2i-demographic-bias", "date": "2026-02-18", "categories": []}
{"id": "2606.02776", "title": "Topics as Proxies for Sociodemographics: How Conversational Context Affects LLM Answers", "abstract": "When large language models (LLMs) are used in high-stakes scenarios, such as legal, medical and financial advice, even a single conversation history is enough to drive differences in outcomes between users. Prior work has demonstrated that this results in outcome disparities between sociodemographic groups, with some groups receiving more advantageous outcomes than others. In this work, we demonstrate that LLMs actually struggle to infer user sociodemographics from a single conversation history and that although there are disparities between sociodemographic groups, they are minimal in magnitude. To investigate what the main driver of these disparities is, we compare user sociodemographics to a range of (psycho)linguistic features of conversations, including conversation topic, emotions, and readability. We find that conversation topics are most predictive of LLM-generated advice within a conversational context, which, to some extent, function as proxies for sociodemographic groups and often affect advice in unpredictable ways. This is cause for concern and highlights the need for future research to better understand and, if needed, mitigate the effect of conversational context on LLM outputs in high-stakes scenarios.", "date": "2026-06-01", "categories": []}
{"id": "2606.04777", "title": "UniFair: A unified fair clustering approach based on separation and compactness", "abstract": "Clustering is increasingly used to support high-impact decisions, yet standard objectives such as k-means can produce clusterings that treat demographic groups unequally. Existing fair clustering methods typically optimize a single notion of fairness and often overlook how clustering costs interact with the geometry of the induced decision boundaries. We propose UniFair, a unified framework that jointly optimizes separation fairness and social fairness. Separation fairness encourages protected groups to lie farther from the induced decision boundaries, while social fairness reduces disparities in within-cluster distortion by penalizing group-wise clustering costs. We develop gradient-based optimization procedures for separation-fair and unified k-means objectives, and extend them to deep clustering by enforcing the same criteria in the latent space of an autoencoder. Experiments on tabular and image datasets show that UniFair reduces both boundary-related and cost-based group disparities with only a modest increase in clustering loss.", "date": "2026-06-03", "categories": []}
{"id": "2606.05187", "title": "Geographic Bias and Diversity in AI Evaluation", "abstract": "Among the many challenges hindering the responsible development and deployment of AI, arguably none has faced more intense scrutiny than bias in its various forms. This underscores the widespread concerns across AI researchers that model outputs, e.g., from generative AI, may encode structural distributional imbalances (stemming from training data or model design) that may amplify social inequality or introduce systemic distortions across application domains ranging from biodiversity to disaster mitigation. Yet, relatively little work has investigated the geographical nature of bias or developed measurable benchmarks for what it means for (generative) AI to be unbiased. In this chapter, we investigate this issue through a literature review. As foundation models are reshaping the landscape of bias research, we examine work spanning both the pre-generative AI and generative AI periods. First, we identify a range of geographic biases. These biases span from representation bias in the training data and regional disparities in the factual recall of language models to the tendency of generative AI to over-proportionally favor prototypical places (called defaults). Then, we showcase how recent studies address the latter bias by evaluating geographic diversity in the outputs of generative AI across various cognitive levels, parameter settings, and output modalities.", "date": "2026-04-28", "categories": []}
{"id": "2606.05937", "title": "Large Language Models are Perplexed by some Political Parties", "abstract": "Large Language Models (LLMs) are increasingly used, including in political applications, but their political fairness has been little studied. We assess it using perplexity, posing that a fair model should give equal probability to all political groups. However, we find, across ten LLMs and three datasets covering 37 languages, that LLMs are more perplexed by the texts of far right and nationalist parties than of social-democratic parties. We find this to be consistent with previous work on translation fairness, to the point that perplexity correlates with downstream translation metrics. Our method is applicable to both base LLMs as well as their instruction-tuned counterpart, and we find that both are highly correlated, suggesting that the political fairness of LLMs stems from their pretraining, and is hardly affected by instruction-tuning.", "date": "2026-06-04", "categories": []}
{"id": "2511.03304", "title": "Extending Fair Null-Space Projections for Continuous Attributes to Kernel Methods", "abstract": "With the on-going integration of machine learning systems into the everyday social life of millions the notion of fairness becomes an ever increasing priority in their development. Fairness notions commonly rely on protected attributes to assess potential biases. Here, the majority of literature focuses on discrete setups regarding both target and protected attributes. The literature on continuous attributes especially in conjunction with regression -- we refer to this as \\emph{continuous fairness} -- is scarce. A common strategy is iterative null-space projection which as of now has only been explored for linear models or embeddings such as obtained by a non-linear encoder. We improve on this by extending this to kernel induced feature spaces by means of the ``empirical feature space''. We theoretically derive this as a direct transformation of the kernel matrix yielding a model and fairness-score agnostic method applicable to continuous protected attributes. We demonstrate that our novel approach in conjunction with Support Vector Regression (SVR) provides competitive or improved performance across multiple datasets in comparison to other contemporary methods.", "date": "2025-11-05", "categories": []}
{"id": "2601.07144", "title": "Optimal Transport under Group Fairness Constraints", "abstract": "Ensuring fairness in matching algorithms is a key challenge in allocating scarce resources and positions. Focusing on Optimal Transport (OT), we introduce a novel notion of group fairness requiring that the probability of matching two individuals from any two given groups in the OT plan satisfies a predefined target. We first propose a modified Sinkhorn algorithm to compute perfectly fair transport plans efficiently. Since exact fairness can significantly degrade matching quality in practice, we then develop two relaxation strategies. The first one involves solving a penalized OT problem, for which we derive novel finite-sample complexity guarantees. Our second strategy leverages bilevel optimization to learn a ground cost that induces a fair OT solution, and we establish a bound on the deviation of fairness when matching unseen data. Finally, we present empirical results illustrating the performance of our approaches and the trade-off between fairness and transport cost.", "date": "2026-01-12", "categories": []}
{"id": "2510.21011", "title": "Generating the Modal Worker: A Cross-Model Audit of Race and Gender in LLM-Generated Personas Across 41 Occupations", "abstract": "As generative AI tools are increasingly used to portray people in professional roles, understanding their racial and gender representational biases is critical. We audit over 1.5 million occupational personas generated by four major large language models (GPT-4, Gemini 2.5, DeepSeek V3.1, and Mistral-medium) across 41 U.S. occupations. Comparing these personas against U.S. Bureau of Labor Statistics (BLS) data, we find that models generate demographics with less variation than real-world data, functionally compressing each occupation toward a dominant demographic profile rather than representing population-level variation. A shift/exaggeration decomposition reveals the structure of these distortions: White (-31 percentage points) and Black (-9 pp) workers are consistently underrepresented, while Hispanic (+17 pp) and Asian (+12 pp) workers are overrepresented, with stereotype exaggeration amplifying existing occupational segregation. These distortions are often extreme, including near-total portrayals of housekeepers as Hispanic and the near-erasure of Black workers from many occupations. Because these patterns recur across models with different institutional and cultural origins, they suggest shared structural sources of bias rather than model-specific artifacts. We argue that auditing generative AI requires evaluation frameworks that examine how synthetic populations systematically reshape demographic visibility across social roles.", "date": "2025-10-23", "categories": []}
{"id": "2510.23469", "title": "Towards Fair Graph Prompting: A Dual-Prompt Mechanism for Mitigating Attribute and Structural Bias", "abstract": "Self-supervised pre-training on unlabeled graph data has become a common paradigm for Graph Neural Networks (GNNs). However, an objective gap often remains between pre-training objectives and downstream tasks. To bridge this gap, graph prompting methods adapt frozen pre-trained GNNs to specific downstream tasks through learnable prompts. Despite its effectiveness, most existing graph prompting methods primarily focus on improving model performance and largely overlook fairness concerns. As downstream graph data inherently contains biases in both node attributes and graph structures, pre-trained GNNs may produce representations that differ across demographic subgroups. To address this limitation, we propose Adaptive Dual Prompting (ADPrompt), a fairness-aware graph prompting framework for adapting pre-trained GNNs. ADPrompt incorporates two complementary components: Adaptive Feature Rectification, which learns personalized attribute prompts to suppress sensitive information at the input level, and Adaptive Message Calibration, which introduces layer-wise structure prompts to dynamically regulate information propagation from neighboring nodes. By jointly optimizing these two modules, ADPrompt adapts the pre-trained GNN while mitigating both attribute-level and structural bias. Experiments on four benchmark datasets with multiple pre-training strategies demonstrate that ADPrompt consistently outperforms seven competitive baselines in node classification tasks.", "date": "2025-10-27", "categories": []}
{"id": "2606.02902", "title": "Fairness Definitions and Metrics in Deep Reinforcement Learning for Drug Discovery in Healthcare: A Rapid Evidence Review", "abstract": "Deep reinforcement learning (DRL) is increasingly applied to de novo molecular design, but choices in data, rewards, and evaluation can yield uneven performance across disease areas and chemotypes. Despite this, there is no concise synthesis of how fairness is defined, measured, and tested in DRL-based drug discovery. In this rapid evidence review, we synthesize fairness definitions and metrics for DRL-driven molecule generation in healthcare. We focus on three questions: (i) how dataset composition and split strategies, especially scaffold versus random splits, affect evaluation and distribution shift; (ii) how reward design (e.g., QED, docking, toxicity, synthetic accessibility) can create or mitigate bias, with emphasis on cancer targets; and (iii) which measurable metrics best capture fairness. This includes parity across cancer versus non-cancer indications and across cancer subtypes. It also includes distributional balance in key physicochemical descriptors, scaffold/chemotype diversity, groupwise validity, toxicity, and synthetic accessibility. From 2017 onward, we searched major biomedical, computer science, and engineering literature databases and used arXiv for horizon scanning. Records were screened using PRISMA-style procedures and analyzed via content coding to link reported parity outcomes to dataset and reward choices. Our review provides a concise set of fairness definitions and metrics for DRL molecule generation. It offers practical guidance for reporting distribution parity and outcome parity. It also summarizes how dataset and reward choices relate to observed parity effects and identifies open gaps relevant to trustworthy, cancer-relevant DRL generation.", "date": "2026-06-01", "categories": []}
{"id": "2404.01356", "title": "Perturbation Effects on Accuracy and Fairness among Similar Individuals", "abstract": "Deep neural networks are vulnerable to adversarial perturbations that can simultaneously degrade prediction robustness and individual fairness across diverse application settings. However, existing evaluation protocols typically assess these dimensions in isolation, thereby obscuring critical failure modes. To bridge this gap, we formalize Robust Individual Fairness (RIF): under semantic-preserving (truth-condition-preserving) perturbations, predictions should remain both correct with respect to the ground truth and invariant across semantically equivalent individuals. To surface RIF violations in practice, we introduce RIFair, a black-box adversarial framework that leverages a decoupled perturbation strategy to construct semantically preserved yet unrobust and/or unfair instance pairs. Experiments across multiple model architectures and real-world textual datasets show that robustness-only or fairness-only metrics often miss Robust Biased and Unrobust Fair behaviors. RIFair}reliably exposes these hidden vulnerabilities, supporting RIF as a necessary criterion for trustworthy model assessment. The experimental code is publicly available at https://github.com/Xuran-LI/RIFair.", "date": "2024-04-01", "categories": []}
{"id": "2508.08337", "title": "Position: Beyond Sensitive Attributes, ML Fairness Should Quantify Structural Injustice via Social Determinants", "abstract": "Algorithmic fairness research has largely framed unfairness as discrimination along sensitive attributes. However, this approach limits visibility into unfairness as structural injustice instantiated through social determinants, which are contextual variables that shape attributes and outcomes without pertaining to specific individuals. This position paper argues that the field should quantify structural injustice via social determinants, beyond sensitive attributes. Drawing on cross-disciplinary insights, we argue that prevailing technical paradigms fail to adequately capture unfairness as structural injustice, because contexts are potentially treated as noise to be normalized rather than signal to be audited. We further demonstrate the practical urgency of this shift through a theoretical model of college admissions, a demographic study using U.S. census data, and a high-stakes domain application regarding breast cancer screening within an integrated U.S. healthcare system. Our results indicate that mitigation strategies centered solely on sensitive attributes can introduce new forms of structural injustice. We contend that auditing structural injustice through social determinants must precede mitigation, and call for new technical developments that move beyond sensitive-attribute-centered notions of fairness as non-discrimination.", "date": "2025-08-10", "categories": []}
{"id": "2510.10943", "title": "The Social Cost of Intelligence: Emergence, Propagation, and Amplification of Stereotypical Bias in Multi-Agent Systems", "abstract": "Bias in large language models (LLMs) remains a persistent challenge, often leading to stereotyping and unfair treatment across social groups. While prior work has mainly focused on individual LLMs, the emergence of multi-agent systems (MAS), where multiple LLMs collaborate and communicate, introduces new and underexplored dynamics in how bias emerges, propagates, and amplifies. To systematically investigate these dynamics, we propose a simple evaluation framework with three agent-level metrics that quantify bias emergence, propagation, and amplification throughout multi-agent interaction. We evaluate MAS across three bias benchmarks under varying LLM backbones, social-group configurations, communication behaviors, and adversarial settings. Our results show that communication can trigger up to 70\\% new bias emergence, propagate bias across over 80\\% of agents, and amplify stereotypes by more than 3$\\times$. We further find that denser and competitive communication generally increases bias. Finally, we demonstrate that MAS are highly vulnerable to simple bias injection attacks, and existing defense strategies provide only limited protection. Our findings provide important insights into the fairness and robustness of multi-agent LLM systems.", "date": "2025-10-13", "categories": []}
{"id": "2511.06676", "title": "How AI Fails: An Interactive Pedagogical Tool for Demonstrating Dialectal Bias in Automated Toxicity Models", "abstract": "Now that AI-driven moderation has become pervasive in everyday life, we often hear claims that \"the AI is biased\". While this is often said jokingly, the light-hearted remark reflects a deeper concern. How can we be certain that an online post flagged as \"inappropriate\" was not simply the victim of a biased algorithm? This paper investigates this problem using a dual approach. First, I conduct a quantitative benchmark of a widely used toxicity model (unitary/toxic-bert) to measure performance disparity between text in African-American English (AAE) and Standard American English (SAE). The benchmark reveals a clear, systematic bias: on average, the model scores AAE text as 1.8 times more toxic and 8.8 times higher for \"identity hate\". Second, I introduce an interactive pedagogical tool that makes these abstract biases tangible. The tool's core mechanic, a user-controlled \"sensitivity threshold,\" demonstrates that the biased score itself is not the only harm; instead, the more-concerning harm is the human-set, seemingly neutral policy that ultimately operationalises discrimination. This work provides both statistical evidence of disparate impact and a public-facing tool designed to foster critical AI literacy.", "date": "2025-11-10", "categories": []}
{"id": "2602.16794", "title": "Beyond Procedure: Substantive Fairness in Conformal Prediction", "abstract": "Conformal prediction (CP) offers distribution-free uncertainty quantification for machine learning models, yet its interplay with fairness in downstream decision-making remains underexplored. Moving beyond CP as a standalone operation (procedural fairness), we analyze the holistic decision-making pipeline to evaluate substantive fairness-the equity of downstream outcomes. Theoretically, we derive an upper bound that decomposes prediction-set size disparity into interpretable components, clarifying how label-clustered CP helps control method-driven contributions to unfairness. To facilitate scalable empirical analysis, we introduce an LLM-in-the-loop evaluator that approximates human assessment of substantive fairness across diverse modalities. Our experiments show that label-clustered CP often provides a favorable balance between utility and substantive fairness, while reducing set-size disparities in line with our theory. Finally, we empirically show that equalized set sizes, rather than coverage, strongly correlate with improved substantive fairness, enabling practitioners to design more fair CP systems. Our code is available at https://github.com/layer6ai-labs/llm-in-the-loop-conformal-fairness.", "date": "2026-02-18", "categories": []}
{"id": "2603.13373", "title": "Ethical Fairness in Ubiquitous Health Sensing without Known Attributes", "abstract": "In ubiquitous and mobile health systems, computational models infer human states from wearable, behavioral, and physiological sensing data. In these settings, high accuracy alone is insufficient; models must act ethically and equitably across diverse people, contexts, and devices. However, fairness methods that rely on demographic or heterogeneous attributes during training are difficult to enforce because such attributes are often unavailable, privacy-sensitive, regulated, or undesirable to collect. Conventional parity-based fairness can also violate ethical principles by trading off subgroup performance. To address this challenge, we present Flare, Fisher-guided LAtent-subgroup learning with do-no-harm REgularization, a demographic- and heterogeneous-attribute-agnostic framework that aligns human-centered fairness with ethical principles for ubiquitous and mobile sensing. Flare leverages optimization geometry, particularly Fisher Information, to regularize curvature and uncover latent disparities in model behavior without demographic or heterogeneous attributes. By integrating representation, loss, and curvature signals, it identifies hidden performance strata and refines them through collaborative but do-no-harm optimization, enhancing subgroup performance while preserving ethical balance. We also introduce BHE (Beneficence-Harm Avoidance-Equity), a metric suite that operationalizes ethical fairness beyond statistical parity. Across mobile physiological, behavioral, and clinical sensing datasets, including EDA, OhioT1DM, IHS, and Percept-R, Flare improves ethical fairness over state-of-the-art baselines. Ablation, interpretability, and loss-landscape analyses show that these gains arise from flatter optimization geometry, simpler decision rules, and do-no-harm latent-subgroup adaptation. Runtime analysis supports the practicality of Flare for resource-constrained sensing deployments.", "date": "2026-03-10", "categories": []}
{"id": "2604.14514", "title": "Perspective on Bias in Biomedical AI: Preventing Downstream Healthcare Disparities", "abstract": "Healthcare disparities persist across socioeconomic boundaries, often attributed to unequal access to screening, diagnostics, and therapeutics. However, this perspective highlights that critical biases can emerge much earlier, during data collection and research prioritization, long before clinical implementation, particularly in studies focused on molecular and omics data. A vast number of studies focus on collecting omics data, but the demographic information associated with these datasets is often not reported, and when it is reported, it reveals substantial biases. An automated analysis of 4514 PubMed-indexed omics publications from 2015 to 2024, examining reporting across multiple demographic dimensions, reveals limited reporting overall; for example, only 2.7% of studies report ancestry or ethnicity information and geographic origin reporting is limited to 2.5%. Analysis of large-scale datasets commonly used for model training, such as CellxGene and GEO, reveals substantial population bias where European-ancestry data dominates. As biomedical foundation models become central to biomedical discovery with a paradigm in which base models are pretrained on large datasets and reusing them repeatedly for many different downstream tasks, they risk perpetuating or amplifying these early-stage biases, leading to cascading inequities that regulatory interventions cannot fully reverse. We propose a community-wide focus on three foundational principles: Provenance, Openness, and Reliability through Evaluation Transparency. Together, these principles can help make biases and limitations more visible to model developers and users, supporting more informed model development, evaluation, and deployment decisions in biomedical AI.", "date": "2026-04-16", "categories": []}
{"id": "2606.00039", "title": "Beyond Categories of Caste: Examining Caste Bias and Morality in Text-to-Image AI Models", "abstract": "Text-to-Image (T2I) models have shown promising utility across various domains. However, such models are also amplifying harmful societal biases in their outputs. In the context of South Asia, recent work has shown caste biases and stereotypes are being perpetuated through Generative AI (GenAI) systems. While this research offers extremely relevant insight into invisibilized narratives of caste discrimination through the GenAI system, they often treat caste as an identity category. Therefore, in this work we shift our ontology to focus on the relational aspect of caste. This enables us to develop a more nuanced understanding of the mechanics of caste discrimination by and through T2I models. Combining an algorithmic audit with critical discourse analysis, we draw on a conceptual frame challenging Brahminical Normativity to show how caste biases are perpetuated beyond the simple binaries of upper vs lower-caste categories. Our contributions are two-fold. Beyond challenging the categorical understanding of caste as a category, we propose an anti-caste approach to tackle the issue of caste bias and fairness in AI systems.", "date": "2026-04-28", "categories": []}
{"id": "2606.00049", "title": "Measuring and Mitigating Bias in Code Generated by Large Language Models", "abstract": "Large language models (LLMs) are widely recognised for their applications in natural language generation and are increasingly used for code generation tasks. However, concerns about bias in their generated outputs remain significant. This paper focuses on GPT-4o and Gemini, mainstream tools for code generation, and proposes a framework for evaluating bias in LLM-generated code, specifically examining the influence of protected attributes, prompts and web-search capability. We use two metrics: the code bias score (CBS) and the attribute change ratio (ACR), to quantify the prevalence of bias and the degree of influence of different attributes, respectively. In addition, we investigate four lightweight mitigation strategies: Few-Shot, Chain-of-Thought, Few-Shot Chain-of-Thought, and Multi-agent, aimed at mitigating bias in generated code. Our findings reveal that bias remains prevalent across different protected attributes and datasets even after applying mitigation strategies, highlighting the need for more effective approaches to reduce bias in AI-driven code generation systems.", "date": "2026-05-05", "categories": []}
{"id": "2606.00656", "title": "Demystifying the Optimal Fair Classifier in Multi-Class Classification", "abstract": "Ensuring fair and equitable treatment across diverse groups, particularly in multi-class classification tasks, poses a significant challenge due to the persistent biases inherent in machine learning models. Most existing bias mitigation techniques are tailored to binary settings, and the presence of multi-dimensional outputs and complex fairness mechanisms makes their extension to multi-class scenarios neither straightforward nor effective. In this paper, we investigate two fundamental, unresolved challenges in fair classification: (i) characterizing the optimal accuracy-fairness frontier in multi-class settings, and (ii) designing practical algorithms that attain this optimum in different training phases. To tackle these challenges, we first specify an analytically tractable probabilistic formulation of the optimal classifier under fairness constraints. Building upon this, we propose two attribute-blind algorithms to enforce fairness requirements in practice: an in-processing approach for fairness intervention during training via the reduction approach, and a post-processing approach for fine-tuning output probabilities with plug-in estimation. Theoretical analysis reveals that both methods converge to the optimal accuracy-fairness Pareto frontier. Experiments conducted on multiple datasets demonstrate the superior performance of our methods in balancing accuracy and fairness.", "date": "2026-05-30", "categories": []}
{"id": "2606.00700", "title": "COPF: An Online Framework for Deployment-Stable Counterfactual Fairness in Evolving Graphs", "abstract": "Online link recommendation on evolving graphs is performative: by choosing which candidate links to show users, the system changes which links form and what feedback it later observes. Consequently, fairness estimates from logged outcomes can be misleading and may drift after deployment when the recommendation policy is updated. We introduce COPF (Counterfactual Online Performative Fairness), a decision-layer framework for deployment-stable fairness monitoring and control in online link recommendation. COPF (i) defines group-level opportunity gaps over exposure (shown vs. not shown) counterfactuals, (ii) makes them estimable by explicit exploration and by logging the probability (propensity) that each candidate is shown, and (iii) audits and controls fairness using residual outcome indistinguishability (OI) over a configurable auditor family with graph-aware doubly robust (GA-DR) estimators. We provide a noisy transfer theorem showing that Residual-OI on estimated GA-DR residuals implies bounds on exposure-counterfactual group gaps under temporal mixing and bounded local interference, and we instantiate an online multicalibration auditor together with a primal-dual controller. Experiments on two TGB streams and a controlled synthetic bipartite stream show that COPF reduces worst-case spikes in exposure-counterfactual group disparities with modest impact on ranking utility. Our code is available at https://github.com/lsnnnnnnnn/COPF.", "date": "2026-05-30", "categories": []}
{"id": "2606.01260", "title": "IndoBias: A Dual Track Culturally Grounded Benchmark for LLMs Bias Evaluation in Indonesian Languages", "abstract": "Despite being home to more than 1300 ethnic groups and 700 indigenous languages, bias in Large Language Models has not been fully studied in Indonesia, thus leaving a critical gap in evaluating representational fairness and localized stereotypes within its uniquely vast, multilingual, and diverse sociocultural landscape. To address this, we introduce IndoBias as a culturally-grounded bias benchmark to assess LLMs bias in Indonesian and three local languages: Javanese, Sundanese, and Makasar. IndoBias features dual perspective evaluation tracks: depth-oriented (with contrastive-pairs) and breadth-oriented (with generation-based), where the latter is grounded in social science frameworks (SPI, O*NET, and WGI). Our results show that existing LLMs -- particularly decoder models -- exhibit strong bias towards prototypical sentences in Indonesian, while local languages suffer higher bias under Ideology and Religion category. We also find that LLMs responses exhibit a non-uniform Stereotype Polarity when prompted with various local entities. Finally, we discover that, in Indonesian, Common Crawl texts introduce more bias during pretraining, compared to human-reviewed article texts (e.g., Wikipedia, News), whereas introducing local languages to pretraining generally increases bias. This work highlights the importance of studying bias in culture-specific context. Warning: This paper contains example data that may be offensive, harmful, or biased.", "date": "2026-05-31", "categories": []}
{"id": "2606.01282", "title": "KG-FairDiff: Knowledge Graph-Guided Prompt Refinement for Demographically Fair Text-to-Image Generation", "abstract": "Text-to-Image (TTI) systems are now everyday infrastructure for journalism, education, advertising, and public communication, and the demographic and cultural stereotypes they inherit from training data (rendering women, people of colour, older adults, and non-Western cultures as under-represented or caricatured) become a population-level harm at deployment scale. Existing mitigations either require costly retraining, infeasible for the closed-source backbones that dominate consumer products, or rely on fixed demographic templates that ignore cultural context. We present KG-FairDiff, a model-agnostic, inference-time framework that formalises fairness-aware prompt refinement as a constrained optimisation problem and operationalises it as a closed-loop pipeline: a knowledge graph of ~1,200 culture- and bias-related triples retrieves structured context, an LLM rewriter proposes refinements, and a validator accepts only prompts that reduce a divergence-based fairness loss while preserving semantic fidelity to the user's original intent. We prove a finite-termination bound for the refinement loop, contribute a mathematically consistent evaluation suite linking Bias-P/Bias-W to divergence from target distributions and ENS to KL divergence, and audit eight widely-deployed backbone generators. KG-FairDiff substantially reduces gender, race, age, and intersectional disparities while preserving prompt semantics, offering a practical, deployment-ready route to more equitable generative AI.", "date": "2026-05-31", "categories": []}
{"id": "2606.02198", "title": "Model Multiplicity and Predictive Arbitrariness in Recidivism Risk Assessment", "abstract": "Prediction tasks over individual futures, which are inherently noisy, often admit multiple similarly accurate models. When these models produce different predictions for the same individual, they raise concerns of arbitrariness in decision-making. How severe can this arbitrariness be, in theory and in practice? How can it be resolved to support high-stakes risk assessment?\n  We address these questions through a study of a machine learning-based decision support system for recidivism risk assessment that has been in use for over 15 years. By translating complex legal rules into an algorithm for labeling post release outcomes (recidivist or non-recidivist), we first construct a dataset of thousands of inmate releases. Using this dataset, we learn interpretable models that improve predictive performance, reduce error-rate disparities between groups, and ensure that rehabilitative progress lowers risk scores. Next, we study predictive multiplicity, by first deriving a tight lower bound on the expected predictive agreement of any finite set of models over a dataset, and then by evaluating the extent to which structural diversity (e.g., different model coefficients) within this set translates to predictive multiplicity (i.e., different predictions for the same individual). Our experiments indicate that the existence of many similarly accurate models with comparable error-rate disparities does not necessarily translate into severe predictive multiplicity. Empirically, similarly performant models can exhibit substantially higher predictive agreement than worst-case theoretical guarantees suggest. We find that a simple policy that assigns each inmate the lowest risk among these models is effective for addressing predictive arbitrariness.", "date": "2026-06-01", "categories": []}
{"id": "2606.02214", "title": "Do Gender Cues Affect LLM Value Trade-offs? Evidence from a Controlled Decision Benchmark", "abstract": "Large language models are increasingly used in value-sensitive decision settings, where irrelevant demographic cues should not alter judgments. We construct the Realistic Value Decision Benchmark (RVDB), a controlled benchmark that varies only the role-gender configuration while holding the scenario, ordered value pair, roles, candidate decisions, Value Distance, and Decision Severity fixed. Using a position-balanced evaluation across seven models, we test whether models preserve decision invariance under gender perturbations and whether their self-attributions reflect observed behavioral changes. We find that explicit gender cues induce bounded but systematic decision flips, including under an explicit gender-attribution prompt that asks models to report whether gender influenced their choice. Cross-gender role swaps reveal a consistent female-proposed-decision asymmetry, while models often attribute flipped decisions to No Influence or other non-gender factors. Further analysis shows that gender effects concentrate near less determinate value boundaries and under more severe decision contexts, suggesting that gender cues act as local boundary-shifting factors rather than global overrides of value reasoning. Value rankings remain largely stable, but ordered value-pair trade-offs shift unevenly across role-gender configurations. These results show that gender can enter LLM value trade-offs behaviorally while remaining obscured in self-attribution, motivating controlled behavioral audits beyond explanation-based evaluation.", "date": "2026-06-01", "categories": []}
{"id": "2501.12020", "title": "On the Illusion of Gender Bias in Face Recognition: Explaining the Fairness Issue Through Non-demographic Attributes", "abstract": "Face recognition systems (FRS) exhibit significant accuracy differences based on the user's gender. Since such a gender gap reduces the trustworthiness of FRS, more recent efforts have tried to find the causes. However, these studies make use of manually selected, correlated, and small-sized sets of facial features to support their claims. In this work, we analyze gender bias in face recognition by successfully extending the search domain to decorrelated combinations of 40 non-demographic facial characteristics. First, we introduce a toolchain to effectively decorrelate and aggregate facial attributes to enable a less-biased gender analysis on large-scale data. Second, we tailor two specialized metrics to quantify the effect of facial attributes on absolute and relative fairness. Based on these grounds, we thirdly present a novel unsupervised joint investigation framework capable of identifying attribute combinations leading to vanishing bias when used as filter predicates for balanced testing datasets. Experiments show the gender gap vanishing when images of male and female subjects share specific attributes, clearly indicating that the disparate performance is not a question of biology but of the social definition of appearance. These findings could reshape our understanding of fairness in face biometrics and provide insights into FRS, helping to address gender bias issues.", "date": "2025-01-21", "categories": []}
{"id": "2602.07285", "title": "Fair Decisions from Calibrated Scores: Achieving Optimal Classification While Satisfying Sufficiency", "abstract": "Binary classification based on predicted probabilities (scores) is a fundamental task in supervised machine learning. While thresholding scores is Bayes-optimal in the unconstrained setting, using a single threshold generally violates statistical group fairness constraints. Under independence (statistical parity) and separation (equalized odds), such thresholding suffices when the scores already satisfy the corresponding criterion. However, this does not extend to sufficiency: even perfectly group-calibrated scores -- including true class probabilities -- violate predictive parity after thresholding. In this work, we present an exact solution for optimal binary (randomized) classification under sufficiency, assuming finite sets of group-calibrated scores. We provide a geometric characterization of the feasible pairs of positive predictive value (PPV) and false omission rate (FOR) achievable by such classifiers, and use it to derive a simple post-processing algorithm that attains the optimal classifier using only group-calibrated scores and group membership. Finally, since sufficiency and separation are generally incompatible, we identify the classifier that minimizes deviation from separation subject to sufficiency, and show that it can also be obtained by our algorithm, often achieving performance comparable to the optimum.", "date": "2026-02-07", "categories": []}
{"id": "2602.11802", "title": "Structural Bias Beyond Homophily: A Study of Fairness in Link Prediction", "abstract": "Graph link prediction (LP) plays a critical role in socially impactful applications such as job recommendation and friendship formation, making fairness a critical concern in this task. While many fairness-aware methods manipulate graph structures to mitigate prediction disparities, the topological biases inherent to social graphs remain poorly understood and are consistently conflated with homophily alone. In this work, we study the relationship between structural biases and fairness outcomes in LP. To this end, we formalize a taxonomy of topological bias measures and introduce a graph generation method producing a diverse corpus of synthetic graphs with controlled structural properties. Using this corpus, we show empirically that fairness outcomes are strongly correlated with graph topology, and that current fairness-aware methods remain sensitive to structural biases beyond homophily. These findings highlight the need for structurally grounded evaluations in fair graph learning.", "date": "2026-02-12", "categories": []}
{"id": "2605.30641", "title": "COFT: Counterfactual-Conformal Decoding for Fair Chain-of-Thought Reasoning in Large Language Models", "abstract": "Large language models (LLMs) can reveal and amplify societal biases during chain-of-thought (CoT) generation. We present COFT (Chain of Fair Thought), a training-free decoding method that applies token-level fairness control at decode time, with distribution-free marginal validity guarantees (under exchangeability) for any frozen causal language model. COFT operates in three stages. First, it creates a masked counterfactual prompt by replacing sensitive spans with neutral tokens. Second, it compares the factual and masked logit distributions through lightweight logit fusion to attenuate attribute-driven biases. Third, it uses dual-branch split-conformal calibration to certify per-step candidate token sets at a user-chosen risk level. We evaluate COFT across six models and multiple bias benchmarks. Our method reduces standard bias metrics by 30-55% (median 38%) while preserving task utility and language quality. Reasoning accuracies remain unchanged within run-to-run noise margins. The computational overhead is modest, equivalent to one additional cached forward pass (<=11%). COFT offers a clear, auditable path to safer CoT generation with significant bias reduction, negligible utility loss, and no requirement for retraining, auxiliary classifiers, or weight access.", "date": "2026-05-28", "categories": []}
{"id": "2508.03253", "title": "Approximate Proportionality in Online Fair Division", "abstract": "We study the online fair division problem, where indivisible goods arrive sequentially and must be allocated immediately and irrevocably. Prior work establishes strong impossibility results for approximating classic notions such as envy-freeness up to one good (EF1) and maximin share (MMS) in this setting, but the approximability of proportionality up to one good (PROP1) has remained unresolved. We resolve this gap in two steps. First, we show that three natural greedy allocation rules (standard baselines in fair division) fail to guarantee any multiplicative approximation to PROP1 against an adaptive adversary. These limitations motivate two relaxations: (i) restricting attention to a non-adaptive adversary, and (ii) incorporating coarse predictions in the spirit of learning-augmented algorithms. Under a non-adaptive adversary, we show that the uniform random allocation achieves a meaningful PROP1 approximation with high probability, and this guarantee is essentially tight for this approach; moreover, when item values are sufficiently small, the allocation is near-PROP1 with high probability. Finally, given maximum item value (MIV) predictions, we design an online algorithm that achieves robust approximation guarantees for PROP1, and degrades gracefully under one-sided prediction error. In contrast, we show that EF1, MMS, and PROPX remain inapproximable even with perfect MIV predictions.", "date": "2025-08-05", "categories": []}
{"id": "2605.29827", "title": "Fairness Beyond Demographics: Optimizing Performance Across Appearance-Based Hidden Cohorts in Medical Imaging", "abstract": "Medical image analysis models can exhibit performance disparities across patient subgroups, threatening clinical safety and fairness. Existing methods typically address this issue by optimizing accuracy and fairness metrics for visible demographic attributes (e.g., sex or age) considered in isolation. This strategy not only overlooks potentially more informative latent stratifications, which may reveal deeper sources of model error and inequity, but also fails to scale when multiple demographic attributes are considered simultaneously due to the resulting sparsity of training data within each subgroup. We deal with these issues by introducing the label-free hidden-cohort fairness (LHCF) training paradigm that instead of maximizing fairness over visible demographic attributes, it optimizes fairness across latent subpopulations discovered from image appearance. By clustering images into K appearance-based cohorts and applying fairness optimization over them, LHCF uncovers underlying sources of model error and avoids the combinatorial sparsity of multi-demographic attributes, reducing disparities across both single and multiple demographic attributes. We demonstrate on our proposed fairness benchmark, HIDFairBench, that LHCF provides state-of-the-art fairness results on single and multiple demographic attributes, despite never using demographic labels for training. Our results position hidden-cohort fairness as a practical, scalable, and robust alternative to demographic-based fairness optimization for trustworthy medical image analysis.", "date": "2026-05-28", "categories": []}
{"id": "2605.27407", "title": "Benchmarking Fairness in Spiking Neural Networks: Data Bias, Spurious Features, and Hardware Effects", "abstract": "Evaluating fairness in Spiking Neural Networks (SNNs) demands rigorous benchmarks that reflect real-world complexities, yet existing assessments remain limited by superficial dataset diversity and idealized hardware assumptions. This work introduces the first systematic fairness benchmark for SNNs, addressing three critical dimensions of realism: (1) demographic coverage gaps in training data, (2) spurious feature leakage (e.g., skin tone as a proxy for class labels), and (3) deployment-environment mismatches (e.g., edge devices with constrained spike encoding). Our framework integrates four cross-demographic datasets with controlled bias injections and three neuromorphic hardware simulators (Loihi 2, SpiNNaker), enabling isolated analysis of fairness-performance trade-offs under resource constraints. Standardized evaluations of 12 state-of-the-art SNNs reveal stark disparities: models trained on biased data exhibit 23\\% higher false positive rates for underrepresented groups, while hardware limitations (e.g., reduced spike precision) further amplify accuracy gaps by up to 41\\% in edge deployments. Critically, bias mitigation strategies developed for cloud-based SNNs often degrade under resource constraints, highlighting the need for co-design principles that jointly optimize fairness and hardware efficiency. By bridging algorithmic fairness research with neuromorphic engineering, our benchmark provides a foundation for trustworthy SNNs in socially critical applications such as healthcare and autonomous systems. Our code is available at: https://anonymous.4open.science/r/SNN-Benchmarks-8017.", "date": "2026-05-08", "categories": []}
{"id": "2605.27689", "title": "Test-Time Collective Action: Proxy-Based Perturbations for Correcting Algorithmic Harms", "abstract": "When machine learning systems under-perform for particular subgroups, affected users typically have no way to correct these disparities without relying on platform-level fixes. Existing approaches to algorithmic fairness rely on provider-centric approaches to correct these failures, leaving users with no external lever when faced with harm. Recent work in Algorithmic Collective Action shows that coordinated users can steer an algorithmic system toward a collective goal, but the existing mechanisms require the provider to retrain on the collective's modified data which users may not have control over. We propose Test-Time Collective Action (TTCA), a framework through which a group of users who share query access to the platform, can correct disparities affecting under-served subgroup without participating in the platform's training loop. We implement this through a proxy-based mechanism where the collective pools query access to a black-box API to extract a proxy of the platform, then optimizes a per-class universal perturbation against the proxy. Each member applies this perturbation to their own inputs at submission time, requiring no cooperation from the platform. We empirically evaluate the mechanism on CIFAR-10, CIFAR-100, and FairFace, showing that modestly-sized collectives close most of the subgroup accuracy gap, transfer across architectures (a small proxy can attack a larger platform), and improve worst-group accuracy, equal-opportunity gap, and disparate impact. A query-budget analysis comparing a per-user black-box attack baseline shows that pooling is cheaper than each subgroup member attacking alone. Test-time collective action thus offers corrective intervention to users when platform-side remediation is unavailable or delayed.", "date": "2026-05-26", "categories": []}
{"id": "2605.28036", "title": "Stay Fair! Ensuring Group Fairness in Diffusion Models Across Guidance Scales", "abstract": "Diffusion models steer conditional generation with a tunable guidance scale to trade off prompt alignment and diversity. However, existing debiasing techniques are optimized for a single scale, degrading fairness when users adjust this parameter. We trace this behavior to a previously overlooked source by decomposing total bias into two components: a model bias and a guidance bias. While prior work primarily targets the former, we show that the guidance bias grows monotonically with the guidance scale, eventually dominating the high-guidance regimes users prefer. To address this, we extend Strong Demographic Parity to guidance and derive a condition under which the target distribution retains its group ratio across guidance scales. We propose StayFair, which leverages this condition to design fair guidance algorithms in both regimes. For classifier guidance, it equalizes the classifier's output distributions across groups; for classifier-free guidance, it shifts the null embedding by a prompt-dependent offset. Because StayFair modifies only the guidance step, it is orthogonal to model debiasing and can be layered onto existing fair diffusion models to extend their fairness across guidance scales. Across class-conditional and text-to-image generation, StayFair decouples fairness from the guidance scale without sacrificing image quality.", "date": "2026-05-27", "categories": []}
{"id": "2605.28098", "title": "Examining Agents' Bias Amplification versus Suppression in Multi-Agent Systems", "abstract": "Multi-agent systems are increasingly deployed to support various tasks where agents interact to achieve individual and collective objectives. Although these systems can enhance task performance and decision-making, fairness preservation through bias reduction remains challenging. This study examines how agent-level biases shift and impact system-wide fairness. We use prompts to expose individual agents to group-favoring bias, then assess downstream impacts at the system level. To quantify the impact, we propose Favor Bias Strength (FBS), a zero-centered metric that decomposes bias alteration between favored-group uplift and disfavored-group suppression. Using multiple agent designs, benchmarks, and up-to-date large language models, we show that agents endowed with bias can substantially affect system-wide fairness. Interestingly, when agents are exposed to bias uniformly, the system-wide bias elevates, even exceeding the additive sum of the individual agents' biases. The empirical evidence underscores the criticality of fairness in multi-agent systems, which warrants further analyses and empirical tests.", "date": "2026-05-27", "categories": []}
{"id": "2605.28233", "title": "Geometry of Relaxed Fair Regression: A Unified Framework for Aware and Unaware Settings", "abstract": "Fairness-accuracy trade-offs are a central concern in the deployment of fairness-aware machine learning methods. When sensitive attributes are unavailable at inference time-the so called unawareness setting, principled methods for obtaining accurate predictions under relaxed fairness constraints are largely missing. In this work, we address this gap by formulating regression under a demographic parity penalty as an optimal transport problem. Our framework unifies both the \\emph{aware} and \\emph{unaware} settings and characterizes optimal prediction functions via optimal transport maps, under both squared Wasserstein-2 and Total Variation penalties. These results reveal that the choice of penalty reflects fundamentally different fairness philosophies: the Wasserstein penalty induces a smooth, population-wide compromise, while Total Variation enforces exact parity for a subset of individuals. Building on these theoretical characterizations, we propose an algorithm that is simple to implement, computationally efficient, and consistently matches or outperforms state-of-the-art baselines on real-world benchmarks.", "date": "2026-05-27", "categories": []}
{"id": "2605.28251", "title": "Counterfactually Fair Regression via Optimal Transport", "abstract": "We consider the problem of learning a counterfactually fair regressor. We adopt a causal uncertainty view in which counterfactual fairness is defined with resampled noise. We focus on obtaining theoretical fairness guarantees for a new post-processing estimator. We begin by showing that counterfactual fairness is equivalent to satisfying demographic parity conditional on the latent variable. This allows us to provide a closed-form expression of the optimal fair regressor via a barycentric quantile map. In order to handle continuous latent variables, we propose a discretized post-processing method. Then, under mild regularity assumptions, we prove high-probability finite-sample fairness guarantees for our estimator, providing an unfairness decay at rate $\\tilde O(n^{-1/3})$, and establishing a matching risk bound of order $\\tilde O(n^{-1/3})$. We provide a matching lower bound on the excess risk of almost fair predictions. Finally, we extend our results to the setting of relaxed counterfactual fairness. We validate our approach on real-world and synthetic data.", "date": "2026-05-27", "categories": []}
{"id": "2605.28450", "title": "BiasEdit: A Training-Free Bias-Detect-and-Edit Framework for Learning Fair Visual Classifiers", "abstract": "Visual data from the Web power image classifiers, which often underpin many web services, such as recommendation and content moderation. However, the raw Web data often contain spurious correlations and social biases, and neural networks are known for their tendency to learn biases present in data. This can reinforce unfairness in web services and the web data, leading to a vicious cycle. In the context of image classification, networks learn bias attributes for a specific class when a majority of images contain the same attribute only for a given class. Hence, training a fair and debiased classifier from a biased dataset demands handling an imbalanced problem between a majority of images with bias attributes (bias-aligned samples) and a minority without (bias-conflict samples). In this work, we introduce BiasEdit, a modular framework that automatically detects bias attributes from the original dataset and edits them to construct a debiased dataset. Specifically, BiasEdit first detects unknown bias attributes via statistical dependence and mutual information analysis of visual-linguistic representations, and then explicitly edits those attributes using text-guided image editing to generate realistic bias-conflict samples. Unlike prior works that assume known bias attributes or relies on synthetic mixing, our method operates without manual annotations and can leverage off-the-shelf vision-language and editing models. BiasEdit addresses a fundamental challenge in Web-sourced visual AI, mitigating dataset-induced bias and achieving state-of-the-art debiasing performance even when training data are fully biased.", "date": "2026-05-27", "categories": []}
{"id": "2605.28626", "title": "When Interpretability Is Unequally Distributed: Fairness in Hybrid Interpretable Models", "abstract": "Hybrid interpretable models combine a transparent component with a black-box model by assigning some examples to the former and deferring the rest to the latter. While this design enables flexible tradeoffs between accuracy and interpretability, it also raises a distinct procedural fairness concern: some demographic groups may systematically receive interpretable decisions, while others are disproportionately routed to a black box.\n  We formalize this issue as Interpretability Coverage Disparity (ICD), a demographic-parity-style measure applied to the routing decision of hybrid interpretable models. Using tools from predictive multiplicity, we study ICD across four hybrid interpretable learning methods, three standard fairness benchmark datasets, and multiple sensitive attributes. Our experiments reveal substantial ICD in intermediate transparency regimes, where both the interpretable and black-box components are actively used. We further show that simple coverage-disparity constraints can significantly reduce ICD in exact hybrid learning methods, with marginal impact on accuracy and sparsity. In several settings, ICD mitigation also improves standard algorithmic fairness metrics. These results show that hybrid interpretable models should be audited not only for predictive fairness, but also for how they allocate interpretability across individuals and groups.", "date": "2026-05-27", "categories": []}
{"id": "2605.28780", "title": "Bias Leaves a Gradient Trail: Label-Free Bias Identification via Gradient Probes on Concept Decompositions", "abstract": "Vision classifiers can exploit spurious correlations, achieving high in-distribution accuracy yet failing under distribution shift. Existing approaches to bias mitigation and analysis often depend on curated datasets, spurious-attribute or group labels, or retraining, which may be infeasible once a model is deployed or the relevant bias is unknown. We present a bias-label-free, post-hoc method for identifying spurious concepts in frozen vision models, relying only on standard class labels from a held-out audit dataset. For each target class, we collect patches from inputs predicted as that class and apply non-negative matrix factorization to intermediate activations to obtain a bank of interpretable concept vectors. Candidate concepts are then ranked with a bias estimator derived from their interaction with backpropagated gradients on misclassified examples: bias concepts tend to get activated when correcting false negatives and suppressed when correcting false positives. On Colored MNIST and Waterbirds the method recovers concepts aligned with the known spurious cue, and on CelebA it surfaces decision-relevant directions that only partially coincide with the annotated gender attribute; suppressing the top-ranked concepts at inference time improves worst-group accuracy by up to 17.9 percentage points on Waterbirds and 10.4 on CelebA without any retraining or parameter updates. Our method identifies decision-relevant spurious directions that need not coincide with annotated ones, providing both an interpretable auditing tool and an actionable debiasing handle for frozen vision models. Code is available at https://github.com/vitryt/label-free-bias-identification.", "date": "2026-05-27", "categories": []}
{"id": "2605.27371", "title": "Algorithmic Monocultures in Hiring", "abstract": "Many employers screen job applicants with algorithms built by the same few algorithm vendors. We hypothesize that algorithmic monoculture leads to the same individuals and members of the same racial groups facing rejection. We acquire and analyze a novel dataset of 3 million applicants submitting 4 million applications where all the applications are screened by algorithms built by the same vendor. We find clear racial disparities in applicant outcomes. Of all applications submitted by Asian and Black applicants, 14.74% and 25.87% are submitted to positions that adversely impact Asian and Black applicants, respectively, according to U.S. employment discrimination standards. Individuals also receive homogeneous outcomes: 4% of all applicants who apply to 10 positions are recommended for rejection from all positions, a rate higher than expected by chance. To better understand this homogeneity, we leverage the deterministic replicability of hiring algorithms to generate the outcomes applicants would have received if they applied to all positions. We show that applicants would need to apply widely in order to ensure their applications are considered by a human", "date": "2026-05-26", "categories": []}
{"id": "2510.22827", "title": "FairJudge: Abstention-Aware Multimodal Judges for Fairness and Alignment Evaluation in Text-to-Image Models", "abstract": "Evaluating text-to-image (T2I) systems requires judging not only whether an image matches a prompt, but also whether socially salient attributes are represented faithfully and without unsupported inference. Existing automated evaluators typically rely on face-centric recognizers or contrastive image--text similarity, which provide limited diagnostic feedback and often force predictions even when visual evidence is ambiguous or absent. For fairness-sensitive attributes such as religion and disability, where cues may be contextual, indirect, or intentionally unspecified, these evaluators can therefore miss failure modes that careful human reviewers would notice.\n  We introduce \\textsc{FairJudge}, an abstention-aware evaluation protocol that uses instruction-following multimodal LLMs as structured judges for social-attribute prediction, profession grounding, and prompt--image alignment. The protocol constrains outputs to closed label sets, requires visible-evidence rationales, supports an explicit \\textsc{unspecified} decision when cues are insufficient, and maps rubric-based alignment judgments to $[-1,1]$. These constraints turn MLLM judging from open-ended assessment into a parseable, auditable evaluation procedure.\n  Across four attribute-prediction benchmarks and three profession/alignment benchmarks, \\textsc{FairJudge} outperforms or complements CLIP, DeepFace, VIEScore, and VQAScore. Ablations show that closed labels, abstention, and evidence reporting are central to reliability. We further introduce \\textsc{DIVERSIFY} and \\textsc{DIVERSIFY-Professions}, two context-rich resources for evaluating social representation and profession grounding beyond face-visible or iconic cues. We release code, prompts, datasets, parser logs, and per-image judge outputs to support reproducible auditing.", "date": "2025-10-26", "categories": []}
{"id": "2603.00777", "title": "DUCX: Decomposing Unfairness in Tool-Using Chest X-ray Agents", "abstract": "Fairness in medical agents is becoming critical as tool-using clinical AI systems orchestrate specialized vision and language modules for tasks such as chest X-ray question answering. While these medical AI agents can improve flexibility, their added pipeline complexity also creates new pathways for demographic bias beyond standalone models. We present DUCK, Decomposing Unfairness in Chest X-ray agents, a systematic audit of fairness in tool-using chest X-ray agents instantiated with MedRAX. To localize where disparities arise, we introduce a stage-wise fairness decomposition that separates end-to-end bias from three agent-specific sources: tool exposure bias, or utility gaps conditioned on tool presence; tool transition bias, or subgroup differences in tool-routing patterns; and model reasoning bias, or subgroup differences in synthesis behaviors. Extensive experiments on tool-using agentic frameworks across five driver backbones reveal that demographic gaps persist in end-to-end performance, with equalized odds up to 20.79% and the lowest fairness-utility tradeoff down to 28.65%. Intermediate behaviors, including tool usage, transition patterns, and reasoning traces, exhibit distinct subgroup disparities that are not predictable from end-to-end evaluation alone. For example, conditioned on segmentation-tool availability, the subgroup utility gap reaches as high as 50%. Our findings underscore the need for process-level fairness auditing and debiasing to ensure the equitable deployment of clinical agentic systems. Code: https://github.com/Nanboy-Ronan/DUCK.", "date": "2026-02-28", "categories": []}
{"id": "2605.24458", "title": "Balancing Fairness, Privacy, and Accuracy: A Multitask Adversarial Framework for Centralized Data-Driven Systems", "abstract": "The integration of fairness and privacy in centralized data-driven applications is critical, especially as these systems increasingly influence sectors with significant societal impact. Current methods rarely address privacy, fairness, and accuracy together, which can potentially compromise ethical standards and privacy regulations. However, balancing these three objectives is quite challenging since each of objective often imposes conflicting requirements on the design and training of models, making it difficult to optimize one without compromising the others. This paper introduces a novel multitask adversarial model that treats fairness and privacy as integral objectives rather than afterthoughts, and learns a latent representation that hides sensitive attributes while preserving essential task-related information. Our approach dynamically balances fairness with accuracy and privacy through an optimized cost function with minimal performance loss even under strict conditions. Extensive testing on diverse datasets shows the ability of our model to achieve high standards of fairness and privacy without significant sacrifice to accuracy. Benchmarking against state-of-the-art privacy and fairness standards shows that our method enhances the robustness of privacy, fairness, and accuracy optimization, proving its adaptability across various datasets.", "date": "2026-05-23", "categories": []}
{"id": "2605.24687", "title": "HoloFair: Unified T2I Fairness Evaluation and Fair-GRPO Debiasing", "abstract": "Text-to-Image (T2I) models have made significant strides in visual realism and semantic consistency, yet they often perpetuate and amplify societal biases. Existing evaluation methods typically address only single-dimensional biases, lacking perspectives to uncover model biases at social-related deeper semantic levels. We introduce HoloFair, a comprehensive benchmark framework for multidimensional demographic bias analysis. Built upon our large-scale fairness-oriented dataset and the SpaFreq (Spatial-Frequency) attribute classifier, this framework proposes the Multi-attribute, Group-wise Bias Index (MGBI) metric, designed to assess both intrinsic diversity and conditional biases. Beyond evaluation, we further introduce Fair-GRPO, a reinforcement-learning-based debiasing method that alters the distribution of generative models through a designed multi-objective reward function. E.g., experiments on the SD3.5-Medium model demonstrate that Fair-GRPO significantly improves multidimensional fairness while maintaining high image quality. We also analyze potential reward hacking phenomena and provide corresponding mitigation strategies. Code and dataset are available at https://github.com/1059684669/HoloFair", "date": "2026-05-23", "categories": []}
{"id": "2605.24926", "title": "Energy Shields for Fairness", "abstract": "Runtime fairness is not a one-time constraint but a dynamic property evaluated over a sequence of decisions.\n  To ensure fairness at runtime, it is necessary to account for past decisions, information neglected by conventional, static classifiers.\n  Traditional fairness shields enforce runtime fairness abruptly,\n  by intervening \\emph{deterministically} whenever a sequence of decisions violates the target for a running fairness measure. This motivates our \\emph{main conceptual contribution: \\textbf{energy shields}.}\n  An energy shield is a novel, lightweight, adaptive controller that monitors a sequence of decisions and intervenes \\emph{probabilistically} to ensure runtime fairness smoothly, by utilizing physics-inspired energy functions to nudge the sequence toward fairness:\n  the more unfair the decisions, the stronger the nudging force becomes. This makes energy shields the \\emph{\\textbf{first}} fairness shields to provide both \\emph{short-term safety and long-term liveness guarantees}.\n  Safety ensures that the running fairness measure stays within a running target interval with high probability,\n  and liveness ensures that the limit of the fairness measure lies within the limit target interval.\n  Intuitively, the short-term specifies the tolerated fairness values and the long-term specifies the desired fairness values.\n  We also provide a synthesis procedure for constructing the least intrusive energy shield for a given target specification, and demonstrate its efficiency experimentally.\n  We evaluate our energy shields against existing fairness shields through the lens of short- and long-term fairness.", "date": "2026-05-24", "categories": []}
{"id": "2605.25228", "title": "A Blended Likelihood Approach for Achieving Fairness Using Naive Bayes", "abstract": "Concerns about algorithmic bias and fairness have increased as artificial intelligence has been incorporated into high-stakes decision-making. Traditional Naive Bayes classifiers, while efficient and interpretable, lack fairness-awareness mechanisms and perpetuate historical biases in sensitive domains such as hiring, credit scoring, and criminal justice. This study develops a fairness-aware extension of the Naive Bayes classifier that mitigates bias while maintaining computational efficiency. We propose the Bias Mitigating Naive Bayes (BMNB) classifier, integrating in-processing and post-processing interventions. The in-processing stage employs a blended likelihood approach combining group-specific and pooled likelihood estimates through a tunable blending parameter alpha to balance fairness and accuracy. The post-processing stage applies output calibration with adaptive thresholding to fine-tune group-specific decision boundaries. Experimental results indicate that BMNB attains Disparate Impact (DI) values of 1.000, 1.171, and 0.997 and Equal Opportunity Difference (EOD) values of -0.217, -0.226, and -0.053 on the Adult, ProPublica, and Framingham datasets, respectively, while maintaining computational efficiency. Ablation studies confirm that the combination of blended likelihood and adaptive thresholding yields superior performance compared to either technique in isolation.", "date": "2026-05-24", "categories": []}
{"id": "2605.23145", "title": "Operationalizing Individual Fairness via Gradient Descent and Bradley-Terry Models", "abstract": "Individual fairness, the notion that \"similar individuals should be treated similarly,\" provides a strong and flexible fairness guarantee for algorithmic decision makers. However, a barrier to implementing individual fairness in practice is the difficulty of learning the similarity metric over individuals. In this work, we present an algorithm for learning a Mahalanobis similarity metric from triplet queries of the form \"is individual $i$ more similar to individual $j$ or $k$?\" We work in the standard Bradley-Terry model for pairwise comparisons. Our algorithm consists of a spectral initialization step followed by gradient descent. We provide extensive theoretical guarantees on our algorithm, showing that it converges quickly to the ground truth metric despite the non-convexity of the loss in our model. Because our focus is on fairness, we also show that individual fairness with respect to an estimated metric is sufficient to achieve similar fairness with respect to the true metric. We also discuss potential applications of our work to AI model tuning. Finally, we present experimental results that demonstrate the convergence of our algorithm and the fairness performance of downstream fair predictors trained on our estimated metric.", "date": "2026-05-22", "categories": []}
{"id": "2605.00185", "title": "Fair Dataset Distillation via Cross-Group Barycenter Alignment", "abstract": "Dataset Distillation aims to compress a large dataset into a small synthetic one while maintaining predictive performance. We show that as different demographic groups exhibit distinct predictive patterns, the distillation process struggles to simultaneously preserve informative signals for all subgroups, regardless of whether group sizes are mildly or severely imbalanced. Consequently, models trained on distilled data can experience substantial performance drops for certain subgroups, leading to fairness gaps. Crucially, these gaps do not disappear by merely correcting group imbalance, since they stem from fundamental mismatches in subgroup predictive patterns rather than from sample-size disparities alone. We therefore formally analyze the interaction between these two sources of bias and cast the solution as identifying a group-imbalance-agnostic barycenter of the predictive information that induces similar representations across all subgroups. By distilling toward this shared aggregate representation, we show that group fairness concerns can be reduced. Our approach is compatible with existing distillation methods, and empirical results show that it substantially reduces bias introduced by dataset distillation. Code is available at https://github.com/mhmoslemi/COBRA.", "date": "2026-04-30", "categories": []}
{"id": "2605.20069", "title": "Smooth Partial Lotteries for Stable Randomized Selection", "abstract": "Competitive selection processes, from scientific funding to admissions and hiring, use evaluations to score candidates, and eventually choose a subset of them based on those scores. Recently, many organizations have adopted partial lotteries, which randomize selection based on evaluation scores. However, existing lottery designs are inherently unstable, as a small change to a single candidate's score can cause large shifts in their selection probabilities. This instability undermines a key goal of lotteries: reducing the influence of fine-grained score distinctions near the decision boundary. We propose smoothness as a design principle for partial lotteries, formalizing it as a Lipschitz condition on the mapping from review scores over candidates to selection probabilities. We introduce the Clipped Linear Lottery, a simple mechanism in which selection probabilities scale linearly with estimated quality between an upper threshold, above which we always accept, and a lower threshold, below which we always reject. We prove that the Clipped Linear Lottery's worst-case regret matches a lower bound for any smooth selection rule up to a factor of $(1 - k/n)$, where $k/n$ is the acceptance rate. We compare smooth selection to other stability notions like Individual Fairness and Differential Privacy, showing that the Clipped Linear Lottery achieves a better smoothness-regret tradeoff than alternatives. Experiments on real peer review data from ICLR 2025, NeurIPS 2024, and the Swiss National Science Foundation demonstrate that existing lottery designs are highly unstable in practice even under perturbations to a single score. Our experiments also confirm the tightness of our theoretical analysis and show that our proposed Clipped Linear Lottery achieves a better smoothness-utility tradeoff than alternatives in practice.", "date": "2026-05-19", "categories": []}
{"id": "2605.22291", "title": "Long-term Fairness with Selective Labels", "abstract": "Long-term fairness algorithms aim to satisfy fairness beyond static and short-term notions by accounting for the dynamics between decision-making policies and population behavior. Most previous approaches evaluate performance and fairness measures from observable features and a label, which is assumed to be fully observed. However, in scenarios such as hiring or lending, the labels (e.g., ability to repay the loan) are selective labels as they are only revealed based on positive decisions (e.g., when a loan is granted). In this paper, we study long-term fairness in the selective labels setting and analytically show that naive solutions do not guarantee fairness. To address this gap, we then introduce a novel framework that leverages both the observed data and a label predictor model to estimate the true fairness measure value by decomposing it into the observed fairness and bias from label predictions. This allows us to derive sufficient conditions to satisfy true fairness from observable quantities by using the confidence in the predictor model. Finally, we rely on our theoretical results to propose a novel reinforcement learning algorithm for effective long-term fair decision-making with selective labels. In semisynthetic environments, the proposed algorithm reached comparable fairness and performance to an agent with oracle access to the true labels.", "date": "2026-05-21", "categories": []}
{"id": "2501.01785", "title": "Can Synthetic Data be Fair and Private? A Comparative Study of Synthetic Data Generation and Fairness Algorithms", "abstract": "  The increasing use of machine learning in learning analytics (LA) has raised significant concerns around algorithmic fairness and privacy. Synthetic data has emerged as a dual-purpose tool, enhancing privacy and improving fairness in LA models. However, prior research suggests an inverse relationship between fairness and privacy, making it challenging to optimize both. This study investigates which synthetic data generators can best balance privacy and fairness, and whether pre-processing fairness algorithms, typically applied to real datasets, are effective on synthetic data. Our results highlight that the DEbiasing CAusal Fairness (DECAF) algorithm achieves the best balance between privacy and fairness. However, DECAF suffers in utility, as reflected in its predictive accuracy. Notably, we found that applying pre-processing fairness algorithms to synthetic data improves fairness even more than when applied to real data. These findings suggest that combining synthetic data generation with fairness pre-processing offers a promising approach to creating fairer LA models.", "date": "2025-01-03", "categories": []}
{"id": "2512.23943", "title": "Statistical Guarantees in the Search for Less Discriminatory Algorithms", "abstract": "U.S. discrimination law can impose liability on firms that fail to adopt a less discriminatory alternative (LDA): a decision policy that achieves the same business objectives while reducing disparate impact on legally protected groups. Recent scholarship argues that this doctrine has direct implications for algorithmic decision-making in high-stakes domains such as employment, lending, and housing, potentially obligating firms to search for \"less discriminatory algorithms\" (Black et al., 2024). Regulators have at times encouraged proactive LDA searches, reinforcing the expectation of a good-faith effort to identify equally performant models with lower disparate impact.\n  Model multiplicity makes such searches plausible: retraining with different random seeds can yield models with comparable predictive performance but materially different disparate impacts. Yet firms cannot retrain indefinitely, raising a central question: when is the search sufficient to demonstrate good faith?\n  We formalize LDA search under multiplicity as an optimal stopping problem in which a developer seeks to produce evidence that further search is unlikely to yield meaningful improvements. Our main contribution is an adaptive stopping algorithm that provides a high-probability upper bound on the best disparate-impact gains attainable through continued retraining, enabling developers to certify (e.g., to a court) that additional search is unlikely to help. We also show how stronger distributional assumptions over the model space can yield tighter bounds, and we validate the approach on real-world credit and housing datasets.", "date": "2025-12-30", "categories": []}
{"id": "2603.28675", "title": "Why Aggregate Accuracy is Inadequate for Evaluating Fairness in Law Enforcement Facial Recognition Systems", "abstract": "Facial recognition systems are increasingly deployed in law enforcement and security contexts, where algorithmic decisions can carry significant societal consequences. Despite high reported accuracy, growing evidence demonstrates that such systems often exhibit uneven performance across demographic groups, leading to disproportionate error rates and potential harm. This paper argues that aggregate accuracy is an insufficient metric for evaluating the fairness and reliability of facial recognition systems in high-stakes environments. Through analysis of subgroup-level error distribution, including false positive rate (FPR) and false negative rate (FNR), the paper demonstrates how aggregate performance metrics can obscure critical disparities across demographic groups. Empirical observations show that systems with similar overall accuracy can exhibit substantially different fairness profiles, with subgroup error rates varying significantly despite a single aggregate metric. The paper further examines the operational risks associated with accuracy-centric evaluation practices in law enforcement applications, where misclassification may result in wrongful suspicion or missed identification. It highlights the importance of fairness-aware evaluation approaches and model-agnostic auditing strategies that enable post-deployment assessment of real-world systems. The findings emphasise the need to move beyond accuracy as a primary metric and adopt more comprehensive evaluation frameworks for responsible AI deployment.", "date": "2026-03-30", "categories": []}
{"id": "2604.15038", "title": "When Fairness Metrics Disagree: Evaluating the Reliability of Demographic Fairness Assessment in Machine Learning", "abstract": "The evaluation of fairness in machine learning systems has become a central concern in high-stakes applications, including biometric recognition, healthcare decision-making, and automated risk assessment. Existing approaches typically rely on a small number of fairness metrics to assess model behaviour across group partitions, implicitly assuming that these metrics provide consistent and reliable conclusions. However, different fairness metrics capture distinct statistical properties of model performance and may therefore produce conflicting assessments when applied to the same system. In this work, we investigate the consistency of fairness evaluation by conducting a systematic multi-metric analysis of demographic bias in machine learning models. Using face recognition as a controlled experimental setting, we evaluate model performance across multiple group partitions under a range of commonly used fairness metrics, including error-rate disparities and performance-based measures. Our results demonstrate that fairness assessments can vary significantly depending on the choice of metrics, leading to contradictory conclusions regarding model bias. To quantify this phenomenon, we introduce the Fairness Disagreement Index (FDI), a measure designed to capture the degree of inconsistency across fairness metrics. We further show that disagreement remains high across thresholds and model configurations. These findings highlight a critical limitation in current fairness evaluation practices and suggest that single-metric reporting is insufficient for reliable bias assessment.", "date": "2026-04-16", "categories": []}
{"id": "2605.21455", "title": "Mitigating Label Bias with Interpretable Rubric Embeddings", "abstract": "Statistical decision algorithms are increasingly deployed in domains where ground-truth labels are hard to obtain, such as hiring, university admissions, and content moderation. In these settings, models are typically trained on historical human evaluations -- for example, using past hiring decisions as a proxy for true applicant quality. However, if past evaluations unjustly favor certain groups, models trained on these labels may inherit those biases. To address this problem, we propose basing predictions on rubric embeddings, a representation framework that replaces standard black-box embeddings with features derived from expert-defined criteria that align with the underlying construct of interest. By anchoring predictions to semantically meaningful dimensions, this approach guards against biased proxy signals. We provide both theoretical and empirical evidence that rubric embeddings mitigate label bias under plausible conditions. Empirically, we evaluate our method on a novel dataset of applications to a large master's program. We find that models trained on rubric embeddings reduce group disparities while improving measures of cohort quality. Our results suggest that basing predictions on interpretable, domain-grounded representations offers a practical approach to learning in the presence of biased labels.", "date": "2026-05-20", "categories": []}
{"id": "2508.06649", "title": "Measuring Stereotype and Deviation Biases in Large Language Models", "abstract": "Large language models (LLMs) are widely applied across diverse domains, raising concerns about their limitations and potential risks. In this study, we investigate two types of bias that LLMs may display: stereotype bias and deviation bias. Stereotype bias refers to when LLMs consistently associate specific traits with a particular demographic group. Deviation bias reflects the disparity between the demographic distributions extracted from LLM-generated content and real-world demographic distributions. By asking four advanced LLMs to generate profiles of individuals, we examine the associations between each demographic group and attributes such as political affiliation, religion, and sexual orientation. Our experimental results show that all examined LLMs exhibit both significant stereotype bias and deviation bias towards multiple groups. Our findings uncover the biases that occur when LLMs infer user attributes and shed light on the potential harms of LLM-generated outputs.", "date": "2025-08-08", "categories": []}
{"id": "2510.23507", "title": "A Deep Latent Factor Graph Clustering with Fairness-Utility Trade-off Perspective", "abstract": "Fair graph clustering seeks partitions that respect network structure while maintaining proportional representation across sensitive groups, with applications spanning community detection, team formation, resource allocation, and social network analysis. Many existing approaches enforce rigid constraints or rely on multi-stage pipelines (e.g., spectral embedding followed by $k$-means), limiting trade-off control, interpretability, and scalability. We introduce \\emph{DFNMF}, an end-to-end deep nonnegative tri-factorization tailored to graphs that directly optimizes cluster assignments with a soft statistical-parity regularizer. A single parameter $\\lambda$ tunes the fairness--utility balance, while nonnegativity yields parts-based factors and transparent soft memberships. The optimization uses sparse-friendly alternating updates and scales near-linearly with the number of edges. Across synthetic and real networks, DFNMF achieves substantially higher group balance at comparable modularity, often dominating state-of-the-art baselines on the Pareto front. The code is available at https://github.com/SiamakGhodsi/DFNMF.git.", "date": "2025-10-27", "categories": []}
{"id": "2605.19214", "title": "Worst-Group Equalized Odds Regularization for Multi-Attribute Fair Medical Image Classification", "abstract": "Diagnostic performance in medical AI varies systematically across demographic groups, yet subgroup AUC can mask clinically important disparities. At a fixed inference-time operating point, some groups may exhibit over-diagnostic behaviour, characterized by elevated true and false positive rates, while others show under-diagnostic patterns with reduced true and false positive rates. These opposing tendencies can cancel in aggregate AUCs while producing meaningful inequities in clinical decision-making. Motivated by the need to assess and mitigate such disparities at the operating point and across multiple demographic attributes simultaneously, we propose a worst-group equalized-odds margin regularizer. The proposed regularizer explicitly targets subgroup-level deviations on both the true positive and false positive sides at inference. At each update, the method identifies subgroups defined by explicit demographic attributes (e.g., age, sex, and race) that exhibit the most extreme margin deviations and applies a unified penalty, enabling fairness optimization across multiple demographic axes without requiring explicit intersectional constraints. Across two medical imaging datasets in realistic multi-label settings, our method consistently reduces disparities in Equalized Odds and Equalized Opportunity with minimal impact on AUC, preserving diagnostic performance while improving fairness.", "date": "2026-05-19", "categories": []}
{"id": "2605.19393", "title": "Neuron Incidence Redistribution for Fairness in Medical Image Classification", "abstract": "Deep learning models for medical image classification are susceptible to subgroup performance disparities across demographic attributes such as age, gender, and race. We identify a latent representational mechanism underlying these disparities: in transfer-learned models, the dominant penultimate-layer activation channel under positive predictions is co-activated by both disease-positive samples and privileged demographic groups (male, older patients), producing over-diagnosis; conversely, the dominant channel under negative predictions is co-activated by disadvantaged groups (female, younger patients), producing systematic under-diagnosis. To address this, we propose Neuron Incidence Redistribution (NIR), a lightweight regularization method that penalizes the variance of predicted-probability-weighted mean activations across penultimate-layer neurons, requiring no demographic labels at training time. On HAM10000, TPR disparity drops from 10.81% to 0.93% across age groups and from 12.04% to 0.74% across gender, with a marginal AUC improvement of 0.51 points. On Harvard OCT-RNFL, NIR reduces FPR disparity for race (from 15.68% to 10.66%) and age (from 12.69% to 1.80%), demonstrating that distributing latent disease evidence across the full penultimate layer is a principled and effective strategy for improving demographic fairness in medical AI.", "date": "2026-05-19", "categories": []}
{"id": "2411.10636", "title": "Mitigating Extrinsic Gender Bias for Bangla Classification Tasks", "abstract": "In this study, we investigate extrinsic gender bias in Bangla pretrained language models, a largely underexplored area in low-resource languages. To assess this bias, we construct four manually annotated, task-specific benchmark datasets for sentiment analysis, toxicity detection, hate speech detection, and sarcasm detection. Each dataset is augmented using nuanced gender perturbations, where we systematically swap gendered names and terms while preserving semantic content, enabling minimal-pair evaluation of gender-driven prediction shifts. We then propose RandSymKL, a randomized debiasing strategy integrated with symmetric KL divergence and cross-entropy loss to mitigate the bias across task-specific pretrained models. RandSymKL is a refined training approach to integrate these elements in a unified way for extrinsic gender bias mitigation focused on classification tasks. Our approach was evaluated against existing bias mitigation methods, with results showing that our technique not only effectively reduces bias but also maintains competitive accuracy compared to other baseline approaches. To promote further research, we have made both our implementation and datasets publicly available: https://github.com/sajib-kumar/Mitigating-Bangla-Extrinsic-Gender-Bias", "date": "2024-11-16", "categories": []}
{"id": "2601.14506", "title": "Compounding Disadvantage: Auditing Intersectional Bias in LLM-Generated Explanations Across Indian and American STEM Education", "abstract": "Large language models are increasingly deployed in STEM education for personalized instruction and feedback across institutions in high- and low-income countries. These systems are designed to adapt content to student needs, but whether they adapt based on demonstrated ability or demographic signals remains untested at scale. Here we establish that LLM-generated STEM content systematically disadvantages marginalized student profiles across two cultural contexts, with the gap between the most privileged and most marginalized profiles reaching 2.55 grade levels. We audited four LLMs (Qwen 2.5-32B-Instruct, GPT-4o, GPT-4o-mini, GPT-OSS 20B) using synthetic profiles crossing dimensions specific to Indian education (caste, medium of instruction, college tier) and American education (race, HBCU attendance, school type), alongside income, gender, and disability, across ranking and generation tasks with FDR-corrected significance testing and SHAP feature attribution. Income produces significant effects across every model and context, medium of instruction drives the largest single effect in the Indian context, and disability status triggers simpler explanations. Effects compound non-additively: marginalization across multiple dimensions produces gaps larger than any single dimension predicts, and biases persist within elite institutions. Bias is consistent across all four architectures and persists through model selection, making intersectional, cross-cultural auditing a structural requirement before deployment.", "date": "2026-01-20", "categories": []}
{"id": "2605.11365", "title": "Causal Bias Detection in Generative Artificial Intelligence", "abstract": "Automated systems built on artificial intelligence (AI) are increasingly deployed across high-stakes domains, raising critical concerns about fairness and the perpetuation of demographic disparities that exist in the world. In this context, causal inference provides a principled framework for reasoning about fairness, as it links observed disparities to underlying mechanisms and aligns naturally with human intuition and legal notions of discrimination. Prior work on causal fairness primarily focuses on the standard machine learning setting, where a decision-maker constructs a single predictive mechanism $f_{\\widehat Y}$ for an outcome variable $Y$, while inheriting the causal mechanisms of all other covariates from the real world. The generative AI setting, however, is markedly more complex: generative models can sample from arbitrary conditionals over any set of variables, implicitly constructing their own beliefs about all causal mechanisms rather than learning a single predictive function. This fundamental difference requires new developments in causal fairness methodology. We formalize the problem of causal fairness in generative AI and unify it with the standard ML setting under a common theoretical framework. We then derive new causal decomposition results that enable granular quantification of fairness impacts along both (a) different causal pathways and (b) the replacement of real-world mechanisms by the generative model's mechanisms. We establish identification conditions and introduce efficient estimators for causal quantities of interest, and demonstrate the value of our methodology by analyzing race and gender bias in large language models across different datasets.", "date": "2026-05-12", "categories": []}
{"id": "2605.17118", "title": "Differentiable Optimization Layers for Guaranteed Fairness in Deep Learning", "abstract": "Differentiable optimization layers are traditionally integrated in predict-then-optimize frameworks where a neural model estimates parameters that subsequently serve as fixed inputs to downstream decision-making optimization problems. In this work, we introduce the concept of a \"fairness layer\": a differentiable optimization layer appended to a model's output layer that guarantees a chosen notion of output parity is satisfied when integrated into a neural network. Additionally, we introduce an online primal-dual inference algorithm that provides provable aggregate fairness guarantees for streaming predictions with arbitrarily small batch sizes, where traditional per-batch constraints become overly restrictive. Numerical experiments demonstrate the effectiveness of the fairness layer and associated algorithm, and theoretical analysis characterizes the layer's differentiability and stability properties during model training and backpropagation. Our code for these experiments is publicly available on GitHub (https://github.com/dtroxell19/FairDL-ICML-2026.git) and our public Python package documentation can be found online: https://dtroxell19.github.io/fairness_training/.", "date": "2026-05-16", "categories": []}
{"id": "2603.13452", "title": "MESD: A Risk-Sensitive Metric for Explanation Fairness Across Intersectional Subgroups", "abstract": "Fairness in machine learning is predominantly evaluated through outcome-oriented metrics, such as Demographic parity, which measure whether predictions are statistically consistent across protected groups. However, these metrics cannot detect whether a model uses systematically different reasoning for different demographic groups, which violates procedural fairness principles. This problem is compounded by intersectionality, where models may appear fair on individual attributes (e.g., race) while exhibiting significant disparities for intersectional subgroups (e.g., race $\\times$ gender), a phenomenon known as fairness gerrymandering. In this work, we introduce Multi-category Explanation Stability Disparity (MESD), a procedural fairness metric that quantifies disparities in explanation quality across intersectional subgroups formed by the Cartesian product of multiple protected attributes. MESD integrates three components, which are label-aware aggregation aligned with outcome-conditional fairness, empirical-Bayes shrinkage to stabilize estimates for small intersectional groups, and Conditional Value-at-Risk (CVaR) weighting to emphasize worst-case subgroup disparities. We integrate MESD within a multi-objective optimization framework (UEF) that jointly optimizes utility, outcome fairness, and procedural fairness using NSGA-II. We evaluated MESD and UEF on three benchmark datasets along with four state-of-the-art methods in several experiments, and we demonstrate that MESD reveals procedural disparities invisible to outcome metrics alone. We position our contribution within procedural justice theory and discuss implications for regulatory compliance and intersectional equity.", "date": "2026-03-13", "categories": []}
{"id": "2605.15208", "title": "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels", "abstract": "Large Language Models are routinely compressed via post-training quantization to reduce inference costs and memory footprint for cloud and edge deployment, yet the impact of this compression on model quality remains poorly understood. Existing studies typically compare only two conditions (full-precision vs. a single quantized variant), rely on aggregate bias metrics, and evaluate a single model family, making it impossible to distinguish gradual degradation from threshold-dependent safety failures. We conduct a controlled empirical study of three instruction-tuned models (Qwen2.5-7B, Mistral-7B, Phi-3.5-mini) at five precision levels (BF16 through 3-bit) on 12,148 BBQ bias benchmark items across 5 random seeds, totaling 911,100 inference records. Our results reveal that 3-bit quantization causes 6-21% of previously unbiased items to develop new stereotypical behaviors, following a clear dose-response pattern confirmed via logistic regression, while models' willingness to select \"unknown\" answers declines by 17.4%. Crucially, these item-level changes are invisible to standard quality metrics: perplexity increases by less than 0.5% at 8-bit and under 3% at 4-bit across all three models, yet 2.5-5.6% of items already develop new biases at 4-bit. These findings demonstrate that aggregate evaluation metrics systematically miss fairness-critical degradation, underscoring the need for quality-aware compression protocols that explicitly test for bias emergence before deployment.", "date": "2026-05-02", "categories": []}
{"id": "2605.15295", "title": "GESD: Beyond Outcome-Oriented Fairness", "abstract": "Machine learning (ML) algorithms are increasingly deployed in high-stakes decision-making domains such as loan approvals, hiring, and recidivism predictions. While existing fairness metrics (e.g., statistical parity, equal opportunity) effectively quantify outcome-oriented disparities, they offer limited insight into the procedure or explanation behind biased decisions. To address this gap, we propose Group-level Explanation Stability Disparity (GESD), a \\textit{procedural-oriented} fairness metric that measures disparities in the stability, robustness, and sensitivity of model explanations across different subgroups in a protected category. %GESD is explainer-agnostic, model-agnostic, and extends the scope of fairness analyses to the level of explainability. We further integrate GESD into a multi-objective optimization framework that jointly optimizes for utility, outcome-based fairness, and explanation-based fairness called FEU (Fairness--Explainability--Utility). Empirical results on multiple benchmark datasets show that GESD effectively captures group-wise discrepancies in explanation quality, and that FEU improves both utility and fairness over state-of-the-art methods. By bridging outcome-based and explanation-based fairness, GESD offers a comprehensive tool for diagnosing and mitigating bias in predictive modeling. Our code and datasets are available on GitHub {\\hyperlink{https://github.com/horlahsunbo/GESD}{https://github.com/horlahsunbo/GESD}}", "date": "2026-05-14", "categories": []}
{"id": "2605.15312", "title": "Beyond Performance Disparities: A Three-Level Audit of Representational Harm in CelebA", "abstract": "Large-scale facial datasets like CelebA are widely used in computer vision, yet the cultural biases embedded in their labels remain underexplored. Fairness research has distinguished representational from allocational harms, but audits of computer vision datasets have mostly examined categorical labels, leaving open how such harms appear in learned features and model attention. This paper examines CelebA at three levels: dataset structure, learned feature weights, and spatial attention, focusing on how gendered double standards of ageing and beauty are encoded in the data and reproduced in model behaviour. First, hierarchical clustering of 202,599 images shows that the 39 attributes organise into latent trait bundles aligned with cultural archetypes: performative femininity (youth, makeup, adornment) and professional masculinity (ageing, facial hair, formal attire). Female faces, though more often rated attractive overall, incur steep penalties when assigned to ageing or masculine-coded clusters. Second, XGBoost with SHAP analysis reveal gender-specific effects, such as adiposity reducing attractiveness only for females. Third, Grad-CAM finds that predictions for female and younger male subgroups concentrate on mid-face cues, whereas predictions for older males drift toward peripheral cues such as hair and clothing. Older males attain the highest accuracy but the lowest average precision, indicating categorical exclusion of groups outside the dataset's evaluative templates. Cultural double standards thus pass from media representation into dataset labels, feature weights, and model attention, producing two representational harms: hyper-scrutiny of women under a narrow evaluative template, and exclusion of older men from the scheme entirely. Fairness metrics focused on performance disparities mask both, underscoring the need to address representational harm in fairness research.", "date": "2026-05-14", "categories": []}
{"id": "2605.16113", "title": "DebiasRAG: A Tuning-Free Path to Fair Generation in Large Language Models through Retrieval-Augmented Generation", "abstract": "Large language models (LLMs) have achieved unprecedented success due to their exceptional generative capabilities. However, because they depend on knowledge encapsulated from training corpora, they may produce hallucinations, stereotypes, and socially biased content. In particular, LLMs are prone to prejudiced responses involving race, gender, and age, which are collectively referred to as social biases. Prior studies have used fine-tuning and prompt engineering to mitigate such biases in LLMs, but these methods require additional training resources or domain knowledge to design the framework. Moreover, they may degrade the original capabilities of LLMs and often overlook the need for dynamic debiasing contexts for fairer inference. In this paper, we propose DebiasRAG, a novel tuning-free and dynamic query-specific debiasing framework based on retrieval-augmented generation (RAG). DebiasRAG improves fairness while preserving the intrinsic properties of LLMs, such as representation ability. DebiasRAG consists of three stages: (1) query-specific debiasing candidate generation; (2) context candidate pool construction; and (3) gradient-updated debiasing-guided context piece reranking. First, DebiasRAG leverages self-diagnosed bias contexts relevant to the query through regular retrieval, where the bias contexts are prepared offline by the DebiasRAG provider. Given the query-specific bias contexts, DebiasRAG reversely produces debiasing contexts, which are provided as additional fairness constraints for LLM outputs. Second, a regular RAG retrieval process produces query-related contexts from the regular RAG document database, such as a chunked Wikipedia dataset.", "date": "2026-05-15", "categories": []}
{"id": "2512.01766", "title": "On the Unreasonable Effectiveness of Last-layer Retraining", "abstract": "Last-layer retraining (LLR) methods -- wherein the last layer of a neural network is reinitialized and retrained on a held-out set following ERM training -- have garnered interest as an efficient approach to rectify dependence on spurious correlations and improve performance on minority groups. Surprisingly, LLR has been found to improve worst-group accuracy even when the held-out set is an imbalanced subset of the training set. We initially hypothesize that this ``unreasonable effectiveness'' of LLR is explained by its ability to mitigate neural collapse through the held-out set, resulting in the implicit bias of gradient descent benefiting robustness. Our empirical investigation does not support this hypothesis. Instead, we present strong evidence for an alternative hypothesis: that the success of LLR is primarily due to better group balance in the held-out set. We conclude by showing how the recent algorithms CB-LLR and AFR perform implicit group-balancing to elicit a robustness improvement.", "date": "2025-12-01", "categories": []}
{"id": "2605.14074", "title": "Fair and Calibrated Toxicity Detection with Robust Training and Abstention", "abstract": "Fairness in toxicity classification involves three integrated axes: ranking, calibration, and abstention. Training-time interventions and post-hoc safety mechanisms cannot be evaluated independently because the former determines the efficacy of the latter. We compare Empirical Risk Minimization (ERM), instance-level reweighting, and Group DRO across these axes, combined with temperature scaling, confidence-based abstention, and per-identity threshold optimization. Evaluation uses subgroup AUC, BPSN/BNSP AUC, error gaps, and per-subgroup Expected Calibration Error (ECE) with bootstrap CIs ($n = 1000$).\n  We report four findings. (1) Calibration disparity is a hidden fairness violation. ERM has near-perfect aggregate calibration ($0.013$) but is significantly miscalibrated across all identity subgroups ($+0.029$ to $+0.134$). (2) Training interventions reshape rather than eliminate disparity. Reweighted ERM improves ranking (BPSN AUC $+0.06$ to $+0.12$) but worsens the calibration-fairness gap by up to $+0.232$. Group DRO eliminates calibration disparity but only by becoming uniformly miscalibrated globally (ECE $0.118$). (3) Post-hoc methods inherit training failure modes. Temperature scaling fails because miscalibration is non-uniform. Confidence-based abstention works under ERM but breaks under DRO, where the risk-coverage curve rises with deferral. (4) Abstention itself is unfair. Confidence-based deferral helps background content far more than identity-mentioning content. We argue that SRAI fairness requires a multi-axis framework: methods that differ only in aggregate ranking can differ sharply in failure modes that determine real-world harm.", "date": "2026-05-13", "categories": []}
{"id": "2603.05582", "title": "Bias In, Bias Out? Finding Unbiased Subnetworks in Vanilla Models", "abstract": "The issue of algorithmic biases in deep learning has led to the development of various debiasing techniques, many of which perform complex training procedures or dataset manipulation. However, an intriguing question arises: is it possible to extract fair and bias-agnostic subnetworks from standard vanilla-trained models without relying on additional data, such as unbiased training set? In this work, we introduce Bias-Invariant Subnetwork Extraction (BISE), a learning strategy that identifies and isolates \"bias-free\" subnetworks that already exist within conventionally trained models, without retraining or finetuning the original parameters. Our approach demonstrates that such subnetworks can be extracted via pruning and can operate without modification, effectively relying less on biased features and maintaining robust performance. Our findings contribute towards efficient bias mitigation through structural adaptation of pre-trained neural networks via parameter removal, as opposed to costly strategies that are either data-centric or involve (re)training all model parameters. Extensive experiments on common benchmarks show the advantages of our approach in terms of the performance and computational efficiency of the resulting debiased model.", "date": "2026-03-05", "categories": []}
{"id": "2605.12530", "title": "In-Situ Behavioral Evaluation for LLM Fairness, Not Standardized-Test Scores", "abstract": "LLM fairness should be evaluated through in-situ conversational behavior rather than standardized-test Q&A benchmarks. We show that the standardized-test paradigm can be structurally unreliable: surface-level prompt construction choices, although entirely orthogonal to the fairness question being tested, account for the majority of score variance, shift fairness conclusions in both the direction and the magnitude, and result in severe discordance in model rankings. We develop MAC-Fairness, a multi-agent conversational framework that embeds controlled variation factors into multi-round dialogue for in-situ behavior evaluation, examining how models' conversational behavior shifts when identity is varied as part of natural multi-agent interaction. Repurposing standardized-test questions as conversation seeds rather than as the evaluation instrument, we evaluate position persistence (how they hold positions, from the self-perspective) and peer receptiveness (how receptive they are to peers, from the other-perspective) across 8 million conversation transcripts spanning multiple models and identity presence configurations. In-situ behavioral evaluation reveals stable, model-specific behavioral signatures that could generalize across benchmarks differing in fairness targets and evaluation methodologies, a form of evidence the standardized-test paradigm does not offer.", "date": "2026-04-21", "categories": []}
{"id": "2605.12701", "title": "Do Fair Models Reason Fairly? Counterfactual Explanation Consistency for Procedural Fairness in Credit Decisions", "abstract": "Machine learning algorithms in socially sensitive domains (e.g., credit decisions) often focus on equalizing predictive outcomes. However, satisfying these metrics does not guarantee that models use the same reasoning for different groups. We show that existing outcome-fair models can still apply fundamentally different reasoning to individuals, a ``hidden procedural bias'' missed by standard fairness metrics and algorithms. We propose Counterfactual Explanation Consistency (CEC), a framework that detects and mitigates this bias by aligning feature attributions between individuals and their counterfactual counterparts. Key contributions include a nearest-neighbor counterfactual generation method, a modified baseline for integrated gradient comparisons, an individual-level procedural fairness metric, and a corresponding training loss. We introduce a taxonomy identifying ``Regime B'' (same outcome, different reasoning) as a critical blind spot. Experiments on synthetic data, German Credit, Adult Income, and HMDA mortgage data demonstrate that outcome-fair baselines exhibit substantial hidden bias, while CEC substantially reduces it with modest utility cost.", "date": "2026-05-12", "categories": []}
{"id": "2605.13113", "title": "Context Matters: Auditing Gender Bias in T2I Generation through Risk-Tiered Use-Case Profiles", "abstract": "Text-to-image (T2I) generative models are increasingly used to produce content for education, media, and public-facing communication, and are starting to be integrated into higher-impact pipelines. Since generated images tend to reinforce stereotypes, producing representational erasure via \"default\" depictions and shaping perceptions of who belongs in certain roles, a growing body of work has proposed metrics to quantify gender bias in T2I outputs. Yet existing evaluations remain fragmented. Metrics are often reported without a shared view of what they measure, what assumptions they entail, or how their results should be interpreted under different deployment contexts. This limits the usefulness of gender bias measurement for both technical auditing and emerging governance discussions. We propose a risk-aligned auditing framework for gender bias in T2I models composed of three constituents that connects risk categories, evaluation metrics, and harms. First, we identify risk-tiered use-case profiles aligned with the EU AI Act's risk categories to motivate why auditing expectations may vary with deployment contexts and stakeholder exposure. Second, we construct a metric catalog that consolidates gender-bias evaluation methods and organizes them in three measurement categories: gender prediction, embedding similarity, and downstream task. Third, we introduce a harm typology that maps context-dependent harm categories (e.g., representational, quality-of-service) to specific risk-tired scenarios. Finally, we introduce THUMB cards (Text-to-image Harms-informed Use-case-aligned Metrics of gender Bias) that help formulate auditing systematically by the incorporation of context, scenario and bias manifestation, harm hypotheses, and audit strategy.", "date": "2026-05-13", "categories": []}
{"id": "2605.13759", "title": "Fast and effective algorithms for fair clustering at scale", "abstract": "Clustering is an unsupervised machine learning task that consists of identifying groups of similar objects. It has numerous applications and is increasingly used in fairness-sensitive domains where objects represent individuals, such as customers, employees, or students. We address a fair clustering problem in which objects belong to protected groups. The problem consists of partitioning the objects into a predefined number of clusters while attaining a user-defined target level of fairness, meaning that each protected group is sufficiently represented in each cluster. The objective is to minimize the clustering cost, defined as the sum of squared Euclidean distances between the objects and the centers of their clusters. Since clustering cost and fairness are generally in conflict, managing the trade-off between them is essential in practical applications. Existing methods provide limited control over this trade-off and either fail to scale to large datasets or, when they scale, produce low-quality solutions. We propose a general framework for fair clustering that provides precise control over the cost-fairness trade-off and introduce three heuristics based on it. The first heuristic focuses on solution quality and the flexibility to incorporate additional constraints, the second improves scalability while retaining high solution quality, and the third is designed for maximum scalability, producing solutions for instances with millions of objects in seconds. The proposed heuristics outperform existing approaches in comprehensive numerical experiments on benchmark datasets. The source code of our heuristics and instructions for reproducing the experiments are publicly available on GitHub.", "date": "2026-05-13", "categories": []}
{"id": "2605.11195", "title": "How Does Differential Privacy Affect Social Bias in LLMs? A Systematic Evaluation", "abstract": "Large language models (LLMs) trained on web-scale corpora can memorize sensitive training data, posing significant privacy risks. Differential privacy (DP) has emerged as a principled framework that limits the influence of individual data points during training, yet the relationship between differential privacy and social bias in LLMs remains poorly understood. To investigate this, we present a systematic evaluation of social bias in a pretrained LLM trained with DP-SGD, comparing a DP model against non-DP baselines across four complementary paradigms: sentence scoring, text completion, tabular classification, and question answering. We find that DP reduces bias in sentence scoring tasks, where bias is measured through controlled likelihood comparisons, yet this improvement does not generalize across all tasks. Our results reveal a discrepancy between logit-level bias and output-level bias. Moreover, decreasing memorization does not necessarily reduce unfairness, underscoring the importance of multi-paradigm evaluation when assessing fairness in LLMs.", "date": "2026-05-11", "categories": []}
{"id": "2605.11362", "title": "Causal Fairness for Survival Analysis", "abstract": "In the data-driven era, large-scale datasets are routinely collected and analyzed using machine learning (ML) and artificial intelligence (AI) to inform decisions in high-stakes domains such as healthcare, employment, and criminal justice, raising concerns about the fairness behavior of these systems. Existing works in fair ML cover tasks such as bias detection, fair prediction, and fair decision-making, but largely focus on static settings. At the same time, fairness in temporal contexts, particularly survival/time-to-event (TTE) analysis, remains relatively underexplored, with current approaches to fair survival analysis adopting statistical fairness definitions, which, even with unlimited data, cannot disentangle the causal mechanisms that generate disparities. To address this gap, we develop a causal framework for fairness in TTE analysis, enabling the decomposition of disparities in survival into contributions from direct, indirect, and spurious pathways. This provides a human-understandable explanation of why disparities arise and how they evolve over time. Our non-parametric approach proceeds in four steps: (1) formalizing the necessary assumptions about censoring and lack of confounding using a graphical model; (2) recovering the conditional survival function given covariates; (3) applying the Causal Reduction Theorem to reframe the problem in a form amenable to causal pathway decomposition; (4) estimating the effects efficiently. Finally, our approach is used to analyze the temporal evolution of racial disparities in outcome after admission to an intensive care unit (ICU).", "date": "2026-05-12", "categories": []}
{"id": "2605.12195", "title": "Fair Conformal Classification via Learning Representation-Based Groups", "abstract": "Conformal prediction methods provide statistically rigorous marginal coverage guarantees for machine learning models, but such guarantees fail to account for algorithmic biases, thereby undermining fairness and trust. This paper introduces a fair conformal inference framework for classification tasks. The proposed method constructs prediction sets that guarantee conditional coverage on adaptively identified subgroups, which can be implicitly defined through nonlinear feature combinations. By balancing effectiveness and efficiency in producing compact, informative prediction sets and ensuring adaptive equalized coverage across unfairly treated subgroups, our approach paves a practical pathway toward trustworthy machine learning. Extensive experiments on both synthetic and real-world datasets demonstrate the effectiveness of the framework.", "date": "2026-05-12", "categories": []}
{"id": "2605.12299", "title": "GKnow: Measuring the Entanglement of Gender Bias and Factual Gender", "abstract": "Recent works have analyzed the impact of individual components of neural networks on gendered predictions, often with a focus on mitigating gender bias. However, mechanistic interpretations of gender tend to (i) focus on a very specific gender-related task, such as gendered pronoun prediction, or (ii) fail to distinguish between the production of factually gendered outputs (the correct assumption of gender given a word that carries gender as a semantic property) and gender biased outputs (based on a stereotype). To address these issues, we curate \\gknow, a benchmark to assess gender knowledge and gender bias in language models across different types of gender-related predictions. \\gknow allows us to identify and analyze circuits and individual neurons responsible for gendered predictions. We test the impact of neuron ablation on benchmarks for disentangling stereotypical and factual gender (DiFair and the test set of GKnow), as well as StereoSet. Results show that gender bias and factual gender are severely entangled on the level of both circuits and neurons, entailing that ablation is an unreliable debiasing method. Furthermore, we show that benchmarks for evaluating gender bias can hide the decrease in factual gender knowledge that accompanies neuron ablation. We curate GKnow as a contribution to the continuous development of robust gender bias benchmarks.", "date": "2026-05-12", "categories": []}
{"id": "2312.08413", "title": "Privacy Constrained Fairness Estimation for Decision Trees", "abstract": "The protection of sensitive data becomes more vital, as data increases in value and potency. Furthermore, the pressure increases from regulators and society on model developers to make their Artificial Intelligence (AI) models non-discriminatory. To boot, there is a need for interpretable, transparent AI models for high-stakes tasks. In general, measuring the fairness of any AI model requires the sensitive attributes of the individuals in the dataset, thus raising privacy concerns. In this work, the trade-offs between fairness, privacy and interpretability are further explored. We specifically examine the Statistical Parity (SP) of Decision Trees (DTs) with Differential Privacy (DP), that are each popular methods in their respective subfield. We propose a novel method, dubbed Privacy-Aware Fairness Estimation of Rules (PAFER), that can estimate SP in a DP-aware manner for DTs. DP, making use of a third-party legal entity that securely holds this sensitive data, guarantees privacy by adding noise to the sensitive data. We experimentally compare several DP mechanisms. We show that using the Laplacian mechanism, the method is able to estimate SP with low error while guaranteeing the privacy of the individuals in the dataset with high certainty. We further show experimentally and theoretically that the method performs better for DTs that humans generally find easier to interpret.", "date": "2023-12-13", "categories": []}
{"id": "2506.21095", "title": "FeDa4Fair: Client-Level Federated Datasets for Fairness Evaluation", "abstract": "Federated Learning (FL) enables collaborative training while preserving privacy, yet it introduces a critical challenge: the \"illusion of fairness''. A global model, usually evaluated on the server, appears fair on average while keeping persistent discrimination at the client level. Current fairness-enhancing FL solutions often fall short, as they typically mitigate biases for a single, usually binary, sensitive attribute, while ignoring two realistic and conflicting scenarios: attribute-bias (where clients are unfair toward different sensitive attributes) and value-bias (where clients exhibit conflicting biases toward different values of the same attribute). To support more robust and reproducible fairness research in FL, we introduce FeDa4Fair, the first benchmarking framework designed to stress-test fairness methods under these heterogeneous conditions. Our contributions are three-fold: (1) We introduce FeDa4Fair, a library designed to create datasets tailored to evaluating fair FL methods under heterogeneous client bias; (2) we release a benchmark suite generated by the FeDa4Fair library to standardize the evaluation of fair FL methods; (3) we provide ready-to-use functions for evaluating fairness outcomes for these datasets.", "date": "2025-06-26", "categories": []}
{"id": "2605.08137", "title": "Weight Pruning Amplifies Bias: A Multi-Method Study of Compressed LLMs for Edge AI", "abstract": "Weight pruning is widely advocated for deploying Large Language Models on resource-constrained IoT and edge devices, yet its impact on model fairness remains poorly understood. We conduct a controlled empirical study of three instruction-tuned models (Gemma-2-9b-it, Mistral-7B-Instruct-v0.3, Phi-3.5-mini-instruct) across three pruning methods (Random, Magnitude, Wanda) at four sparsity levels (10-70%) on 12,148 BBQ bias benchmark items with 5 random seeds, totaling 2,368,860 inference records. Our results reveal a Smart Pruning Paradox: activation-aware pruning (Wanda) preserves perplexity nearly perfectly (just 3.5% increase at 50% sparsity for Mistral-7B), yet produces the highest bias amplification, with Stereotype Reliance Score increasing 83.7% and 47-59% of previously unbiased items developing new stereotypical behaviors at 70% sparsity. Random pruning destroys language capability entirely (perplexity exceeding $10^4$ and reaching $10^8$) but produces only random-chance bias. We further show that unstructured pruning provides zero storage savings and zero inference latency reduction on real edge hardware, undermining the primary motivation for its use in IoT deployment. Of 180 dense-vs-pruned comparisons, 141 (78.3%) are significant ($p < 0.05$) with mean $|h| = 0.305$. Published quantization studies report up to 21% of responses flipping between biased and unbiased states; our pruning results show transition rates nearly three times higher (47-59%), suggesting pruning poses a categorically greater risk to alignment than quantization. These findings demonstrate that perplexity-based evaluation provides false assurance of behavioral equivalence, and that IoT deployment pipelines require bias-aware validation before deploying pruned models at the edge.", "date": "2026-05-02", "categories": []}
{"id": "2605.09221", "title": "The Pok\\'emon Theorem and other Fairness Impossibility Results", "abstract": "Fairness impossibility results often look like distinct scalar incompatibility statements. We show that several share one RKHS geometry: fairness criteria are linear constraints on conditional mean embeddings, and unequal base rates make the law of total expectation overdetermine those constraints.\n  This view yields four results. The Kleinberg--Mullainathan--Raghavan dichotomy needs only group-conditional unbiasedness, not full calibration. The \\emph{Pok\\'emon theorem} shows that a distinct group pair satisfying any finite collection of linear mean-fairness criteria leaves a residual violation witnessed by the MMD, decaying at the Kolmogorov $m$-width rate under spectral regularity. The same tools prove an impossibility for fair feature learning: parity and class-conditional separation in representation space force class collapse under unequal base rates. The approximate relaxations yield signal and error frontiers, allowing a trade-off between real-world estimators and fairness goals. Experiments on standard fairness benchmarks are consistent with our bounds.", "date": "2026-05-09", "categories": []}
{"id": "2605.09852", "title": "Fairness of Explanations in Artificial Intelligence (AI): A Unifying Framework, Axioms, and Future Direction toward Responsible AI", "abstract": "Machine learning algorithms are being used in high-stakes decisions, including those in criminal justice, healthcare, credit, and employment. The research community has responded with two largely independent research fields: \\emph{algorithmic fairness}, which targets equitable outcomes, and \\emph{explainable AI} (XAI), which targets interpretable reasoning. This survey identifies and maps a novel blind spot at their intersection, which is a model that can satisfy every standard fairness criterion in its outputs while being profoundly unfair in its \\emph{reasoning process}. We refer to this as the procedural bias, and mitigating it requires treating the fairness of explanations as a distinct object of scientific study.\n  To our knowledge, we provide the first unified theoretical and literature review of this emerging field and elucidate the drawbacks of post-hoc explainers in certifying explanation fairness. Our central contribution is a \\emph{conditional invariance framework} formalizing explanation fairness as the requirement that explanations should be indifferent regardless of the protected attributes $ P(E(X) \\in \\cdot \\mid X_\\text{rel} = x_\\text{rel},\\, A = a) = P(E(X) \\in \\cdot \\mid X_\\text{rel} = x_\\text{rel},\\, A = b)$ for all task-relevant $x$, a single principle from which all existing explanation fairness metrics emerge as partial operationalizations. We introduce a seven-dimensional taxonomy, identify three generative mechanisms of explanation inequity (representation-driven, explanation-model mismatch, actionability-driven), and propose a canonical six-step evaluation workflow for operationalizing explanation fairness audits in practice.", "date": "2026-05-11", "categories": []}
{"id": "2605.09945", "title": "Selection of the Best Policy under Fairness Constraints for Subpopulations", "abstract": "Many high-stakes decisions in health care, public policy, and clinical development require committing to a single policy that will be applied uniformly across a heterogeneous population. Regulatory and fairness standards sometime requires that the chosen policy performs adequately in every pre-specified subpopulation, not only on average. We formalize this as a Selection of the Best with Fairness Constraints (SBFC) problem, in order to identify the policy with the highest average performance among those policies that meet a minimum per-subpopulation threshold. We establish an instance-specific lower bound on sample complexity of the SBFC problem. We then develop a Track-and-Stop with Constraints on Subpopulation (T-a-S-CS) algorithm that achieves the lower bound asymptotically. We extend the framework to general closed-set and penalty-based fairness specifications with matching guarantees. Numerical experiments and a case study using the International Stroke Trial demonstrate substantial efficiency gains over policy-level allocation baselines.", "date": "2026-05-11", "categories": []}
{"id": "2605.10604", "title": "Fairness vs Performance: Characterizing the Pareto Frontier of Algorithmic Decision Systems", "abstract": "Designing fair algorithmic decision systems requires balancing model performance with fairness toward affected individuals: More fairness might require sacrificing some performance and vice versa, yet the space of possible trade-offs is still poorly understood. We investigate fairness in binary prediction-based decision problems by conceptualizing decision making as a multi-objective optimization problem that simultaneously considers decision-maker utility and group fairness. We investigate the set of Pareto-optimal decision rules for arbitrary utility functions for decision maker, arbitrary population distributions, and a wide range of group fairness metrics.\n  We find that the Pareto frontier consists of deterministic, group-specific threshold rules applied to individuals' success probability. This complements existing optimality theorems from literature which, for specific fairness constraints, posit lower-bound threshold rules only. However we also show that, depending on the used fairness metric, the Pareto frontier may include upper-bound threshold rules, thus preferring individuals with lower success probabilities. We show that the location of the Pareto frontier depends only on population characteristics, utility functions and fairness score, but not on the technical design of the algorithm - our findings hold for pre-, in-, and post-processing approaches alike.\n  Our results generalize existing optimality theorems for fairness-constrained classification and extend them to generalized fairness metrics and fairness principles, and to partial fairness regimes. This paper connects formal fairness research with legal and ethical requirements to search for less discriminatory alternatives, offering a principled foundation for evaluating and comparing algorithmic decision systems.", "date": "2026-05-11", "categories": []}
{"id": "2605.10615", "title": "Responsible Benchmarking of Fairness for Automatic Speech Recognition", "abstract": "Many studies have shown automatic speech processing (ASR) systems have unequal performance across speakergroups (SG's). However, the manner in which such studies arrive at this conclusion is inconsistent. To pave the wayfor more reliable results in future studies, we lay out best practices for benchmarking ASR fairness based on literaturefrom machine learning fairness, social sciences, and speech science. We first describe the importance of preciselythe fairness hypothesis being interrogated, and tailoring fairness metrics to apply specifically to said hypothesis.We then examine several benchmarks used to rate ASR systems on fairness and discuss how their results can bemisconstrued without assiduous oversight into the intersections between SG's. We find that evaluating fairnessbased on single heterogeneous SG's, such as they are defined in fairness benchmarks, can lead to misidentifyingwhich SG's are actually being mistreated by ASR systems. We advocate for as fine-grained an analysis as possibleof the intersectionality of as many demographic variables as are available in the metadata of fairness corpora in orderto tease out such spurious correlations", "date": "2026-05-11", "categories": []}
{"id": "2605.06891", "title": "Towards Fairness under Label Bias in Image Segmentation: Impact, Measurement and Mitigation", "abstract": "Labeled datasets reflect the biases of their annotation pipelines, which sometimes introduce label bias: group-conditional label errors that cause systematic performance disparities across demographic subgroups. Label bias in image segmentation remains underexplored, as even detecting it typically requires clean, unbiased annotations, which are not readily available. We present a data-centric adaptation of Confident Learning to segmentation, allowing detection of label bias directly in the training data without a clean, unbiased ground truth. By comparing the provided training labels to the model's confident predictions, we isolate directional errors that quantify the presence and nature of bias, where standard overlap metrics like Dice fail. We further show that label bias influences subgroup separability in the encoder's feature space, an artifact we leverage for bias mitigation rather than suppressing it. We evaluate three datasets, spanning from synthetic to real-life bias, showing how our framework reliably detects and mitigates bias without access to clean labels, achieving equitable performance across experimental conditions.", "date": "2026-05-07", "categories": []}
{"id": "2506.06816", "title": "How do datasets, developers, and models affect biases in a low-resourced language?: The Case of the Bengali Language", "abstract": "Sociotechnical systems, such as language technologies, frequently exhibit identity-based biases. These biases exacerbate the experiences of historically marginalized communities and remain understudied in low-resource contexts. While models and datasets specific to a language or with multilingual support are commonly recommended to address these biases, this paper empirically tests the effectiveness of such approaches in the context of gender, religion, and nationality-based identities in Bengali, a widely spoken but low-resourced language. We conducted an algorithmic audit of sentiment analysis models built on mBERT and BanglaBERT, which were fine-tuned using all Bengali sentiment analysis (BSA) datasets from Google Dataset Search. Our analyses showed that BSA models exhibit biases across different identity categories despite having similar semantic content and structure. We also examined the inconsistencies and uncertainties arising from combining pre-trained models and datasets created by individuals from diverse demographic backgrounds. We connected these findings to the broader discussions on epistemic injustice, AI alignment, and methodological decisions in algorithmic audits.", "date": "2025-06-07", "categories": []}
{"id": "2605.05882", "title": "Tuning Derivatives for Causal Fairness in Machine Learning", "abstract": "Artificial-intelligence systems are becoming ubiquitous in society, yet their predictions typically inherit biases with respect to protected attributes such as race, gender, or age. Classical fairness notions, most notably Statistical Parity (SP), demand that predictions be independent of the protected attributes, but are overly restrictive when these attributes influence mediating variables that are considered business necessities. Recent causal formulations relax SP by distinguishing allowed from not-allowed causal paths and by complementing SP with Predictive Parity (PP), requiring the predictor to replicate the legitimate influence of business-necessities. Existing path-based definitions are mainly practical when applied to categorical attributes. This paper introduces a new framework for fairness in structural causal models that is tailored to continuous protected attributes. We formalize SP and PP through path-specific partial derivatives, establish conditions under which these criteria coincide with prior causal definitions, and characterize when a fair predictor, one that satisfies SP along not-allowed paths while achieving PP along allowed paths, exists. Building on this theory, we propose a fair tuning algorithm that either constructs such a predictor or, when not possible, allows for a trade-off between SP and PP. We present experiments on simulated and real data to evaluate our proposal, compare it with previously proposed methods, and show that it performs better when PP is considered.", "date": "2026-05-07", "categories": []}
{"id": "2605.06227", "title": "Price of Fairness in Short-Term and Long-Term Algorithmic Selections", "abstract": "Algorithmic decision-making in high-stakes settings can have profound impacts on individuals and populations. While much prior work studies fairness in static settings, recent results show that enforcing static fairness constraints may exacerbate long-run disparities. Motivated by this tension, we study a stylized sequential selection problem in which a decision-maker repeatedly selects individuals, affecting both immediate utility and the population distribution over time. We introduce notions of group fairness for both the short and long term and theoretically analyze the trade-off between fairness and utility via the Price of Fairness (PoF). We characterize optimal and fair policies in the short term and show that the PoF can be large even when group distributions are nearly identical. In contrast, we show that long-term disparities can vanish under simple investment policies that achieve a low PoF. We also empirically validate these theoretical observations using both synthetic and real datasets.", "date": "2026-05-07", "categories": []}
{"id": "2605.06324", "title": "Gaming the Metric, Not the Harm: Certifying Safety Audits against Strategic Platform Manipulation", "abstract": "Online-safety regulation under the UK Online Safety Act and the EU Digital Services Act increasingly treats scalar metrics as compliance evidence. Once announced, such a metric also becomes an optimization target: a strategic platform can improve its score by routing recommendations through semantically equivalent content variants, without reducing true harm. We ask when such an audit metric can still certify a genuine reduction in harm. The protocol is modeled as a published transformation graph whose connected components form semantic classes, and the metric itself is treated as a security object. Three results follow. First, any metric that scores variants directly is manipulable as soon as two equivalent variants in a harmful class disagree in score. Second, the semantic-envelope lift, which assigns each variant the maximum score in its class, is the unique pointwise minimum among conservative classwise-constant repairs. Third, a class-stratified certificate, $H^\\star(x) \\le (1/\\hat\\alpha) M_{\\mathrm{Env}(m)}(x) + \\bar\\eta$, holds for every platform strategy, with $\\bar\\eta$ absorbing annotation and protocol error. We check the claims at three levels: exhaustive enumeration on a finite-state grid of mixed strategies, an SMT encoding in Z3 cross-replayed in cvc5, and a bounded single-player MDP encoded in PRISM-games. The fragile metric fails manipulation invariance and cannot support the same useful predeclared class-coverage certificate; under the envelope-level certificate, it produces large violations at every tested instance, with a large mean gaming gap across random catalogs at a fixed audit budget. The semantic-envelope metric exhibits no such violation in the tested instances.", "date": "2026-05-07", "categories": []}
{"id": "2605.06479", "title": "Risk-Controlled Post-Processing of Decision Policies", "abstract": "Predictive models are often deployed through existing decision policies that stakeholders are reluctant to change unless a risk constraint requires intervention. We study risk-controlled post-processing: given a deterministic baseline policy, choose a new policy that maximizes agreement with the baseline subject to a chance constraint on a user-specified loss. At the population level, we show that the optimal policy has a threshold structure: it follows the baseline except on contexts where switching to the oracle fallback policy yields a large reduction in conditional violation risk. At the finite-sample level, given a fitted fallback policy and score, we develop a post-processing algorithm that uses calibration data to select a threshold. Leveraging tools from algorithmic stability and stochastic processes, we show that under regularity conditions, in the i.i.d. setting, the expected excess risk of the post-processed policy is $O(\\log n/n)$. In the special case when an exact-safe fallback policy is available, the algorithm achieves precise expected risk control under exchangeability. In this setting, we also give high-probability near-optimality guarantees on the post-processed policy. Experiments on a COVID-19 radiograph diagnosis task, an LLM routing problem, and a synthetic multiclass decision task show that targeted post-processing can meet or nearly meet risk budgets while preserving substantially more agreement with the baseline than score-blind random mixing.", "date": "2026-05-07", "categories": []}
{"id": "2605.04063", "title": "Investigating Trustworthiness of Nonparametric Deep Survival Models for Alzheimer's Disease Progression Analysis", "abstract": "Alzheimer's Dementia (AD) is a progressive neurodegenerative disease marked by irreversible decline, making reliable modeling of its progression essential for effective patient care. Progression-aware methods such as survival analysis are therefore crucial tools for the early detection and monitoring of AD. Recent advancements in deep learning have demonstrated remarkable performance in survival tasks, but alarmingly fewer studies have been conducted in the domain of AD. Further, the studies that do exist do not consider learned bias within the model itself, which could result in unfair and unreliable predictions toward certain marginalized groups. As such, we conduct a rigorous study of fairness in AD progression analysis along with a thorough feature importance study to determine the characteristics which are most important for reliable AD predictions. Furthermore, we propose two novel fairness metrics, called Time-Dependent Concordance Impurity and Kaplan-Meier Fairness, to quantify bias with respect to sensitive attributes such as sex, race, and education in nonparametric survival models. Our study demonstrates that while deep learning powered survival models are robust tools which can aid clinicians in AD care decisions, they often exhibit considerable bias, representing important avenues for future research.", "date": "2026-04-10", "categories": []}
{"id": "2502.04419", "title": "Understanding and Mitigating Bias Inheritance in LLM-based Data Augmentation on Downstream Tasks", "abstract": "Generating synthetic datasets via large language models (LLMs) has emerged as a promising approach to improve LLM performance. However, LLMs inherently reflect biases in their training data, leading to a critical challenge: when models are trained on synthetic data, they may propagate and amplify the inherent biases that can significantly impact fairness and robustness on downstream tasks-a phenomenon we term bias inheritance. This work presents the first systematic investigation in understanding, analyzing, and mitigating bias inheritance. We fine-tune LLMs with a combined dataset of real and LLM-augmented data with varied bias ratio as the proportion of augmented data. Through systematic experiments across 10 classification and generation tasks, we analyze how 6 different types of biases manifest. Our results indicate that bias inheritance harms downstream task performance in bias directly-related classification and generation tasks. Then, our analysis identifies three key misalignment factors: misalignment of values, group data, and data distributions. Based on these insights, we propose three mitigation strategies: token-based, mask-based, and loss-based approaches, which can work differently on various tasks and bias, indicating the substantial challenges to mitigate bias inheritance. We hope this work can provide insights to the research of LLM data augmentation.", "date": "2025-02-06", "categories": []}
{"id": "2605.00382", "title": "Social Bias in LLM-Generated Code: Benchmark and Mitigation", "abstract": "Large Language Models (LLMs) are increasingly deployed to generate code for human-centered applications where demographic fairness is critical. However, existing evaluations focus almost exclusively on functional correctness, leaving social bias in LLM-generated code largely unexamined. Extending our prior work on Solar, we conduct a comprehensive empirical study using SocialBias-Bench, a benchmark of 343 real-world coding tasks spanning seven demographic dimensions. We evaluate four prominent LLMs and find severe bias across all models, with Code Bias Scores reaching up to 60.58%. We further show that standard prompt-level interventions, such as Chain-of-Thought reasoning and fairness persona assignment, inadvertently amplify bias rather than reduce it. We then investigate whether structured multi-agent software process frameworks can improve fairness, finding that structured pipelines reduce bias when early roles correctly scope what the code should and should not consider. However, adding explicit fairness instructions to all agent roles produces worse outcomes than providing none, suggesting that diffused responsibility goes unaddressed. To address these limitations, we propose the Fairness Monitor Agent (FMA), a modular component that plugs into any existing code generation pipeline without modifying it. FMA analyzes the task description to determine which attributes should be considered or restricted, then detects and corrects violations through an iterative review process, without requiring an executable test suite. Evaluated on all 343 tasks, FMA reduces bias by 65.1% compared to a developer agent alone and improves functional correctness from 75.80% to 83.97%, outperforming all other studied approaches.", "date": "2026-05-01", "categories": []}
{"id": "2605.01451", "title": "Auditing demographic bias in AI-based emergency police dispatch: a cross-lingual evaluation of eleven large language models", "abstract": "Large language models (LLMs) are rapidly being integrated into high-stakes public safety systems, including emergency call triage and dispatch decision support, yet their demographic fairness in this context remains largely untested. Here we introduce a cross-lingual audit framework that operationalizes the Police Priority Dispatch System as a five-level ordinal classification task and applies a controlled minimal-pair design to isolate the effect of demographic cues. Across 19,800 model outputs spanning 11 frontier models, 15 scenario pairs, three demographic categories (religious appearance, gender, and race), and two languages (English and Mandarin Chinese), we find that demographic bias emerges systematically when incident severity is ambiguous but largely disappears when the operational priority is clearly determined by call content. Bias magnitude varies by demographic axis, with the largest effects observed for religious appearance, followed by gender and race. Critically, bias does not transfer consistently across languages: gender bias is substantially amplified in Mandarin Chinese, whereas race bias is more pronounced in English, revealing cross-lingual asymmetries that aggregate analyses obscure. In several scenarios, demographic cues produce counter-directional effects, challenging simple stereotype-amplification accounts of model behavior. These findings suggest that bias in LLM-based dispatch is not a fixed property of models alone, but arises from the interaction between demographic signals, contextual ambiguity, and language. Beyond these empirical results, the proposed framework provides a scalable audit infrastructure that enables deploying agencies to evaluate candidate models on jurisdiction-relevant scenarios prior to real-world adoption.", "date": "2026-05-02", "categories": []}
{"id": "2605.01971", "title": "ProtoFair: Fair Self-Supervised Contrastive Learning via Pseudo-Counterfactual Pairs", "abstract": "Self-supervised learning methods learn high-quality visual representations, yet recent studies show that these representations often capture demographic biases present in the training data. Existing fairness-aware methods address this by redesigning the self-supervised objective itself, limiting portability across the rapidly evolving landscape of self-supervised learning (SSL) frameworks. We propose ProtoFair, a fairness-aware contrastive loss designed to work alongside existing SSL objectives without modifying them. ProtoFair leverages unsupervised prototype clustering to identify pseudo-counterfactual pairs: samples sharing the same cluster assignment but belonging to different sensitive groups. By pulling these content-matched, cross-group samples together in the embedding space, ProtoFair encourages the encoder to learn representations that are invariant to the sensitive attribute. The method requires only sensitive attribute annotations, no target labels, and integrates seamlessly with both SimCLR and SupCon. Experiments on CelebA and UTKFace demonstrate consistent fairness improvements while maintaining competitive accuracy.", "date": "2026-05-03", "categories": []}
{"id": "2605.02348", "title": "Decoding-Time Debiasing via Process Reward Models: From Controlled Fill-in to Open-Ended Generation", "abstract": "Large language models pick up social biases from the data they are trained on and carry those biases into downstream applications, often reinforcing stereotypes around gender, race, religion, disability, age, and socioeconomic status. The standard fixes (retraining on curated data or fine-tuning with human feedback) are expensive, need access to model weights, and risk degrading the model on other tasks. In this paper we take a different route: we debias the model at decoding time, treating bias mitigation as a structured search over candidate tokens without ever touching model weights. A separate Process Reward Model (PRM) acts as a judge, scoring each candidate for both fairness and fluency. We design three schemes of increasing sophistication (Best-of-N selection, Sequential critique-and-revise, and Constitutional self-audit) and evaluate them on four models (GPT-4o-mini, Llama 3.2 3B, Gemma 3 4B, Qwen 2.5 3B) across a 200-prompt bilingual benchmark in English and Urdu covering eight bias categories. Sequential debiasing proves the most effective, raising mean bias scores by up to +0.40 over baseline while preserving (and sometimes improving) fluency. We then extend all three schemes to open-ended generation, where each token is debiased on the fly, and introduce a lightweight Bias Guard gate that fires only on potentially biased words, keeping overhead near 2x for well-calibrated models. A formal overhead metric that separates generator cost from judge cost reveals that Best-of-N is effectively free on the generator side in a native implementation. GPT-4o-mini, included as a strong proprietary anchor, confirms that the framework scales with model capability; the three open-weight models show where current small-scale LLMs still struggle.", "date": "2026-05-04", "categories": []}
{"id": "2404.07475", "title": "Laissez-Faire Harms: Algorithmic Biases in Generative Language Models", "abstract": "The rapid deployment of generative language models (LMs) has raised concerns about social biases affecting the well-being of diverse consumers. The extant literature on generative LMs has primarily examined bias via explicit identity prompting. However, prior research on bias in earlier language-based technology platforms, including search engines, has shown that discrimination can occur even when identity terms are not specified explicitly. Studies of bias in LM responses to open-ended prompts (where identity classifications are left unspecified) are lacking and have not yet been grounded in end-consumer harms. Here, we advance studies of generative LM bias by considering a broader set of natural use cases via open-ended prompting. In this \"laissez-faire\" setting, we find that synthetically generated texts from five of the most pervasive LMs (ChatGPT3.5, ChatGPT4, Claude2.0, Llama2, and PaLM2) perpetuate harms of omission, subordination, and stereotyping for minoritized individuals with intersectional race, gender, and/or sexual orientation identities (AI/AN, Asian, Black, Latine, MENA, NH/PI, Female, Non-binary, Queer). We find widespread evidence of bias to an extent that such individuals are hundreds to thousands of times more likely to encounter LM-generated outputs that portray their identities in a subordinated manner compared to representative or empowering portrayals. We also document a prevalence of stereotypes (e.g. perpetual foreigner) in LM-generated outputs that are known to trigger psychological harms that disproportionately affect minoritized individuals. These include stereotype threat, which leads to impaired cognitive performance and increased negative self-perception. Our findings highlight the urgent need to protect consumers from discriminatory harms caused by language models and invest in critical AI education programs tailored towards empowering diverse consumers.", "date": "2024-04-11", "categories": []}
{"id": "2405.13693", "title": "Mutatis Mutandis: Revisiting the Comparator in Discrimination Testing", "abstract": "Testing for individual discrimination involves deriving a profile, the comparator, similar to the one making the discrimination claim, the complainant, based on a protected attribute, such as race or gender, and comparing their decision outcomes. The complainant-comparator pair is central to discrimination testing. Most discrimination testing tools rely on this pair to establish evidence for discrimination. In this work, we revisit the role of the comparator in discrimination testing. We first argue for the inherent causal modeling nature of deriving the comparator. We then introduce a two-kind classification for the comparator: the ceteris paribus, or \"with all else equal,\" (CP) comparator and the mutatis mutandis, or \"with the appropriate adjustments being made,\" (MM) comparator. The CP comparator is the standard comparator, representing an idealized comparison for establishing discrimination as it aims for a complainant-comparator pair that only differs in membership in the protected attribute. As an alternative to the CP comparator, we define the MM comparator, which requires a comparator that represents the ``what would have been'' of the complainant without the effects of the protected attribute on the non-protected attributes. Under the MM comparator, the complainant-comparator pair can be dissimilar in terms of the non-protected attributes, departing from the idealized comparison imposed by the CP comparator. Notably, the MM comparator denotes a more complex object and its implementation offers an impactful venue for machine learning methods. We illustrate these two comparators and their impact on discrimination testing using a real-world example.", "date": "2024-05-22", "categories": []}
{"id": "2411.10915", "title": "Bias in Large Language Models: Origin, Evaluation, and Mitigation", "abstract": "Large Language Models (LLMs) have revolutionized natural language processing, but their susceptibility to biases poses significant challenges. This comprehensive review examines the landscape of bias in LLMs, from its origins to current mitigation strategies. We categorize biases as intrinsic and extrinsic, analyzing their manifestations in various NLP tasks. The review critically assesses a range of bias evaluation methods, including data-level, model-level, and output-level approaches, providing researchers with a robust toolkit for bias detection. We further explore mitigation strategies, categorizing them into pre-model, intra-model, and post-model techniques, highlighting their effectiveness and limitations. Ethical and legal implications of biased LLMs are discussed, emphasizing potential harms in real-world applications such as healthcare and criminal justice. By synthesizing current knowledge on bias in LLMs, this review contributes to the ongoing effort to develop fair and responsible AI systems. Our work serves as a comprehensive resource for researchers and practitioners working towards understanding, evaluating, and mitigating bias in LLMs, fostering the development of more equitable AI technologies.", "date": "2024-11-16", "categories": []}
{"id": "2509.06864", "title": "Concolic Testing on Individual Fairness of Neural Network Models", "abstract": "This paper introduces PyFair, a formal framework for evaluating and verifying individual fairness of Deep Neural Networks (DNNs). By adapting the concolic testing tool PyCT, we generate fairness-specific path constraints to systematically explore DNN behaviors. Our key innovation is a dual network architecture that enables comprehensive fairness assessments and provides completeness guarantees for certain network types. We evaluate PyFair on 25 benchmark models, including those enhanced by existing bias mitigation techniques. Results demonstrate PyFair's efficacy in detecting discriminatory instances and verifying fairness, while also revealing scalability challenges for complex models. This work advances algorithmic fairness in critical domains by offering a rigorous, systematic method for fairness testing and verification of pre-trained DNNs.", "date": "2025-09-08", "categories": []}
{"id": "2601.09896", "title": "The Algorithmic Gaze of Image Quality Assessment: An Audit and Trace Ethnography of the LAION-Aesthetics Predictor", "abstract": "Visual generative AI models are trained using a one-size-fits-all measure of aesthetic appeal. However, what is deemed \"aesthetic\" is inextricably linked to personal taste and cultural values, raising the question of whose taste is represented in visual generative AI models. In this work, we study an aesthetic evaluation model--LAION-Aesthetics Predictor (LAP)--that is widely used to curate datasets to train visual generative image models, like Stable Diffusion, and evaluate the quality of AI-generated images. To understand what LAP measures, we audited the model across three datasets. First, we examined the impact of aesthetic filtering on the LAION-Aesthetics Dataset (approximately 1.2B images), which was curated from LAION-5B using LAP. We find that the LAP disproportionally filters in images with captions mentioning women, while filtering out images with captions mentioning men or LGBTQ+ people. Then, we used LAP to score approximately 330k images across two art datasets, finding the model rates realistic images of landscapes, cityscapes, and portraits from western and Japanese artists most highly. In doing so, the algorithmic gaze of this aesthetic evaluation model reinforces the imperial and male gazes found within western art history. In order to understand where these biases may have originated, we performed a digital ethnography of public materials related to the creation of LAP. We find that the development of LAP reflects the biases we found in our audits, such as the aesthetic scores used to train LAP primarily coming from English-speaking photographers and western AI-enthusiasts. In response, we discuss how aesthetic evaluation can perpetuate representational harms and call on AI developers to shift away from prescriptive measures of \"aesthetics\" toward more pluralistic evaluation.", "date": "2026-01-14", "categories": []}
{"id": "2605.00592", "title": "Fairness of Classifiers in the Presence of Constraints between Features", "abstract": "In Machine Learning, an accepted definition of fairness of a decision taken by a classifier is that it should not depend on protected features, such as gender. Unfortunately, when constraints exist between features, such dependencies can be obscured by the constraints.\n  To avoid this problem, we propose that a decision be considered fair if it has a fair explanation. We define a fair explanation as a prime-implicant reason for the decision that does not contain any protected feature (where the constraints are taken into account in the definition of prime-implicant). Surprisingly, ignoring constraints can completely change the fairness of a decision (according to this definition) even in the absence of constraints between protected and unprotected features.\n  Three possible definitions of fairness of a classifier are that for all its decisions (1) there are only fair explanations, (2) there is at least one fair explanation, or (3) changing protected features does not change the outcome. We identify the relationships between these different definitions of fairness and study the computational complexity of testing fairness of classifiers.", "date": "2026-05-01", "categories": []}
{"id": "2605.00762", "title": "Meritocratic Fairness in Budgeted Combinatorial Multi-armed Bandits via Shapley Values", "abstract": "We propose a new framework for meritocratic fairness in budgeted combinatorial multi-armed bandits with full-bandit feedback (BCMAB-FBF). Unlike semi-bandit feedback, the contribution of individual arms is not received in full-bandit feedback, making the setting significantly more challenging. To compute arm contributions in BCMAB-FBF, we first extend the Shapley value, a classical solution concept from cooperative game theory, to the $K$-Shapley value, which captures the marginal contribution of an agent restricted to a set of size at most $K$. We show that $K$-Shapley value is a unique solution concept that satisfies Symmetry, Linearity, Null player, and efficiency properties. We next propose K-SVFair-FBF, a fairness-aware bandit algorithm that adaptively estimates $K$-Shapley value with unknown valuation function. Unlike standard bandit literature on full bandit feedback, K-SVFair-FBF not only learns the valuation function under full feedback setting but also mitigates the noise arising from Monte Carlo approximations. Theoretically, we prove that K-SVFair-FBF achieves $O(T^{3/4})$ regret bound on fairness regret. Through experiments on federated learning and social influence maximization datasets, we demonstrate that our approach achieves fairness and performs more effectively than existing baselines.", "date": "2026-05-01", "categories": []}
{"id": "2503.07878", "title": "A Woman with a Knife or A Knife with a Woman? Measuring Directional Bias Amplification in Image Captions", "abstract": "When we train models on biased datasets, they not only reproduce data biases, but can worsen them at test time - a phenomenon called bias amplification. Many of the current bias amplification metrics (e.g., BA (MALS), DPA) measure bias amplification only in classification datasets. These metrics are ineffective for image captioning datasets, as they cannot capture the language semantics of a caption. Recent work introduced Leakage in Captioning (LIC), a language-aware bias amplification metric that understands caption semantics. However, LIC has a crucial limitation: it cannot identify the source of bias amplification in captioning models. We propose Directional Bias Amplification in Captioning (DBAC), a language-aware and directional metric that can identify when captioning models amplify biases. DBAC has two more improvements over LIC: (1) it is less sensitive to sentence encoders (a hyperparameter in language-aware metrics), and (2) it provides a more accurate estimate of bias amplification in captions. Our experiments on gender and race attributes in the COCO captions dataset show that DBAC is the only reliable metric to measure bias amplification in captions.", "date": "2025-03-10", "categories": []}
{"id": "2604.28030", "title": "MIFair: A Mutual-Information Framework for Intersectionality and Multiclass Fairness", "abstract": "Fairness in machine learning remains challenging due to its ethical complexity, the absence of a universal definition, and the need for context-specific bias metrics. Existing methods still struggle with intersectionality, multiclass settings, and limited flexibility and generality. To address these gaps, we introduce MIFair, a unified framework for bias assessment and mitigation based on mutual information. MIFair provides a flexible metric template and an in-processing mitigation method inspired by the Prejudice Remover, defining group fairness as statistical independence between prediction-derived variables and sensitive attributes. We further strengthen its information-theoretic foundation by establishing equivalences with widely used fairness notions such as independence and separation. MIFair naturally supports intersectionality, complex subgroup structures, and multiclass classification and employs regularization-based training to reduce bias according to the selected metric. Its key advantage is its versatility: it consolidates diverse fairness requirements into a single coherent framework, enabling consistent benchmarking and simplifying practical use. Experiments on real-world tabular and image datasets show that MIFair effectively reduces bias, including previously unaddressed multi-attribute scenarios, while maintaining strong predictive performance across the evaluated settings.", "date": "2026-04-30", "categories": []}
{"id": "2604.25932", "title": "Sociodemographic Biases in Educational Counselling by Large Language Models", "abstract": "As Large Language Models (LLMs) are increasingly integrated into educational settings, understanding their potential biases is critical. This study examines sociodemographic biases in LLM-based educational counselling. We evaluate responses from six LLMs answering questions about 900 vignettes describing students in diverse circumstances. Each vignette is systematically tested across 14 sociodemographic identifiers - spanning race and gender, socioeconomic status, and immigrant background - along with a control condition, yielding 243,000 model responses. Our findings indicate that (1) all models exhibit measurable biases, (2) bias patterns partially align with documented human biases but diverge in notable ways, (3) the magnitude of these biases is strongly influenced by the precision of the student descriptions, where vague or minimal information amplifies disparities nearly threefold, while concrete, individualised metrics substantially reduce them, and (4) bias profiles vary substantially across models. These results demonstrate the importance of context-rich and personalised educational representations, suggesting that AI-driven educational decisions benefit from detailed student-specific information to promote fairness and equity.", "date": "2026-04-03", "categories": []}
{"id": "2604.26188", "title": "Efficient and Interpretable Transformer for Counterfactual Fairness", "abstract": "The growing reliance of machine learning models in high-stakes, highly regulated domains such as finance and insurance has created a growing tension between predictive performance, interpretability, and regulatory fairness requirements. In these settings, models are expected not only to deliver reliable predictions but also to provide transparent decision rationales and comply with strict fairness requirements. Attention-based transformers offer powerful mechanisms for modeling complex data relationships as demonstrated in various language tasks, yet their attention mechanisms alone do not ensure counterfactually fair predictions, even when combined with fairness-aware techniques. To address these limitations, we propose the Feature Correlation Transformer (FCorrTransformer), an attention-light architecture tailored for tabular data. In this design, the attention matrix admits a direct statistical interpretation as pairwise feature dependencies, enhancing both interpretability and efficiency. Leveraging this structure, we introduce Counterfactual Attention Regularization (CAR), a framework that enforces group-invariant fair representations of sensitive features at the attention level, promoting counterfactually fair predictions without relying on explicit causal assumptions. Empirical evaluations on imbalanced classification and regression benchmarks demonstrate that FCorrTransformer combined with CAR achieves strong counterfactual fairness while maintaining competitive predictive performance and substantially reducing model complexity compared with standard transformer-based baselines. Overall, this work bridges a critical gap between fairness theory and machine learning models, offering a practical framework for responsible AI in regulatory-sensitive domains.", "date": "2026-04-29", "categories": []}
{"id": "2409.13869", "title": "Generative AI Carries Non-Democratic Biases and Stereotypes: Representation of Women, Black Individuals, Age Groups, and People with Disability in AI-Generated Images across Occupations", "abstract": "In this study, I investigate how generative artificial intelligence (AI) systems reproduce and reinforce societal biases, with a specific focus on the representation of women, Black individuals, age groups, and people with visible disabilities in AI-generated occupational images. I analyzed 444 images generated by Microsoft Designer, Meta AI, and Ideogram across 37 occupations and found significant disparities in representation. Women are underrepresented in senior and technology roles, Black individuals are nearly absent, and people with visible disabilities are completely absent across all categories. I also observed clear age bias, with younger individuals predominantly depicted. These patterns suggest that generative AI tools replicate, and in some cases amplify, existing workplace inequalities and stereotypes, undermining democratic values of equity and inclusion. My findings highlight the urgent need for algorithmic diversity exposure, and I recommend that AI developers and corporate users audit their tools for equity, diversity, and inclusion (EDI) risks. I argue for the critical inclusion of diverse groups in AI development and governance to foster more democratic and socially responsible technologies.", "date": "2024-09-20", "categories": []}
{"id": "2509.13400", "title": "Justice in Judgment: Unveiling (Hidden) Bias in LLM-assisted Peer Reviews", "abstract": "The adoption of large language models (LLMs) is transforming the peer review process, from assisting reviewers in writing detailed evaluations to generating entire reviews automatically. While these capabilities offer new opportunities, they also raise concerns about fairness and reliability. In this paper, we investigate bias in LLM-generated peer reviews through controlled interventions on author metadata, including affiliation, gender, seniority, and publication history. Our analysis consistently shows a strong affiliation bias favoring authors from highly ranked institutions. We also identify directional preferences associated with seniority and prior publication record, which can influence acceptance decisions for borderline papers. Gender effects are smaller but present in several models. Notably, implicit biases become more pronounced when examining token-level soft ratings, suggesting that alignment may mask but not fully eliminate underlying preferences", "date": "2025-09-16", "categories": []}
{"id": "2511.14135", "title": "AdaFair-MARL: Enforcing Adaptive Fairness Constraints in Multi-Agent Reinforcement Learning", "abstract": "Fair workload enforcement in heterogeneous multi-agent systems that pursue shared objectives remains challenging. Fixed fairness penalties often introduce inefficiencies, training instability, and conflicting agent incentives. Reward-shaping approaches in fair Multi-Agent Reinforcement Learning (MARL) typically incorporate fairness through heuristic penalties or scalar reward modifications and often rely on post-hoc evaluation. However, these methods do not guarantee that a desired fairness level will be satisfied. To address this limitation, we propose the Adaptive Fairness Multi-Agent Reinforcement Learning (AdaFair-MARL) framework, which formulates workload fairness as an explicit constraint so that agents maintain balanced contributions while optimizing team performance. We present AdaFair-MARL, a constrained cooperative MARL framework whose core algorithmic component is a primal-dual update that enforces workload fairness via adaptive Lagrange multiplier updates. Grounding the framework in a cooperative Markov game, we derive the fairness constraint from Jain's Fairness Index (JFI) geometry and show that the resulting feasible set admits a second-order cone representation, enabling principled Lagrangian dual-ascent updates without manual penalty tuning. Experiments in a simulated hospital coordination environment (MARLHospital) demonstrate the effectiveness of AdaFair-MARL compared to reward-shaping and fixed-penalty fairness methods, improving workload balance while maintaining team performance. We found that AdaFair-MARL achieves nearly perfect constraint satisfaction (0.99-1.00) while significantly improving workload fairness compared to fixed-penalty baselines.", "date": "2025-11-18", "categories": []}
{"id": "2601.18572", "title": "One Persona, Many Cues, Different Results: How Sociodemographic Cues Impact LLM Personalization", "abstract": "Personalization of LLMs by sociodemographic subgroup often improves user experience, but can also introduce or amplify biases and unfair outcomes across groups. Prior work has employed so-called personas, sociodemographic user attributes conveyed to a model, to study bias in LLMs by relying on a single cue to prompt a persona, such as user names or explicit attribute mentions. This disregards LLM sensitivity to prompt variation and the rarity of some cues in real interactions (external validity). We compare six commonly used persona cues across seven open and proprietary LLMs on four writing and advice tasks. While cues are overall highly correlated, they produce substantial variance in responses across personas that can change findings on persona-induced differences and bias. We therefore caution against claims based on single persona cues, especially when they are overly explicit and have low external validity.", "date": "2026-01-26", "categories": []}
{"id": "2604.22631", "title": "Identifying and typifying demographic unfairness in phoneme-level embeddings of self-supervised speech recognition models", "abstract": "Modern automatic speech recognition (ASR) systems have been observed to function better for certain speaker groups (SGs) than others, despite recent gains in overall performance. One potential impediment to progress towards fairer ASR is a more nuanced understanding of the types of modeling errors that speech encoder models make, and in particular the difference between the structure of embeddings for high-performance and low-performance SGs. This paper proposes a framework typifying two types of error that can occur in modeling phonemes in ASR systems: random error/high variance in phoneme embedding, vs systematic error/embedding bias. We find that training phoneme classification probes only on a single, typically disadvantaged SG, sometimes improves performance for that SG, which is evidence for the existence of SG-level bias in phoneme embeddings. On the other hand, we find that speakers and SGs with higher levels of phoneme variance are the same as those with worse phoneme prediction accuracy. We conclude that both types of error are present in phoneme embeddings and both are candidate causes for SG-level unfairness in ASR, though random error is likely a greater hindrance to fairness than systematic error. Furthermore, we find that finetuning encoder models using a fairness-enhancing algorithm (domain enhancing and adversarial training) changes neither the benefits of in-domain phoneme classification probe training, nor measured levels of random embedding error.", "date": "2026-04-24", "categories": []}
{"id": "2309.07176", "title": "Mind the Gap: Optimal and Equitable Encouragement Policies", "abstract": "In consequential domains, it is often impossible to compel individuals to take treatment, so that optimal policy rules are merely suggestions in the presence of human non-adherence to treatment recommendations. We study personalized decision problems in which the planner controls recommendations into treatment rather than treatment itself. Under a covariate-conditional no-direct-effect model of encouragement, policy value depends on two distinct objects: responsiveness to encouragement and treatment efficacy. This modeling distinction makes induced treatment take-up, rather than recommendation rates alone, the natural fairness target and yields tractable policy characterizations under budget and access constraints. In settings with deterministic algorithmic recommendations, the same model localizes overlap-robustness to the recommendation-response model rather than the downstream outcome model. We illustrate the methods in case studies based on data from reminders of SNAP benefits recertification, and from pretrial supervised release with electronic monitoring. While the specific remedy to inequities in algorithmic allocation is context-specific, it requires studying both take-up of decisions and downstream outcomes of them.", "date": "2023-09-12", "categories": []}
{"id": "2412.16641", "title": "A Systems Thinking Approach to Algorithmic Fairness", "abstract": "Systems thinking provides us with a way to model the algorithmic fairness problem by allowing us to encode prior knowledge and assumptions about where we believe bias might exist in the data generating process. We can then encode these beliefs as a series of causal graphs, enabling us to link AI/ML systems to politics and the law. This allows us to combine techniques from machine learning, causal inference, and system dynamics in order to capture different emergent aspects of the fairness problem. We can use systems thinking to help policymakers on both sides of the political aisle to understand the complex trade-offs that exist from different types of fairness policies, providing a sociotechnical foundation for designing AI policy that is aligned to their political agendas and with society's shared democratic values.", "date": "2024-12-21", "categories": []}
{"id": "2510.18914", "title": "Fairness Evaluation and Inference Level Mitigation in LLMs", "abstract": "Large language models often display undesirable behaviors embedded in their internal representations, undermining fairness, inconsistency drift, amplification of harmful content, and the propagation of unwanted patterns during extended dialogue and conversations. Although training-time or data-centric methods attempt to reduce these effects, they are computationally expensive, irreversible once deployed, and slow to adapt to new conversational contexts. Pruning-based methods provide a flexible and transparent way to reduce bias by adjusting the neurons responsible for certain behaviors. However, most existing approaches are static; once a neuron is removed, the model loses the ability to adapt when the conversation or context changes. To address this, we propose a dynamic, reversible, pruning-based framework that detects context-aware neuron activations and applies adaptive masking to modulate their influence during generation. Our inference-time solution provides fine-grained, memory-aware mitigation with knowledge-preserved, more coherent behavior across multilingual single- and multi-turn dialogues, enabling dynamic fairness control in real-world conversational AI.", "date": "2025-10-21", "categories": []}
{"id": "2512.09292", "title": "Identifying Bias in Machine-generated Text Detection", "abstract": "The meteoric rise in text generation capability has been accompanied by parallel growth in interest in machine-generated text detection: the capability to identify whether a given text was generated using a model or written by a person. While detection models show strong performance, they have the capacity to cause significant negative impacts. We explore potential biases in English machine-generated text detection systems. We curate a dataset of student essays and assess 16 different detection systems for bias across four attributes: gender, race/ethnicity, English-language learner (ELL) status, and economic status. We evaluate these attributes using regression-based models to determine the significance and power of the effects, as well as performing subgroup analysis. We find that while biases are generally inconsistent across systems, there are several key issues: several models tend to classify disadvantaged groups as machine-generated, ELL essays are more likely to be classified as machine-generated, economically disadvantaged students' essays are less likely to be classified as machine-generated, and non-White ELL essays are disproportionately classified as machine-generated relative to their White counterparts. Finally, we perform human annotation and find that while humans perform generally poorly at the detection task, they show no significant biases on the studied attributes.", "date": "2025-12-10", "categories": []}
{"id": "2604.20677", "title": "Intersectional Fairness in Large Language Models", "abstract": "Large Language Models (LLMs) are increasingly deployed in socially sensitive settings, raising concerns about fairness and biases, particularly across intersectional demographic attributes. In this paper, we systematically evaluate intersectional fairness in six LLMs using ambiguous and disambiguated contexts from two benchmark datasets. We assess LLM behavior using bias scores, subgroup fairness metrics, accuracy, and consistency through multi-run analysis across contexts and negative and non-negative question polarities. Our results show that while modern LLMs generally perform well in ambiguous contexts, this limits the informativeness of fairness metrics due to sparse non-unknown predictions. In disambiguated contexts, LLM accuracy is influenced by stereotype alignment, with models being more accurate when the correct answer reinforces a stereotype than when it contradicts it. This pattern is especially pronounced in race-gender intersections, where directional bias toward stereotypes is stronger. Subgroup fairness metrics further indicate that, despite low observed disparity in some cases, outcome distributions remain uneven across intersectional groups. Across repeated runs, responses also vary in consistency, including stereotype-aligned responses. Overall, our findings show that apparent model competence is partly associated with stereotype-consistent cues, and no evaluated LLM achieves consistently reliable or fair behavior across intersectional settings. These findings highlight the need for evaluation beyond accuracy, emphasizing the importance of combining bias, subgroup fairness, and consistency metrics across intersectional groups, contexts, and repeated runs.", "date": "2026-04-22", "categories": []}
{"id": "2604.21036", "title": "Who Defines Fairness? Target-Based Prompting for Demographic Representation in Generative Models", "abstract": "Text-to-image(T2I) models like Stable Diffusion and DALL-E have made generative AI widely accessible, yet recent studies reveal that these systems often replicate societal biases, particularly in how they depict demographic groups across professions. Prompts such as 'doctor' or 'CEO' frequently yield lighter-skinned outputs, while lower-status roles like 'janitor' show more diversity, reinforcing stereotypes. Existing mitigation methods typically require retraining or curated datasets, making them inaccessible to most users. We propose a lightweight, inference-time framework that mitigates representational bias through prompt-level intervention without modifying the underlying model. Instead of assuming a single definition of fairness, our approach allows users to select among multiple fairness specifications-ranging from simple choices such as a uniform distribution to more complex definitions informed by a large language model(LLM) that cites sources and provides confidence estimates. These distributions guide the construction of demographic specific prompt variants in the corresponding proportions, and we evaluate alignment by auditing adherence to the declared target and measuring the resulting skin tone distribution rather than assuming uniformity as 'fairness'. Across 36 prompts spanning 30 occupations and 6 non-occupational contexts, our method shifts observed skin-tone outcomes in directions consistent with the declared target, and reduces deviation from targets when the target is defined directly in skin-tone space(fallback). This work demonstrates how fairness interventions can be made transparent, controllable, and usable at inference time, directly empowering users of generative AI.", "date": "2026-04-22", "categories": []}
{"id": "2604.21309", "title": "When Bigger Isn't Better: A Comprehensive Fairness Evaluation of Political Bias in Multi-News Summarisation", "abstract": "Multi-document news summarisation systems are increasingly adopted for their convenience in processing vast daily news content, making fairness across diverse political perspectives critical. However, these systems can exhibit political bias through unequal representation of viewpoints, disproportionate emphasis on certain perspectives, and systematic underrepresentation of minority voices. This study presents a comprehensive evaluation of such bias in multi-document news summarisation using FairNews, a dataset of complete news articles with political orientation labels, examining how large language models (LLMs) handle sources with varying political leanings across 13 models and five fairness metrics. We investigate both baseline model performance and effectiveness of various debiasing interventions, including prompt-based and judge-based approaches. Our findings challenge the assumption that larger models yield fairer outputs, as mid-sized variants consistently outperform their larger counterparts, offering the best balance of fairness and efficiency. Prompt-based debiasing proves highly model dependent, while entity sentiment emerges as the most stubborn fairness dimension, resisting all intervention strategies tested. These results demonstrate that fairness in multi-document news summarisation requires multi-dimensional evaluation frameworks and targeted, architecture-aware debiasing rather than simply scaling up.", "date": "2026-04-23", "categories": []}
{"id": "2604.21420", "title": "FairQE: Multi-Agent Framework for Mitigating Gender Bias in Translation Quality Estimation", "abstract": "Quality Estimation (QE) aims to assess machine translation quality without reference translations, but recent studies have shown that existing QE models exhibit systematic gender bias. In particular, they tend to favor masculine realizations in gender-ambiguous contexts and may assign higher scores to gender-misaligned translations even when gender is explicitly specified. To address these issues, we propose FairQE, a multi-agent-based, fairness-aware QE framework that mitigates gender bias in both gender-ambiguous and gender-explicit scenarios. FairQE detects gender cues, generates gender-flipped translation variants, and combines conventional QE scores with LLM-based bias-mitigating reasoning through a dynamic bias-aware aggregation mechanism. This design preserves the strengths of existing QE models while calibrating their gender-related biases in a plug-and-play manner. Extensive experiments across multiple gender bias evaluation settings demonstrate that FairQE consistently improves gender fairness over strong QE baselines. Moreover, under MQM-based meta-evaluation following the WMT 2023 Metrics Shared Task, FairQE achieves competitive or improved general QE performance. These results show that gender bias in QE can be effectively mitigated without sacrificing evaluation accuracy, enabling fairer and more reliable translation evaluation.", "date": "2026-04-23", "categories": []}
{"id": "2604.21711", "title": "Fairness under uncertainty in sequential decisions", "abstract": "Fair machine learning (ML) methods help identify and mitigate the risk that algorithms encode or automate social injustices. Algorithmic approaches alone cannot resolve structural inequalities, but they can support socio-technical decision systems by surfacing discriminatory biases, clarifying trade-offs, and enabling governance. Although fairness is well studied in supervised learning, many real ML applications are online and sequential, with prior decisions informing future ones. Each decision is taken under uncertainty due to unobserved counterfactuals and finite samples, with dire consequences for under-represented groups, systematically under-observed due to historical exclusion and selective feedback. A bank cannot know whether a denied loan would have been repaid, and may have less data on marginalized populations.\n  This paper introduces a taxonomy of uncertainty in sequential decision-making -- model, feedback, and prediction uncertainty -- providing shared vocabulary for assessing systems where uncertainty is unevenly distributed across groups. We formalize model and feedback uncertainty via counterfactual logic and reinforcement learning, and illustrate harms to decision makers (unrealized gains/losses) and subjects (compounding exclusion, reduced access) of policies that ignore the unobserved space. Algorithmic examples show it is possible to reduce outcome variance for disadvantaged groups while preserving institutional objectives (e.g. expected utility). Experiments on data simulated with varying bias show how unequal uncertainty and selective feedback produce disparities, and how uncertainty-aware exploration alters fairness metrics. The framework equips practitioners to diagnose, audit, and govern fairness risks. Where uncertainty drives unfairness rather than incidental noise, accounting for it is essential to fair and effective decision-making.", "date": "2026-04-23", "categories": []}
{"id": "2407.17395", "title": "The Costs of Pretending That There Are Data-Generating Probability Distributions in the Social World", "abstract": "Machine Learning research, including work promoting fair or equitable algorithms, often relies on the concept of a data-generating probability distribution. The standard presumption is that since data points are 'sampled from' such a distribution, one can learn from observed data about this distribution and, thus, predict future data points which are also drawn from it. We argue, however, that such true probability distributions do not exist and that the rhetoric around them is harmful in social settings. We show that alternative frameworks focusing directly on relevant populations rather than abstract distributions are available and leave classical learning theory almost unchanged. Furthermore, we argue that the assumption of true probabilities or data-generating distributions can be misleading and obscure both the choices made and the goals pursued in machine learning practice. Based on these considerations, we suggest avoiding the assumption of data-generating probability distributions in the social world.", "date": "2024-07-24", "categories": []}
{"id": "2411.00585", "title": "Fairness Testing of Large Language Models in Role-Playing", "abstract": "Large Language Models (LLMs) have become foundational in modern language-driven software applications, profoundly influencing daily life. A critical technique in leveraging their potential is role-playing, where LLMs simulate diverse roles to enhance their real-world utility. However, while research has highlighted the presence of social biases in LLM outputs, it remains unclear whether and to what extent these biases emerge during role-playing scenarios. In this paper, we conduct an empirical study on fairness testing of LLMs in role-playing scenarios. To enable this testing, we use LLMs to generate 550 social roles spanning a comprehensive set of 11 demographic attributes, producing 33,000 role-specific questions that target various forms of bias. These questions, covering Yes/No, multiple-choice, and open-ended formats, are designed to prompt LLMs to adopt specific roles and respond accordingly. We employ a combination of rule-based and LLM-based strategies to identify biased responses, rigorously validated through human evaluation. Using the generated questions as the test cases, we conduct extensive evaluations of 10 advanced LLMs. The evaluation reveal 107,580 biased responses across the studied LLMs, with individual models yielding between 7,579 and 16,963 biased responses, underscoring the prevalence of bias in role-playing contexts. To support future research, we have publicly released the dataset, along with all scripts and experimental results.", "date": "2024-11-01", "categories": []}
{"id": "2604.18644", "title": "FASE : A Fairness-Aware Spatiotemporal Event Graph Framework for Predictive Policing", "abstract": "Predictive policing systems that allocate patrol resources based solely on predicted crime risk can unintentionally amplify racial disparities through feedback driven data bias. We present FASE, a Fairness Aware Spatiotemporal Event Graph framework, which integrates spatiotemporal crime prediction with fairness constrained patrol allocation and a closed loop deployment feedback simulator.\n  We model Baltimore as a graph of 25 ZIP Code Tabulation Areas and use 139,982 Part 1 crime incidents from 2017 to 2019 at hourly resolution, producing a sparse feature tensor. The prediction module combines a spatiotemporal graph neural network with a multivariate Hawkes process to capture spatial dependencies and self exciting temporal dynamics. Outputs are modeled using a Zero Inflated Negative Binomial distribution, suitable for overdispersed and zero heavy crime counts. The model achieves a validation loss of 0.4800 and a test loss of 0.4857.\n  Patrol allocation is formulated as a fairness constrained linear optimization problem that maximizes risk weighted coverage while enforcing a Demographic Impact Ratio constraint with deviation bounded by 0.05. Across six simulated deployment cycles, fairness remains within 0.9928 to 1.0262, and coverage ranges from 0.876 to 0.936. However, a persistent detection rate gap of approximately 3.5 percentage points remains between minority and non minority areas. This result shows that allocation level fairness constraints alone do not eliminate feedback induced bias in retraining data, highlighting the need for fairness interventions across the full pipeline.", "date": "2026-04-19", "categories": []}
{"id": "2604.19984", "title": "Bias in the Tails: How Name-conditioned Evaluative Framing in Resume Summaries Destabilizes LLM-based Hiring", "abstract": "Research has documented LLMs' name-based bias in hiring and salary recommendations. In this paper, we instead consider a setting where LLMs generate candidate summaries for downstream assessment. In a large-scale controlled study, we analyze nearly one million resume summaries produced by 4 models under systematic race-gender name perturbations, using synthetic resumes and real-world job postings. By decomposing each summary into resume-grounded factual content and evaluative framing, we find that factual content remains largely stable, while evaluative language exhibits subtle name-conditioned variation concentrated in the extremes of the distribution, especially in open-source models. Our hiring simulation demonstrates how evaluative summary transforms directional harm into symmetric instability that might evade conventional fairness audit, highlighting a potential pathway for LLM-to-LLM automation bias.", "date": "2026-04-21", "categories": []}
{"id": "2604.18729", "title": "Investigating Counterfactual Unfairness in LLMs towards Identities through Humor", "abstract": "Humor holds up a mirror to social perception: what we find funny often reflects who we are and how we judge others. When language models engage with humor, their reactions expose the social assumptions they have internalized from training data. In this paper, we investigate counterfactual unfairness through humor by observing how the model's responses change when we swap who speaks and who is addressed while holding other factors constant. Our framework spans three tasks: humor generation refusal, speaker intention inference, and relational/societal impact prediction, covering both identity-agnostic humor and identity-specific disparagement humor. We introduce interpretable bias metrics that capture asymmetric patterns under identity swaps. Experiments across state-of-the-art models reveal consistent relational disparities: jokes told by privileged speakers are refused up to 67.5% more often, judged as malicious 64.7% more frequently, and rated up to 1.5 points higher in social harm on a 5-point scale. These patterns highlight how sensitivity and stereotyping coexist in generative models, complicating efforts toward fairness and cultural alignment.", "date": "2026-04-20", "categories": []}
{"id": "2604.19204", "title": "Auditing LLMs for Algorithmic Fairness in Casenote-Augmented Tabular Prediction", "abstract": "LLMs are increasingly being considered for prediction tasks in high-stakes social service settings, but their algorithmic fairness properties in this context are poorly understood. In this short technical report, we audit the algorithmic fairness of LLM-based tabular classification on a real housing placement prediction task, augmented with street outreach casenotes from a nonprofit partner. We audit multi-class classification error disparities. We find that a fine-tuned model augmented with casenote summaries can improve accuracy while reducing algorithmic fairness disparities. We experiment with variable importance improvements to zero-shot tabular classification and find mixed results on resulting algorithmic fairness. Overall, given historical inequities in housing placement, it is crucial to audit LLM use. We find that leveraging LLMs to augment tabular classification with casenote summaries can safely leverage additional text information at low implementation burden. The outreach casenotes are fairly short and heavily redacted. Our assessment is that LLM zero-shot classification does not introduce additional textual biases beyond algorithmic biases in tabular classification. Combining fine-tuning and leveraging casenote summaries can improve accuracy and algorithmic fairness.", "date": "2026-04-21", "categories": []}
{"id": "2604.19357", "title": "FairTree: Subgroup Fairness Auditing of Machine Learning Models with Bias-Variance Decomposition", "abstract": "The evaluation of machine learning models typically relies mainly on performance metrics based on loss functions, which risk to overlook changes in performance in relevant subgroups. Auditing tools such as SliceFinder and SliceLine were proposed to detect such groups, but usually have conceptual disadvantages, such as the inability to directly address continuous covariates. In this paper, we introduce FairTree, a novel algorithm adapted from psychometric invariance testing. Unlike SliceFinder and related algorithms, FairTree directly handles continuous, categorical, and ordinal features without discretization. It further decomposes performance disparities into systematic bias and variance, allowing a categorization of changes in algorithm performance. We propose and evaluate two variations of the algorithm: a permutation-based approach, which is conceptually closer to SliceFinder, and a fluctuation test. Through simulation studies that include a direct comparison with SliceLine, we demonstrate that both approaches have a satisfactory rate of false-positive results, but that the fluctuation approach has relatively higher power. We further illustrate the method on the UCI Adult Census dataset. The proposed algorithms provide a flexible framework for the statistical evaluation of the performance and aspects of fairness of machine learning models in a wide range of applications even in relatively small data.", "date": "2026-04-21", "categories": []}
{"id": "2604.19431", "title": "Counting Worlds Branching Time Semantics for post-hoc Bias Mitigation in generative AI", "abstract": "Generative AI systems are known to amplify biases present in their training data. While several inference-time mitigation strategies have been proposed, they remain largely empirical and lack formal guarantees. In this paper we introduce CTLF, a branching-time logic designed to reason about bias in series of generative AI outputs. CTLF adopts a counting worlds semantics where each world represents a possible output at a given step in the generation process and introduces modal operators that allow us to verify whether the current output series respects an intended probability distribution over a protected attribute, to predict the likelihood of remaining within acceptable bounds as new outputs are generated, and to determine how many outputs are needed to remove in order to restore fairness. We illustrate the framework on a toy example of biased image generation, showing how CTLF formulas can express concrete fairness properties at different points in the output series.", "date": "2026-04-21", "categories": []}
{"id": "2604.19468", "title": "Fairness Audits of Institutional Risk Models in Deployed ML Pipelines", "abstract": "Fairness audits of institutional risk models are critical for understanding how deployed machine learning pipelines allocate resources. Drawing on multi-year collaboration with Centennial College, where our prior ethnographic work introduced the ASP-HEI Cycle, we present a replica-based audit of a deployed Early Warning System (EWS), replicating its model using institutional training data and design specifications. We evaluate disparities by gender, age, and residency status across the full pipeline (training data, model predictions, and post-processing) using standard fairness metrics. Our audit reveals systematic misallocation: younger, male, and international students are disproportionately flagged for support, even when many ultimately succeed, while older and female students with comparable dropout risk are under-identified. Post-processing amplifies these disparities by collapsing heterogeneous probabilities into percentile-based risk tiers. This work provides a replicable methodology for auditing institutional ML systems and shows how disparities emerge and compound across stages, highlighting the importance of evaluating construct validity alongside statistical fairness. It contributes one empirical thread to a broader program investigating algorithms, student data, and power in higher education.", "date": "2026-04-21", "categories": []}
{"id": "2601.06931", "title": "Measuring Social Bias in Vision-Language Models with Face-Only Counterfactuals from Real Photos", "abstract": "Vision-Language Models (VLMs) are increasingly deployed in socially consequential settings, raising concerns about social bias driven by demographic cues. A central challenge in measuring such social bias is attribution under visual confounding: real-world images entangle race and gender with correlated factors such as background and clothing, obscuring attribution. We propose a \\textbf{face-only counterfactual evaluation paradigm} that isolates demographic effects while preserving real-image realism. Starting from real photographs, we generate counterfactual variants by editing only facial attributes related to race and gender, keeping all other visual factors fixed. Based on this paradigm, we construct \\textbf{FOCUS}, a dataset of 480 scene-matched counterfactual images across six occupations and ten demographic groups, and propose \\textbf{REFLECT}, a benchmark comprising three decision-oriented tasks: two-alternative forced choice, multiple-choice socioeconomic inference, and numeric salary recommendation. Experiments on five state-of-the-art VLMs reveal that demographic disparities persist under strict visual control and vary substantially across task formulations. These findings underscore the necessity of controlled, counterfactual audits and highlight task design as a critical factor in evaluating social bias in multimodal models.", "date": "2026-01-11", "categories": []}
{"id": "2602.23580", "title": "BRIDGE the Gap: Mitigating Bias Amplification in Automated Scoring of English Language Learners via Inter-group Data Augmentation", "abstract": "In the field of educational assessment, automated scoring systems increasingly rely on deep learning and large language models (LLMs). However, these systems face significant risks of bias amplification, where model prediction gaps between student groups become larger than those observed in training data. This issue is especially severe for underrepresented groups such as English Language Learners (ELLs), as models may inherit and further magnify existing disparities in the data. We identify that this issue is closely tied to representation bias: the scarcity of minority (high-scoring ELL) samples makes models trained with empirical risk minimization favor majority (non-ELL) linguistic patterns. Consequently, models tend to under-predict ELL students who even demonstrate comparable domain knowledge but use different linguistic patterns, thereby undermining the fairness of automated scoring outcomes. To mitigate this, we propose BRIDGE, a Bias-Reducing Inter-group Data GEneration framework designed for low-resource assessment settings. Instead of relying on the limited minority samples, BRIDGE synthesizes high-scoring ELL samples by \"pasting\" construct-relevant (i.e., rubric-aligned knowledge and evidence) content from abundant high-scoring non-ELL samples into authentic ELL linguistic patterns. We further introduce a discriminator model to ensure the quality of synthetic samples. Experiments on California Science Test (CAST) datasets demonstrate that BRIDGE effectively reduces prediction bias for high-scoring ELL students while maintaining overall scoring performance. Notably, our method achieves fairness gains comparable to using additional real human data, offering a cost-effective solution for ensuring equitable scoring in large-scale assessments.", "date": "2026-02-27", "categories": []}
{"id": "2604.16516", "title": "Operationalizing Fairness in Text-to-Image Models: A Survey of Bias, Fairness Audits and Mitigation Strategies", "abstract": "Text-to-Image (T2I) generation models have been widely adopted across various industries, yet are criticized for frequently exhibiting societal stereotypes. While a growing body of research has emerged to evaluate and mitigate these biases, the field at present contends with conceptual ambiguity, for example terms like \"bias\" and \"fairness\" are not always clearly distinguished and often lack clear operational definitions. This paper provides a comprehensive systematic review of T2I fairness literature, organizing existing work into a taxonomy of bias types and fairness notions. We critically assess the gap between \"target fairness\" (normative ideals in T2I outputs) and \"threshold fairness\" (normative standards with actionable decision rules). Furthermore, we survey the landscape of mitigation strategies, ranging from prompt engineering to diffusion process manipulation. We conclude by proposing a new framework for operationalizing fairness that moves beyond descriptive metrics towards rigorous, target-based testing, offering an approach for more accountable generative AI development.", "date": "2026-04-15", "categories": []}
{"id": "2604.16610", "title": "Fairness Constraints in High-Dimensional Generalized Linear Models", "abstract": "Machine learning models often inherit biases from historical data, raising critical concerns about fairness and accountability. Conventional fairness interventions typically require access to sensitive attributes like gender or race, but privacy and legal restrictions frequently limit their use. To address this challenge, we propose a framework that infers sensitive attributes from auxiliary features and integrates fairness constraints into model training. Our approach mitigates bias while preserving predictive accuracy, offering a practical solution for fairness-aware learning. Empirical evaluations validate its effectiveness, contributing to the advancement of more equitable algorithmic decision-making.", "date": "2026-04-17", "categories": []}
{"id": "2604.16780", "title": "FairNVT: Improving Fairness via Noise Injection in Vision Transformers", "abstract": "This paper presents FairNVT, a lightweight debiasing framework for pretrained transformer-based encoders that improves both representation and prediction level fairness while preserving task accuracy. Unlike many existing debiasing approaches that address these notions separately, we argue they are inherently connected: suppressing sensitive information at the representation level can facilitate fairer predictions. Our approach learns task-relevant and sensitive embeddings via lightweight adapters, applies calibrated Gaussian noise to the sensitive embedding, and fuses it with the task representation. Together with orthogonality constraints and fairness regularization, these components jointly reduce sensitive-attribute leakage in the learned embeddings and encourage fairer downstream predictions. The framework is compatible with a wide range of pretrained transformer encoders. Across three datasets spanning vision and language, FairNVT reduces sensitive-attribute attacker accuracy, improves demographic-parity and equalized-odds metrics, and maintains high task performance.", "date": "2026-04-18", "categories": []}
{"id": "2604.17971", "title": "Identifying Ethical Biases in Action Recognition Models", "abstract": "Human Action Recognition (HAR) models are increasingly deployed in high-stakes environments, yet their fairness across different human appearances has not been analyzed. We introduce a framework for auditing bias in HAR models using synthetic video data, generated with full control over visual identity attributes such as skin color. Unlike prior work that focuses on static images or pose estimation, our approach preserves temporal consistency, allowing us to isolate and test how changes to a single attribute affect model predictions. Through controlled interventions using the BEDLAM simulation platform, we show whether some popular HAR models exhibit statistically significant biases on the skin color even when the motion remains identical. Our results highlight how models may encode unwanted visual associations, and we provide evidence of systematic errors across groups. This work contributes a framework for auditing HAR models and supports the development of more transparent, accountable systems in light of upcoming regulatory standards.", "date": "2026-04-20", "categories": []}
{"id": "2604.18249", "title": "Where Do Self-Supervised Speech Models Become Unfair?", "abstract": "Speech encoder models are known to model members of some speaker groups (SGs) better than others. However, there has been little work in establishing why this occurs on a technological level. To our knowledge, we present the first layerwise fairness analysis of pretrained self-supervised speech encoder models (S3Ms), probing each embedding layer for speaker identification (SID) automatic speech recognition (ASR). We find S3Ms produce embeddings biased against certain SGs for both tasks, starting at the very first latent layers. Furthermore, we find opposite patterns of layerwise bias for SID vs ASR for all models in our study: SID bias is minimized in layers that minimize overall SID error; on the other hand, ASR bias is maximized in layers that minimize overall ASR error. The inverse bias/error relationship for ASR is unaffected when probing S3Ms that are finetuned for ASR, suggesting SG-level bias is established during pretraining and is difficult to remove.", "date": "2026-04-20", "categories": []}
{"id": "2604.16061", "title": "Constant-Factor Approximations for Doubly Constrained Fair k-Center, k-Median and k-Means", "abstract": "We study discrete k-clustering problems in general metric spaces that are constrained by a combination of two different fairness conditions within the demographic fairness model. Given a metric space (P,d), where every point in P is equipped with a protected attribute, and a number k, the goal is to partition P into k clusters with a designated center each, such that a center-based objective function is minimized and the attributes are fairly distributed with respect to the following two fairness concepts: 1) group fairness: We aim for clusters with balanced numbers of attributes by specifying lower and upper bounds for the desired attribute proportions. 2) diverse center selection: Clusters have natural representatives, i.e., their centers. We ask for a balanced set of representatives by specifying the desired number of centers to choose from each attribute.\n  Dickerson, Esmaeili, Morgenstern and Zhang (2023) denote the combination of these two constraints as doubly constrained fair clustering. They present algorithms whose guarantees depend on the best known approximation factors for either of these problems. Currently, this implies an 8-approximation with a small additive violation on the group fairness constraint. For k-center, we improve this approximation factor to 4 with a small additive violation. This guarantee also depends on the currently best algorithm for DS-fair k-center given by Jones, Nguyen and Nguyen (2020). For k-median and k-means, we propose the first constant-factor approximation algorithms. Our algorithms transform a solution that satisfies diverse center selection into a doubly constrained fair clustering using an LP-based approach. Furthermore, our results are generalizable to other center-selection constraints, such as matroid k-clustering and knapsack constraints.", "date": "2026-04-17", "categories": []}
{"id": "2604.15078", "title": "Where are the Humans? A Scoping Review of Fairness in Multi-agent AI Systems", "abstract": "Rapid advances in Generative AI are giving rise to increasingly sophisticated Multi-Agent AI (MAAI) systems. While AI fairness has been extensively studied in traditional predictive scenarios, its examination in MAAI remains nascent and fragmented. This scoping review critically synthesizes existing research on fairness in MAAI systems. Through a qualitative content analysis of 23 selected studies, we identify five archetypal approaches. Our findings reveal that fairness in MAAI systems is often addressed superficially, lacks robust normative foundations, and frequently overlooks the complex dynamics introduced by agent autonomy and system-level interactions. We argue that fairness must be embedded structurally throughout the development lifecycle of MAAI, rather than appended as a post-hoc consideration. Meaningful evaluation requires explicit human oversight, normative clarity, and a precise articulation of fairness objectives and beneficiaries. This review provides a foundation for advancing fairness research in MAAI systems by highlighting critical gaps, exposing prevailing limitations, and suggesting pathways.", "date": "2026-04-16", "categories": []}
{"id": "2604.13305", "title": "Bias at the End of the Score", "abstract": "Reward models (RMs) are inherently non-neutral value functions designed and trained to encode specific objectives, such as human preferences or text-image alignment. RMs have become crucial components of text-to-image (T2I) generation systems where they are used at various stages for dataset filtering, as evaluation metrics, as a supervisory signal during optimization of parameters, and for post-generation safety and quality filtering of T2I outputs. While specific problems with the integration of RMs into the T2I pipeline have been studied (e.g. reward hacking or mode collapse), their robustness and fairness as scoring functions remains largely unknown. We conduct a large scale audit of RM robustness with respect to demographic biases during T2I model training and generation. We provide quantitative and qualitative evidence that while originally developed as quality measures, RMs encode demographic biases, which cause reward-guided optimization to disproportionately sexualize female image subjects reinforce gender/racial stereotypes, and collapse demographic diversity. These findings highlight shortcomings in current reward models, challenge their reliability as quality metrics, and underscore the need for improved data collection and training procedures to enable more robust scoring.", "date": "2026-04-14", "categories": []}
{"id": "2604.13705", "title": "Beyond Arrow's Impossibility: Fairness as an Emergent Property of Multi-Agent Collaboration", "abstract": "Fairness in language models is typically studied as a property of a single, centrally optimized model. As large language models become increasingly agentic, we propose that fairness emerges through interaction and exchange. We study this via a controlled hospital triage framework in which two agents negotiate over three structured debate rounds. One agent is aligned to a specific ethical framework via retrieval-augmented generation (RAG), while the other is either unaligned or adversarially prompted to favor demographic groups over clinical need. We find that alignment systematically shapes negotiation strategies and allocation patterns, and that neither agent's allocation is ethically adequate in isolation, yet their joint final allocation can satisfy fairness criteria that neither would have reached alone. Aligned agents partially moderate bias through contestation rather than override, acting as corrective patches that restore access for marginalized groups without fully converting a biased counterpart. We further observe that even explicitly aligned agents exhibit intrinsic biases toward certain frameworks, consistent with known left-leaning tendencies in LLMs. We connect these limits to Arrow's Impossibility Theorem: no aggregation mechanism can simultaneously satisfy all desiderata of collective rationality, and multi-agent deliberation navigates rather than resolves this constraint. Our results reposition fairness as an emergent, procedural property of decentralized agent interaction, and the system rather than the individual agent as the appropriate unit of evaluation.", "date": "2026-04-15", "categories": []}
{"id": "2604.14035", "title": "First-See-Then-Design: A Multi-Stakeholder View for Optimal Performance-Fairness Trade-Offs", "abstract": "Fairness in algorithmic decision-making is often defined in the predictive space, where predictive performance - used as a proxy for decision-maker (DM) utility - is traded off against prediction-based fairness notions, such as demographic parity or equality of opportunity. This perspective, however, ignores how predictions translate into decisions and ultimately into utilities and welfare for both DM and decision subjects (DS), as well as their allocation across social-salient groups.\n  In this paper, we propose a multi-stakeholder framework for fair algorithmic decision-making grounded in welfare economics and distributive justice, explicitly modeling the utilities of both the DM and DS, and defining fairness via a social planner's utility that captures inequalities in DS utilities across groups under different justice-based fairness notions (e.g., Egalitarian, Rawlsian). We formulate fair decision-making as a post-hoc multi-objective optimization problem, characterizing the achievable performance-fairness trade-offs in the two-dimensional utility space of DM utility and the social planner's utility, under different decision policy classes (deterministic vs. stochastic, shared vs. group-specific). Using the proposed framework, we then identify conditions (in terms of the stakeholders' utilities) under which stochastic policies are more optimal than deterministic ones, and empirically demonstrate that simple stochastic policies can yield superior performance-fairness trade-offs by leveraging outcome uncertainty. Overall, we advocate a shift from prediction-centric fairness to a transparent, justice-based, multi-stakeholder approach that supports the collaborative design of decision-making policies.", "date": "2026-04-15", "categories": []}
{"id": "2604.10200", "title": "Edu-MMBias: A Three-Tier Multimodal Benchmark for Auditing Social Bias in Vision-Language Models under Educational Contexts", "abstract": "As Vision-Language Models (VLMs) become integral to educational decision-making, ensuring their fairness is paramount. However, current text-centric evaluations neglect the visual modality, leaving an unregulated channel for latent social biases. To bridge this gap, we present Edu-MMBias, a systematic auditing framework grounded in the tri-component model of attitudes from social psychology. This framework diagnoses bias across three hierarchical dimensions: cognitive, affective, and behavioral. Utilizing a specialized generative pipeline that incorporates a self-correct mechanism and human-in-the-loop verification, we synthesize contamination-resistant student profiles to conduct a holistic stress test on state-of-the-art VLMs. Our extensive audit reveals critical, counter-intuitive patterns: models exhibit a compensatory class bias favoring lower-status narratives while simultaneously harboring deep-seated health and racial stereotypes. Crucially, we find that visual inputs act as a safety backdoor, triggering a resurgence of biases that bypass text-based alignment safeguards and revealing a systematic misalignment between latent cognition and final decision-making. The contributions of this paper are available at: https://anonymous.4open.science/r/EduMMBias-63B2.", "date": "2026-04-11", "categories": []}
{"id": "2604.12337", "title": "Identifying and Mitigating Gender Cues in Academic Recommendation Letters: An Interpretability Case Study", "abstract": "Letters of recommendation (LoRs) can carry patterns of implicitly gendered language that can inadvertently influence downstream decisions, e.g. in hiring and admissions. In this work, we investigate the extent to which Transformer-based encoder models as well as Large Language Models (LLMs) can infer the gender of applicants in academic LoRs submitted to an U.S. medical-residency program after explicit identifiers like names and pronouns are de-gendered. While using three models (DistilBERT, RoBERTa, and Llama 2) to classify the gender of anonymized and de-gendered LoRs, significant gender leakage was observed as evident from up to 68% classification accuracy. Text interpretation methods, like TF-IDF and SHAP, demonstrate that certain linguistic patterns are strong proxies for gender, e.g. \"emotional'' and \"humanitarian'' are commonly associated with LoRs from female applicants. As an experiment in creating truly gender-neutral LoRs, these implicit gender cues were remove resulting in a drop of up to 5.5% accuracy and 2.7% macro $F_1$ score on re-training the classifiers. However, applicant gender prediction still remains better than chance. In this case study, our findings highlight that 1) LoRs contain gender-identifying cues that are hard to remove and may activate bias in decision-making and 2) while our technical framework may be a concrete step toward fairer academic and professional evaluations, future work is needed to interrogate the role that gender plays in LoR review. Taken together, our findings motivate upstream auditing of evaluative text in real-world academic letters of recommendation as a necessary complement to model-level fairness interventions.", "date": "2026-04-14", "categories": []}
{"id": "2604.12481", "title": "T2I-BiasBench: A Multi-Metric Framework for Auditing Demographic and Cultural Bias in Text-to-Image Models", "abstract": "Text-to-image (T2I) generative models achieve impressive visual fidelity but inherit and amplify demographic imbalances and cultural biases embedded in training data. We introduce T2I-BiasBench, a unified evaluation framework of thirteen complementary metrics that jointly captures demographic bias, element omission, and cultural collapse in diffusion models - the first framework to address all three dimensions simultaneously.\n  We evaluate three open-source models - Stable Diffusion v1.5, BK-SDM Base, and Koala Lightning - against Gemini 2.5 Flash (RLHF-aligned) as a reference baseline. The benchmark comprises 1,574 generated images across five structured prompt categories. T2I-BiasBench integrates six established metrics with seven additional measures: four newly proposed (Composite Bias Score, Grounded Missing Rate, Implicit Element Missing Rate, Cultural Accuracy Ratio) and three adapted (Hallucination Score, Vendi Score, CLIP Proxy Score).\n  Three key findings emerge: (1) Stable Diffusion v1.5 and BK-SDM exhibit bias amplification (>1.0) in beauty-related prompts; (2) contextual constraints such as surgical PPE substantially attenuate professional-role gender bias (Doctor CBS = 0.06 for SD v1.5); and (3) all models, including RLHF-aligned Gemini, collapse to a narrow set of cultural representations (CAS: 0.54-1.00), confirming that alignment techniques do not resolve cultural coverage gaps.\n  T2I-BiasBench is publicly released to support standardized, fine-grained bias evaluation of generative models. The project page is available at: https://gyanendrachaubey.github.io/T2I-BiasBench/", "date": "2026-04-14", "categories": []}
{"id": "2604.08963", "title": "Aligned Agents, Biased Swarm: Measuring Bias Amplification in Multi-Agent Systems", "abstract": "While Multi-Agent Systems (MAS) are increasingly deployed for complex workflows, their emergent properties-particularly the accumulation of bias-remain poorly understood. Because real-world MAS are too complex to analyze entirely, evaluating their ethical robustness requires first isolating their foundational mechanics. In this work, we conduct a baseline empirical study investigating how basic MAS topologies and feedback loops influence prejudice. Contrary to the assumption that multi-agent collaboration naturally dilutes bias, we hypothesize that structured workflows act as echo chambers, amplifying minor stochastic biases into systemic polarization. To evaluate this, we introduce Discrim-Eval-Open, an open-ended benchmark that bypasses individual model neutrality through forced comparative judgments across demographic groups. Analyzing bias cascades across various structures reveals that architectural sophistication frequently exacerbates bias rather than mitigating it. We observe systemic amplification even when isolated agents operate neutrally, and identify a 'Trigger Vulnerability' where injecting purely objective context drastically accelerates polarization. By stripping away advanced swarm complexity to study foundational dynamics, we establish a crucial baseline: structural complexity does not guarantee ethical robustness. Our code is available at https://github.com/weizhihao1/MAS-Bias.", "date": "2026-04-10", "categories": []}
{"id": "2604.09620", "title": "LLM Nepotism in Organizational Governance", "abstract": "Large language models are increasingly used to support organizational decisions from hiring to governance, raising fairness concerns in AI-assisted evaluation. Prior work has focused mainly on demographic bias and broader preference effects, rather than on whether evaluators reward expressed trust in AI itself. We study this phenomenon as LLM Nepotism, an attitude-driven bias channel in which favorable signals toward AI are rewarded even when they are not relevant to role-related merit. We introduce a two-phase simulation pipeline that first isolates AI-trust preference in qualification-matched resume screening and then examines its downstream effects in board-level decision making. Across several popular LLMs, we find that resume screeners tend to favor candidates with positive or non-critical attitudes toward AI, discriminating skeptical, human-centered counterparts. These biases suggest a loophole: LLM-based hiring can produce more homogeneous AI-trusting organizations, whose decision-makers exhibit greater scrutiny failure and delegation to AI agents, approving flawed proposals more readily while favoring AI-delegation initiatives. To mitigate this behavior, we additionally study prompt-based mitigation and propose Merit-Attitude Factorization, which separates non-merit AI attitude from merit-based evaluation and attenuates this bias across experiments.", "date": "2026-03-18", "categories": []}
{"id": "2604.09656", "title": "Fairboard: a quantitative framework for equity assessment of healthcare models", "abstract": "Despite there now being more than 1,000 FDA-authorised AI medical devices, formal equity assessments -- whether model performance is uniform across patient subgroups -- are rare. Here, we evaluate the equity of 18 open-source brain tumour segmentation models across 648 glioma patients from two independent datasets (n = 11,664 model inferences) along distinct univariate, Bayesian multivariate, spatial, and representational dimensions. We find that patient identity consistently explains more performance variance than model choice, with clinical factors, including molecular diagnosis, tumour grade, and extent of resection, predicting segmentation accuracy more strongly than model architecture. A voxel-wise spatial meta-analysis identifies neuroanatomically localised biases that are compartment-specific yet often consistent across models. Within a high-dimensional latent space of lesion masks and clinic-demographic features, model performance clusters significantly, indicating that the patient feature space contains axes of algorithmic vulnerability. Although newer models tend toward greater equity, none provide a formal fairness guarantee. Lastly, we release Fairboard, an open-source, no-code dashboard that lowers barriers to equitable model monitoring in medical imaging.", "date": "2026-03-30", "categories": []}
{"id": "2604.10014", "title": "Demographic and Linguistic Bias Evaluation in Omnimodal Language Models", "abstract": "This paper provides a comprehensive evaluation of demographic and linguistic biases in omnimodal language models that process text, images, audio, and video within a single framework. Although these models are being widely deployed, their performance across different demographic groups and modalities is not well studied. Four omnimodal models are evaluated on tasks that include demographic attribute estimation, identity verification, activity recognition, multilingual speech transcription, and language identification. Accuracy differences are measured across age, gender, skin tone, language, and country of origin. The results show that image and video understanding tasks generally exhibit better performance with smaller demographic disparities. In contrast, audio understanding tasks exhibit significantly lower performance and substantial bias, including large accuracy differences across age groups, genders, and languages, and frequent prediction collapse toward narrow categories. These findings highlight the importance of evaluating fairness across all supported modalities as omnimodal language models are increasingly used in real-world applications.", "date": "2026-04-11", "categories": []}
{"id": "2604.10224", "title": "Exploring the impact of fairness-aware criteria in AutoML", "abstract": "Machine Learning (ML) systems are increasingly used to support decision-making processes that affect individuals. However, these systems often rely on biased data, which can lead to unfair outcomes against specific groups. With the growing adoption of Automated Machine Learning (AutoML), the risk of intensifying discriminatory behaviours increases, as most frameworks primarily focus on model selection to maximise predictive performance. Previous research on fairness in AutoML had largely followed this trend, integrating fairness awareness only in the model selection or hyperparameter tuning, while neglecting other critical stages of the ML pipeline. This paper aims to study the impact of integrating fairness directly into the optimisation component of an AutoML framework that constructs complete ML pipelines, from data selection and transformations to model selection and tuning. As selecting appropriate fairness metrics remains a key challenge, our work incorporates complementary fairness metrics to capture different dimensions of fairness during the optimisation. Their integration within AutoML resulted in measurable differences compared to a baseline focused solely on predictive performance. Despite a 9.4% decrease in predictive power, the average fairness improved by 14.5%, accompanied by a 35.7% reduction in data usage. Furthermore, fairness integration produced complete yet simpler final solutions, suggesting that model complexity is not always required to achieve balanced and fair ML solutions.", "date": "2026-04-11", "categories": []}
{"id": "2603.27141", "title": "Routing Sensitivity Without Controllability: A Diagnostic Study of Fairness in MoE Language Models", "abstract": "Mixture-of-Experts (MoE) language models are universally sensitive to demographic content at the routing level, yet exploiting this sensitivity for fairness control is structurally limited. We introduce Fairness-Aware Routing Equilibrium (FARE), a diagnostic framework designed to probe the limits of routing-level stereotype intervention across diverse MoE architectures. FARE reveals that routing-level preference shifts are either unachievable (Mixtral, Qwen1.5, Qwen3), statistically non-robust (DeepSeekMoE), or accompanied by substantial utility cost (OLMoE, -4.4%p CrowS-Pairs at -6.3%p TQA). Critically, even where log-likelihood preference shifts are robust, they do not transfer to decoded generation: expanded evaluations on both non-null models yield null results across all generation metrics. Group-level expert masking reveals why: bias and core knowledge are deeply entangled within expert groups. These findings indicate that routing sensitivity is necessary but insufficient for stereotype control, and identify specific architectural conditions that can inform the design of more controllable future MoE systems.", "date": "2026-03-28", "categories": []}
{"id": "2604.08561", "title": "A Representation-Level Assessment of Bias Mitigation in Foundation Models", "abstract": "We investigate how successful bias mitigation reshapes the embedding space of encoder-only and decoder-only foundation models, offering an internal audit of model behaviour through representational analysis. Using BERT and Llama2 as representative architectures, we assess the shifts in associations between gender and occupation terms by comparing baseline and bias-mitigated variants of the models. Our findings show that bias mitigation reduces gender-occupation disparities in the embedding space, leading to more neutral and balanced internal representations. These representational shifts are consistent across both model types, suggesting that fairness improvements can manifest as interpretable and geometric transformations. These results position embedding analysis as a valuable tool for understanding and validating the effectiveness of debiasing methods in foundation models. To further promote the assessment of decoder-only models, we introduce WinoDec, a dataset consisting of 4,000 sequences with gender and occupation terms, and release it to the general public. (https://github.com/winodec/wino-dec)", "date": "2026-03-17", "categories": []}
{"id": "2512.09665", "title": "OxEnsemble: Fair Ensembles for Low-Data Classification", "abstract": "We address the problem of fair classification in settings where data is scarce and unbalanced across demographic groups. Such low-data regimes are common in domains like medical imaging, where false negatives can have fatal consequences. We propose a novel approach \\emph{OxEnsemble} for efficiently training ensembles and enforcing fairness in these low-data regimes. Unlike other approaches, we aggregate predictions across ensemble members, each trained to satisfy fairness constraints. By construction, \\emph{OxEnsemble} is both data-efficient -- carefully reusing held-out data to enforce fairness reliably -- and compute-efficient, requiring little more compute than used to fine-tune or evaluate an existing model. We validate this approach with new theoretical guarantees. Experimentally, our approach yields more consistent outcomes and stronger fairness-accuracy trade-offs than existing methods across multiple challenging medical imaging classification datasets.", "date": "2025-12-10", "categories": []}
{"id": "2510.20220", "title": "Alternatives to the Laplacian for Scalable Spectral Clustering with Group Fairness Constraints", "abstract": "Recent research has focused on mitigating algorithmic bias in clustering by incorporating fairness constraints into algorithmic design. Notions such as disparate impact, community cohesion, and cost per population have been implemented to enforce equitable outcomes. Among these, group fairness (balance) ensures that each protected group is proportionally represented within every cluster. However, incorporating balance as a metric of fairness into spectral clustering algorithms has led to computational times that can be improved. This study aims to enhance the efficiency of spectral clustering algorithms by reformulating the constrained optimization problem using a new formulation derived from the Lagrangian method and the Sherman-Morrison-Woodbury (SMW) identity, resulting in the Fair-SMW algorithm. Fair-SMW employs three alternatives to the Laplacian matrix with different spectral gaps to generate multiple variations of Fair-SMW, achieving clustering solutions with comparable balance to existing algorithms while offering improved runtime performance. We present the results of Fair-SMW, evaluated using the Stochastic Block Model (SBM) to measure both runtime efficiency and balance across real-world network datasets, including LastFM, FacebookNet, Deezer, and German. We achieve an improvement in computation time that is twice as fast as the state-of-the-art, and also flexible enough to achieve twice as much balance.", "date": "2025-10-22", "categories": []}
{"id": "2604.06378", "title": "Revisiting Fairness Impossibility with Endogenous Behavior", "abstract": "In many real-world settings, institutions can and do adjust the consequences attached to algorithmic classification decisions, such as the size of fines, sentence lengths, or benefit levels. We refer to these consequences as the stakes associated with classification. These stakes can give rise to behavioral responses to classification, as people adjust their actions in anticipation of how they will be classified. Much of the algorithmic fairness literature evaluates classification outcomes while holding behavior fixed, treating behavioral differences across groups as exogenous features of the environment. Under this assumption, the stakes of classification play no role in shaping outcomes.\n  We revisit classic impossibility results in algorithmic fairness in a setting where people respond strategically to classification. We show that, in this environment, the well-known incompatibility between error-rate balance and predictive parity disappears, but only by potentially introducing a qualitatively different form of unequal treatment. Concretely, we construct a two-stage design in which a classifier first standardizes its statistical performance across groups, and then adjusts stakes so as to induce comparable patterns of behavior. This requires treating groups differently in the consequences attached to identical classification decisions. Our results demonstrate that fairness in strategic settings cannot be assessed solely by how algorithms map data into decisions. Rather, our analysis treats the human consequences of classification as primary design variables, introduces normative criteria governing their use, and shows that their interaction with statistical fairness criteria generates qualitatively new tradeoffs. Our aim is to make these tradeoffs precise and explicit.", "date": "2026-04-07", "categories": []}
{"id": "2604.07009", "title": "CAFP: A Post-Processing Framework for Group Fairness via Counterfactual Model Averaging", "abstract": "Ensuring fairness in machine learning predictions is a critical challenge, especially when models are deployed in sensitive domains such as credit scoring, healthcare, and criminal justice. While many fairness interventions rely on data preprocessing or algorithmic constraints during training, these approaches often require full control over the model architecture and access to protected attribute information, which may not be feasible in real-world systems. In this paper, we propose Counterfactual Averaging for Fair Predictions (CAFP), a model-agnostic post-processing method that mitigates unfair influence from protected attributes without retraining or modifying the original classifier. CAFP operates by generating counterfactual versions of each input in which the sensitive attribute is flipped, and then averaging the model's predictions across factual and counterfactual instances. We provide a theoretical analysis of CAFP, showing that it eliminates direct dependence on the protected attribute, reduces mutual information between predictions and sensitive attributes, and provably bounds the distortion introduced relative to the original model. Under mild assumptions, we further show that CAFP achieves perfect demographic parity and reduces the equalized odds gap by at least half the average counterfactual bias.", "date": "2026-04-08", "categories": []}
{"id": "2604.05224", "title": "Attribution Bias in Large Language Models", "abstract": "As Large Language Models (LLMs) are increasingly used to support search and information retrieval, it is critical that they accurately attribute content to its original authors. In this work, we introduce AttriBench, the first fame- and demographically-balanced quote attribution benchmark dataset. Through explicitly balancing author fame and demographics, AttriBench enables controlled investigation of demographic bias in quote attribution. Using this dataset, we evaluate 11 widely used LLMs across different prompt settings and find that quote attribution remains a challenging task even for frontier models. We observe large and systematic disparities in attribution accuracy between race, gender, and intersectional groups. We further introduce and investigate suppression, a distinct failure mode in which models omit attribution entirely, even when the model has access to authorship information. We find that suppression is widespread and unevenly distributed across demographic groups, revealing systematic biases not captured by standard accuracy metrics. Our results position quote attribution as a benchmark for representational fairness in LLMs.", "date": "2026-04-06", "categories": []}
{"id": "2604.05830", "title": "\"OK Aura, Be Fair With Me\": Demographics-Agnostic Training for Bias Mitigation in Wake-up Word Detection", "abstract": "Voice-based interfaces are widely used; however, achieving fair Wake-up Word detection across diverse speaker populations remains a critical challenge due to persistent demographic biases. This study evaluates the effectiveness of demographics-agnostic training techniques in mitigating performance disparities among speakers of varying sex, age, and accent. We utilize the OK Aura database for our experiments, employing a training methodology that excludes demographic labels, which are reserved for evaluation purposes. We explore (i) data augmentation techniques to enhance model generalization and (ii) knowledge distillation of pre-trained foundational speech models. The experimental results indicate that these demographics-agnostic training techniques markedly reduce demographic bias, leading to a more equitable performance profile across different speaker groups. Specifically, one of the evaluated techniques achieves a Predictive Disparity reduction of 39.94\\% for sex, 83.65\\% for age, and 40.48\\% for accent when compared to the baseline. This study highlights the effectiveness of label-agnostic methodologies in fostering fairness in Wake-up Word detection.", "date": "2026-04-07", "categories": []}
{"id": "2505.12530", "title": "Enforcing Fair Predicted Scores on Intervals of Percentiles by Difference-of-Convex Constraints", "abstract": "Fairness in machine learning has become a critical concern. Existing approaches often focus on achieving full fairness across all score ranges generated by predictive models, ensuring fairness in both high- and low-percentile populations. However, this stringent requirement can compromise predictive performance and may not align with the practical fairness concerns of stakeholders. In this work, we propose a novel framework for building partially fair machine learning models that enforce fairness only within a specific percentile interval of interest while maintaining flexibility in other regions. We introduce statistical metrics to evaluate partial fairness within a given percentile interval. To achieve partial fairness, we propose an in-processing method by formulating the model training problem as constrained optimization with difference-of-convex constraints, which can be solved by an inexact difference-of-convex algorithm (IDCA). We provide the complexity analysis of IDCA for finding a nearly KKT point. Through numerical experiments on real-world datasets, we demonstrate that our framework achieves high predictive performance while enforcing partial fairness where it matters most.", "date": "2025-05-18", "categories": []}
{"id": "2511.08887", "title": "FAST-CAD: A Fairness-Aware Framework for Non-Contact Stroke Diagnosis", "abstract": "Stroke is an acute cerebrovascular disease, and timely diagnosis significantly improves patient survival. However, existing automated diagnosis methods suffer from fairness issues across demographic groups, potentially exacerbating healthcare disparities. In this work we propose FAST-CAD, a theoretically grounded framework that combines domain-adversarial training (DAT) with group distributionally robust optimization (Group-DRO) for fair and accurate non-contact stroke diagnosis. Our approach is built on domain adaptation and minimax fairness theory and provides convergence guarantees and fairness bounds. We curate a multimodal dataset covering 12 demographic subgroups defined by age, gender, and posture. FAST-CAD employs self-supervised encoders with adversarial domain discrimination to learn demographic-invariant representations, while Group-DRO optimizes worst-group risk to ensure robust performance across all subgroups. Extensive experiments show that our method achieves superior diagnostic performance while maintaining fairness across demographic groups, and our theoretical analysis supports the effectiveness of the unified DAT + Group-DRO framework. This work provides both practical advances and theoretical insights for fair medical AI systems.", "date": "2025-11-12", "categories": []}
{"id": "2603.29270", "title": "Unbiased Model Prediction Without Using Protected Attribute Information", "abstract": "The problem of bias persists in the deep learning community as models continue to provide disparate performance across different demographic subgroups. Therefore, several algorithms have been proposed to improve the fairness of deep models. However, a majority of these algorithms utilize the protected attribute information for bias mitigation, which severely limits their application in real-world scenarios. To address this concern, we have proposed a novel algorithm, termed as \\textbf{Non-Protected Attribute-based Debiasing (NPAD)} algorithm for bias mitigation, that does not require the protected attribute information. The proposed NPAD algorithm utilizes the auxiliary information provided by the non-protected attributes to optimize the model for bias mitigation. Further, two different loss functions, \\textbf{Debiasing via Attribute Cluster Loss (DACL)} and \\textbf{Filter Redundancy Loss (FRL)} have been proposed to optimize the model for fairness goals. Multiple experiments are performed on the LFWA and CelebA datasets for facial attribute prediction, and a significant reduction in bias across different gender and age subgroups is observed.", "date": "2026-03-31", "categories": []}
{"id": "2604.03323", "title": "InsightBoard: An Interactive Multi-Metric Visualization and Fairness Analysis Plugin for TensorBoard", "abstract": "Modern machine learning systems deployed in safety-critical domains require visibility not only into aggregate performance but also into how training dynamics affect subgroup fairness over time. Existing training dashboards primarily support single-metric monitoring and offer limited support for examining relationships between heterogeneous metrics or diagnosing subgroup disparities during training. We present InsightBoard, an interactive TensorBoard plugin that integrates synchronized multi-metric visualization with slice-based fairness diagnostics in a unified interface. InsightBoard enables practitioners to jointly inspect training dynamics, performance metrics, and subgroup disparities through linked multi-view plots, correlation analysis, and standard group fairness indicators computed over user-defined slices. Through case studies with YOLOX on the BDD100k dataset, we demonstrate that models achieving strong aggregate performance can still exhibit substantial demographic and environmental disparities that remain hidden under conventional monitoring. By making fairness diagnostics available during training, InsightBoard supports earlier, more informed model inspection without modifying existing training pipelines or introducing additional data stores.", "date": "2026-04-02", "categories": []}
{"id": "2604.03478", "title": "Investigating Data Interventions for Subgroup Fairness: An ICU Case Study", "abstract": "In high-stakes settings where machine learning models are used to automate decision-making about individuals, the presence of algorithmic bias can exacerbate systemic harm to certain subgroups of people. These biases often stem from the underlying training data. In practice, interventions to \"fix the data\" depend on the actual additional data sources available -- where many are less than ideal. In these cases, the effects of data scaling on subgroup performance become volatile, as the improvements from increased sample size are counteracted by the introduction of distribution shifts in the training set. In this paper, we investigate the limitations of combining data sources to improve subgroup performance within the context of healthcare. Clinical models are commonly trained on datasets comprised of patient electronic health record (EHR) data from different hospitals or admission departments. Across two such datasets, the eICU Collaborative Research Database and the MIMIC-IV dataset, we find that data addition can both help and hurt model fairness and performance, and many intuitive strategies for data selection are unreliable. We compare model-based post-hoc calibration and data-centric addition strategies to find that the combination of both is important to improve subgroup performance. Our work questions the traditional dogma of \"better data\" for overcoming fairness challenges by comparing and combining data- and model-based approaches.", "date": "2026-04-03", "categories": []}
{"id": "2604.04858", "title": "FairLogue: A Toolkit for Intersectional Fairness Analysis in Clinical Machine Learning Models", "abstract": "Objective: Algorithmic fairness is essential for equitable and trustworthy machine learning in healthcare. Most fairness tools emphasize single-axis demographic comparisons and may miss compounded disparities affecting intersectional populations. This study introduces Fairlogue, a toolkit designed to operationalize intersectional fairness assessment in observational and counterfactual contexts within clinical settings. Methods: Fairlogue is a Python-based toolkit composed of three components: 1) an observational framework extending demographic parity, equalized odds, and equal opportunity difference to intersectional populations; 2) a counterfactual framework evaluating fairness under treatment-based contexts; and 3) a generalized counterfactual framework assessing fairness under interventions on intersectional group membership. The toolkit was evaluated using electronic health record data from the All of Us Controlled Tier V8 dataset in a glaucoma surgery prediction task using logistic regression with race and gender as protected attributes. Results: Observational analysis identified substantial intersectional disparities despite moderate model performance (AUROC = 0.709; accuracy = 0.651). Intersectional evaluation revealed larger fairness gaps than single-axis analyses, including demographic parity differences of 0.20 and equalized odds true positive and false positive rate gaps of 0.33 and 0.15, respectively. Counterfactual analysis using permutation-based null distributions produced unfairness (\"u-value\") estimates near zero, suggesting observed disparities were consistent with chance after conditioning on covariates. Conclusion: Fairlogue provides a modular toolkit integrating observational and counterfactual methods for quantifying and evaluating intersectional bias in clinical machine learning workflows.", "date": "2026-04-06", "categories": []}
{"id": "2604.02342", "title": "Homophily-aware Supervised Contrastive Counterfactual Augmented Fair Graph Neural Network", "abstract": "In recent years, Graph Neural Networks (GNNs) have achieved remarkable success in tasks such as node classification, link prediction, and graph representation learning. However, they remain susceptible to biases that can arise not only from node attributes but also from the graph structure itself. Addressing fairness in GNNs has therefore emerged as a critical research challenge. In this work, we propose a novel model for training fairness-aware GNNs by improving the counterfactual augmented fair graph neural network framework (CAF). Specifically, our approach introduces a two-phase training strategy: in the first phase, we edit the graph to increase homophily ratio with respect to class labels while reducing homophily ratio with respect to sensitive attribute labels; in the second phase, we integrate a modified supervised contrastive loss and environmental loss into the optimization process, enabling the model to jointly improve predictive performance and fairness. Experiments on five real-world datasets demonstrate that our model outperforms CAF and several state-of-the-art graph-based learning methods in both classification accuracy and fairness metrics.", "date": "2026-02-08", "categories": []}
{"id": "2604.02660", "title": "SocioEval: A Template-Based Framework for Evaluating Socioeconomic Status Bias in Foundation Models", "abstract": "As Large Language Models (LLMs) increasingly power decision-making systems across critical domains, understanding and mitigating their biases becomes essential for responsible AI deployment. Although bias assessment frameworks have proliferated for attributes such as race and gender, socioeconomic status bias remains significantly underexplored despite its widespread implications in the real world. We introduce SocioEval, a template-based framework for systematically evaluating socioeconomic bias in foundation models through decision-making tasks. Our hierarchical framework encompasses 8 themes and 18 topics, generating 240 prompts across 6 class-pair combinations. We evaluated 13 frontier LLMs on 3,120 responses using a rigorous three-stage annotation protocol, revealing substantial variation in bias rates (0.42\\%-33.75\\%). Our findings demonstrate that bias manifests differently across themes lifestyle judgments show 10$\\times$ higher bias than education-related decisions and that deployment safeguards effectively prevent explicit discrimination but show brittleness to domain-specific stereotypes. SocioEval provides a scalable, extensible foundation for auditing class-based bias in language models.", "date": "2026-04-03", "categories": []}
{"id": "2604.02772", "title": "Multiple-Debias: A Full-process Debiasing Method for Multilingual Pre-trained Language Models", "abstract": "Multilingual Pre-trained Language Models (MPLMs) have become essential tools for natural language processing. However, they often exhibit biases related to sensitive attributes such as gender, race, and religion. In this paper, we introduce a comprehensive multilingual debiasing method named Multiple-Debias to address these issues across multiple languages. By incorporating multilingual counterfactual data augmentation and multilingual Self-Debias across both pre-processing and post-processing stages, alongside parameter-efficient fine-tuning, we significantly reduced biases in MPLMs across three sensitive attributes in four languages. We also extended CrowS-Pairs to German, Spanish, Chinese, and Japanese, validating our full-process multilingual debiasing method for gender, racial, and religious bias. Our experiments show that (i) multilingual debiasing methods surpass monolingual approaches in effectively mitigating biases, and (ii) integrating debiasing information from different languages notably improves the fairness of MPLMs.", "date": "2026-04-03", "categories": []}
{"id": "2507.14221", "title": "Fair Representation in Parliamentary Summaries: Measuring and Mitigating Inclusion Bias", "abstract": "The The use of Large language models (LLMs) to summarise parliamentary proceedings presents a promising means of increasing the accessibility of democratic participation. However, as these systems increasingly mediate access to political information -- filtering and framing content before it reaches users -- there are important fairness considerations to address. In this work, we evaluate 5 LLMs (both proprietary and open-weight) in the summarisation of plenary debates from the European Parliament to investigate the representational biases that emerge in this context. We develop an attribution-aware evaluation framework to measure speaker-level inclusion and mis-representation in debate summaries. Across all models and experiments, we find that speakers are less accurately represented in the final summary on the basis of (i) their speaking-order (speeches in the middle of the debate were systematically excluded), (ii) language spoken (non-English speakers were less faithfully represented), and (iii) political affiliations (better outcomes for left-of-centre parties). We further show how biases in these contexts can be decomposed to distinguish inclusion bias (systematic omission) from hallucination bias (systematic misrepresentation), and explore the effect of different mitigation strategies. Prompting strategies do not affect these biases. Instead, we propose a hierarchical summarisation method that decomposes the task into simpler extraction and aggregation steps, which we show significantly improves the positional/speaking-order bias across all models. These findings underscore the need for domain-sensitive evaluation metrics and ethical oversight in the deployment of LLMs for multilingual democratic applications.", "date": "2025-07-16", "categories": []}
{"id": "2511.18123", "title": "Bias Is a Subspace, Not a Coordinate: A Geometric Rethinking of Post-hoc Debiasing in Vision-Language Models", "abstract": "Vision-Language Models (VLMs) have become indispensable for multimodal reasoning, yet their representations often encode and amplify demographic biases, resulting in biased associations and misaligned predictions in downstream tasks. Such behavior undermines fairness and distorts the intended alignment between vision and language. Recent post-hoc approaches attempt to mitigate bias by replacing the most attribute-correlated embedding coordinates with neutral values. However, our systematic analysis reveals three critical limitations of this coordinate-wise approach: feature entanglement, poor cross-dataset generalization, and incomplete bias removal. We find that bias is not localized to a few coordinates but is instead distributed across a few linear subspaces. To address these limitations, we propose $\\textbf{S}$ubspace $\\textbf{P}$rojection $\\textbf{D}$ebiasing ($\\textbf{SPD}$), a geometrically principled framework that identifies and removes the entire subspace of linearly decodable bias while reinserting a neutral mean component to preserve semantic fidelity. Extensive experiments across zero-shot classification, text-to-image retrieval, and image generation validate the effectiveness of SPD: our method achieves more robust debiasing with an average improvement of $18.5\\%$ across four fairness metrics, while maintaining minimal loss in task performance compared to the best debiasing baseline.", "date": "2025-11-22", "categories": []}
{"id": "2601.05352", "title": "When the Server Steps In: Calibrated Updates for Fair Federated Learning", "abstract": "Federated learning (FL) has emerged as a transformative distributed learning paradigm, enabling multiple clients to collaboratively train a global model under the coordination of a central server without sharing their raw training data. While FL offers notable advantages, it faces critical challenges in ensuring fairness across diverse demographic groups. To address these fairness concerns, various fairness-aware debiasing methods have been proposed. However, many of these approaches either require modifications to clients' training protocols or lack flexibility in their aggregation strategies. In this work, we address these limitations by introducing EquFL, a novel server-side debiasing method designed to mitigate bias in FL systems. EquFL operates by allowing the server to generate a single calibrated update after receiving model updates from the clients. This calibrated update is then integrated with the aggregated client updates to produce an adjusted global model that reduces bias. Theoretically, we establish that EquFL converges to the optimal global model achieved by FedAvg and effectively reduces fairness loss over training rounds. Empirically, we demonstrate that EquFL significantly mitigates bias within the system, showcasing its practical effectiveness.", "date": "2026-01-08", "categories": []}
{"id": "2604.01925", "title": "ImplicitBBQ: Benchmarking Implicit Bias in Large Language Models through Characteristic Based Cues", "abstract": "Large Language Models increasingly suppress biased outputs when demographic identity is stated explicitly, yet may still exhibit implicit biases when identity is conveyed indirectly. Existing benchmarks use name based proxies to detect implicit biases, which carry weak associations with many social demographics and cannot extend to dimensions like age or socioeconomic status. We introduce ImplicitBBQ, a QA benchmark that evaluates implicit bias through characteristic based cues, culturally associated attributes that signal implicitly, across age, gender, region, religion, caste, and socioeconomic status. Evaluating 11 models, we find that implicit bias in ambiguous contexts is over six times higher than explicit bias in open weight models. Safety prompting and chain-of-thought reasoning fail to substantially close this gap; even few-shot prompting, which reduces implicit bias by 84%, leaves caste bias at four times the level of any other dimension. These findings indicate that current alignment and prompting strategies address the surface of bias evaluation while leaving culturally grounded stereotypic associations largely unresolved. We publicly release our code and dataset for model providers and researchers to benchmark potential mitigation techniques.", "date": "2026-04-02", "categories": []}
{"id": "2604.02017", "title": "Demographic Parity Tails for Regression", "abstract": "Demographic parity (DP) is a widely studied fairness criterion in regression, enforcing independence between the predictions and sensitive attributes. However, constraining the entire distribution can degrade predictive accuracy and may be unnecessary for many applications, where fairness concerns are localized to specific regions of the distribution. To overcome this issue, we propose a new framework for regression under DP that focuses on the tails of target distribution across sensitive groups. Our methodology builds on optimal transport theory. By enforcing fairness constraints only over targeted regions of the distribution, our approach enables more nuanced and context-sensitive interventions. Leveraging recent advances, we develop an interpretable and flexible algorithm that leverages the geometric structure of optimal transport. We provide theoretical guarantees, including risk bounds and fairness properties, and validate the method through experiments in regression settings.", "date": "2026-04-02", "categories": []}
{"id": "2604.00568", "title": "A Japanese Benchmark for Evaluating Social Bias in Reasoning Based on Attribution Theory", "abstract": "In enhancing the fairness of Large Language Models (LLMs), evaluating social biases rooted in the cultural contexts of specific linguistic regions is essential. However, most existing Japanese benchmarks heavily rely on translating English data, which does not necessarily provide an evaluation suitable for Japanese culture. Furthermore, they only evaluate bias in the conclusion, failing to capture biases lurking in the reasoning. In this study, based on attribution theory in social psychology, we constructed a new dataset, ``JUBAKU-v2,'' which evaluates the bias in attributing behaviors to in-groups and out-groups within reasoning while fixing the conclusion. This dataset consists of 216 examples reflecting cultural biases specific to Japan. Experimental results verified that it can detect performance differences across models more sensitively than existing benchmarks.", "date": "2026-04-01", "categories": []}
{"id": "2603.29288", "title": "Sima AIunty: Caste Audit in LLM-Driven Matchmaking", "abstract": "Social and personal decisions in relational domains such as matchmaking are deeply entwined with cultural norms and historical hierarchies, and can potentially be shaped by algorithmic and AI-mediated assessments of compatibility, acceptance, and stability. In South Asian contexts, caste remains a central aspect of marital decision-making, yet little is known about how contemporary large language models (LLMs) reproduce or disrupt caste-based stratification in such settings. In this work, we conduct a controlled audit of caste bias in LLM-mediated matchmaking evaluations using real-world matrimonial profiles. We vary caste identity across Brahmin, Kshatriya, Vaishya, Shudra, and Dalit, and income across five buckets, and evaluate five LLM families (GPT, Gemini, Llama, Qwen, and BharatGPT). Models are prompted to assess profiles along dimensions of social acceptance, marital stability, and cultural compatibility. Our analysis reveals consistent hierarchical patterns across models: same-caste matches are rated most favorably, with average ratings up to 25% higher (on a 10-point scale) than inter-caste matches, which are further ordered according to traditional caste hierarchy. These findings highlight how existing caste hierarchies are reproduced in LLM decision-making and underscore the need for culturally grounded evaluation and intervention strategies in AI systems deployed in socially sensitive domains, where such systems risk reinforcing historical forms of exclusion.", "date": "2026-03-31", "categories": []}
{"id": "2503.05371", "title": "Shifting Perspectives: Steering Vectors for Robust Bias Mitigation in LLMs", "abstract": "We present a novel approach to bias mitigation in large language models (LLMs) by applying steering vectors to modify model activations in forward passes. We compute 8 steering vectors, each corresponding to a different social bias axis, such as age, gender, or race, on a training subset of the BBQ dataset and compare the effectiveness of these to 3 additional bias mitigation methods across 4 datasets. When optimized on the BBQ dataset, our individually tuned steering vectors achieve average improvements of 12.8% on BBQ, 8.3% on CLEAR-Bias, and 1% on StereoSet, and show improvements over prompting and Self-Debias in all cases, and improvements over fine-tuning in 12 out of 17 evaluations. In addition, steering vectors showed the lowest impact on MMLU scores of the four bias mitigation methods tested. The work presents the first systematic investigation of steering vectors for bias mitigation, and we demonstrate that they are a powerful and computationally efficient strategy for reducing bias in LLMs, with broader implications for enhancing AI safety.", "date": "2025-03-07", "categories": []}
{"id": "2603.26476", "title": "Shapley meets Rawls: an integrated framework for measuring and explaining unfairness", "abstract": "Explainability and fairness have mainly been considered separately, with recent exceptions trying the explain the sources of unfairness. This paper shows that the Shapley value can be used to both define and explain unfairness, under standard group fairness criteria. This offers an integrated framework to estimate and derive inference on unfairness as-well-as the features that contribute to it. Our framework can also be extended from Shapley values to the family of Efficient-Symmetric-Linear (ESL) values, some of which offer more robust definitions of fairness, and shorter computation times. An illustration is run on the Census Income dataset from the UCI Machine Learning Repository. Our approach shows that ``Age\", ``Number of hours\" and ``Marital status\" generate gender unfairness, using shorter computation time than traditional Bootstrap tests.", "date": "2026-03-27", "categories": []}
{"id": "2603.27510", "title": "Decomposing Discrimination: Causal Mediation Analysis for AI-Driven Credit Decisions", "abstract": "Statistical fairness metrics in AI-driven credit decisions conflate two causally distinct mechanisms: discrimination operating directly from a protected attribute to a credit outcome, and structural inequality propagating through legitimate financial features. We formalise this distinction using Pearl's framework of natural direct and indirect effects applied to the credit decision setting. Our primary theoretical contribution is an identification strategy for natural direct and indirect effects under treatment-induced confounding -- the prevalent setting in which protected attributes causally affect both financial mediators and the final decision, violating standard sequential ignorability. We show that interventional direct and indirect effects (IDE/IIE) are identified under the weaker Modified Sequential Ignorability assumption, and prove that IDE/IIE provide conservative bounds on the unidentified natural effects under monotone indirect treatment response. We propose a doubly-robust augmented inverse probability weighted (AIPW) estimator for IDE/IIE with semiparametric efficiency properties, implemented via cross-fitting. An E-value sensitivity analysis addresses residual confounding on the direct pathway. Empirical evaluation on 89,465 real HMDA conventional purchase mortgage applications from New York State (2022) demonstrates that approximately 77% of the observed 7.9 percentage-point racial denial disparity operates through financial mediators shaped by structural inequality, while the remaining 23% constitutes a conservative lower bound on direct discrimination. The open-source CausalFair Python package implements the full pipeline for deployment at resource-constrained financial institutions.", "date": "2026-03-29", "categories": []}
{"id": "2603.28321", "title": "FairGC: Fairness-aware Graph Condensation", "abstract": "Graph condensation (GC) has become a vital strategy for scaling Graph Neural Networks by compressing massive datasets into small, synthetic node sets. While current GC methods effectively maintain predictive accuracy, they are primarily designed for utility and often ignore fairness constraints. Because these techniques are bias-blind, they frequently capture and even amplify demographic disparities found in the original data. This leads to synthetic proxies that are unsuitable for sensitive applications like credit scoring or social recommendations. To solve this problem, we introduce FairGC, a unified framework that embeds fairness directly into the graph distillation process. Our approach consists of three key components. First, a Distribution-Preserving Condensation module synchronizes the joint distributions of labels and sensitive attributes to stop bias from spreading. Second, a Spectral Encoding module uses Laplacian eigen-decomposition to preserve essential global structural patterns. Finally, a Fairness-Enhanced Neural Architecture employs multi-domain fusion and a label-smoothing curriculum to produce equitable predictions. Rigorous evaluations on four real-world datasets, show that FairGC provides a superior balance between accuracy and fairness. Our results confirm that FairGC significantly reduces disparity in Statistical Parity and Equal Opportunity compared to existing state-of-the-art condensation models. The codes are available at https://github.com/LuoRenqiang/FairGC.", "date": "2026-03-30", "categories": []}
{"id": "2603.26456", "title": "Fair Data Pre-Processing with Imperfect Attribute Space", "abstract": "Fair data pre-processing is a widely used strategy for mitigating bias in machine learning. A promising line of research focuses on calibrating datasets to satisfy a designed fairness policy so that sensitive attributes influence outcomes only through clearly specified legitimate causal pathways. While effective on clean and information-rich data, these methods often break down in real-world scenarios with imperfect attribute spaces, where decision-relevant factors may be deemed unusable or even missing. To address this gap, we propose LatentPre, a novel framework that enables principled and robust fair data processing in practical settings. Instead of relying solely on observed attributes, LatentPre augments the fairness policy with latent attributes that capture essential but subtle signals, enabling the framework to operate as if the attribute space were perfect. These latent attributes are strategically introduced to guarantee identifiability and are estimated using a tailored expectation-maximization paradigm. The raw data is then carefully refined to conform to this latent-augmented policy, effectively removing biased patterns while preserving justifiable ones. Extensive experiments demonstrate that LatentPre consistently achieves strong fairness-utility trade-offs across diverse scenarios, advancing practical fairness-aware data management.", "date": "2026-03-27", "categories": []}
{"id": "2603.25224", "title": "Fair regression under localized demographic parity constraints", "abstract": "Demographic parity (DP) is a widely used group fairness criterion requiring predictive distributions to be invariant across sensitive groups. While natural in classification, full distributional DP is often overly restrictive in regression and can lead to substantial accuracy loss. We propose a relaxation of DP tailored to regression, enforcing parity only at a finite set of quantile levels and/or score thresholds. Concretely, we introduce a novel (${\\ell}$, Z)-fair predictor, which imposes groupwise CDF constraints of the form F f |S=s (z m ) = ${\\ell}$ m for prescribed pairs (${\\ell}$ m , z m ). For this setting, we derive closed-form characterizations of the optimal fair discretized predictor via a Lagrangian dual formulation and quantify the discretization cost, showing that the risk gap to the continuous optimum vanishes as the grid is refined. We further develop a model-agnostic post-processing algorithm based on two samples (labeled for learning a base regressor and unlabeled for calibration), and establish finite-sample guarantees on constraint violation and excess penalized risk. In addition, we introduce two alternative frameworks where we match group and marginal CDF values at selected score thresholds. In both settings, we provide closed-form solutions for the optimal fair discretized predictor. Experiments on synthetic and real datasets illustrate an interpretable fairness-accuracy trade-off, enabling targeted corrections at decision-relevant quantiles or thresholds while preserving predictive performance.", "date": "2026-03-26", "categories": []}
{"id": "2603.25613", "title": "Demographic Fairness in Multimodal LLMs: A Benchmark of Gender and Ethnicity Bias in Face Verification", "abstract": "Multimodal Large Language Models (MLLMs) have recently been explored as face verification systems that determine whether two face images are of the same person. Unlike dedicated face recognition systems, MLLMs approach this task through visual prompting and rely on general visual and reasoning abilities. However, the demographic fairness of these models remains largely unexplored. In this paper, we present a benchmarking study that evaluates nine open-source MLLMs from six model families, ranging from 2B to 8B parameters, on the IJB-C and RFW face verification protocols across four ethnicity groups and two gender groups. We measure verification accuracy with the Equal Error Rate and True Match Rate at multiple operating points per demographic group, and we quantify demographic disparity with four FMR-based fairness metrics. Our results show that FaceLLM-8B, the only face-specialised model in our study, substantially outperforms general-purpose MLLMs on both benchmarks. The bias patterns we observe differ from those commonly reported for traditional face recognition, with different groups being most affected depending on the benchmark and the model. We also note that the most accurate models are not necessarily the fairest and that models with poor overall accuracy can appear fair simply because they produce uniformly high error rates across all demographic groups.", "date": "2026-03-26", "categories": []}
{"id": "2603.24218", "title": "Who Benefits from RAG? The Role of Exposure, Utility and Attribution Bias", "abstract": "Large Language Models (LLMs) enhanced with Retrieval-Augmented Generation (RAG) have achieved substantial improvements in accuracy by grounding their responses in external documents that are relevant to the user's query. However, relatively little work has investigated the impact of RAG in terms of fairness. Particularly, it is not yet known if queries that are associated with certain groups within a fairness category systematically receive higher accuracy, or accuracy improvements in RAG systems compared to LLM-only, a phenomenon we refer to as query group fairness. In this work, we conduct extensive experiments to investigate the impact of three key factors on query group fairness in RAG, namely: Group exposure, i.e., the proportion of documents from each group appearing in the retrieved set, determined by the retriever; Group utility, i.e., the degree to which documents from each group contribute to improving answer accuracy, capturing retriever-generator interactions; and Group attribution, i.e., the extent to which the generator relies on documents from each group when producing responses. We examine group-level average accuracy and accuracy improvements disparities across four fairness categories using three datasets derived from the TREC 2022 Fair Ranking Track for two tasks: article generation and title generation. Our findings show that RAG systems suffer from the query group fairness problem and amplify disparities in terms of average accuracy across queries from different groups, compared to an LLM-only setting. Moreover, group utility, exposure, and attribution can exhibit strong positive or negative correlations with average accuracy or accuracy improvements of queries from that group, highlighting their important role in fair RAG. Our data and code are publicly available from Github.", "date": "2026-03-25", "categories": []}
{"id": "2601.13698", "title": "Does Privacy Always Harm Fairness? Data-Dependent Trade-offs via Chernoff Information Neural Estimation", "abstract": "Fairness and privacy are two vital pillars of trustworthy machine learning. Despite extensive research on these individual topics, their relationship has received significantly less attention. In this paper, we utilize an information-theoretic measure Chernoff Information to characterize the fundamental trade-off between fairness, privacy, and accuracy, as induced by the input data distribution. We first propose Chernoff Difference, a notion of data fairness, along with its noisy variant, Noisy Chernoff Difference, which allows us to analyze both fairness and privacy simultaneously. Through simple Gaussian examples, we show that Noisy Chernoff Difference exhibits three qualitatively distinct behaviors depending on the underlying data distribution. To extend this analysis beyond synthetic settings, we develop the Chernoff Information Neural Estimator (CINE), the first neural network-based estimator of Chernoff Information for unknown distributions. We apply CINE to analyze the Noisy Chernoff Difference on real-world datasets. Together, this work fills a critical gap in the literature by providing a principled, data-dependent characterization of the fairness-privacy interaction.", "date": "2026-01-20", "categories": []}
{"id": "2603.22714", "title": "PopResume: Causal Fairness Evaluation of LLM/VLM Resume Screeners with Population-Representative Dataset", "abstract": "We present PopResume, a population-representative resume dataset for causal fairness auditing of LLM- and VLM-based resume screening systems. Unlike existing benchmarks that rely on manually injected demographic information and outcome-level disparities, PopResume is grounded in population statistics and preserves natural attribute relationships, enabling path-specific effect (PSE)-based fairness evaluation. We decompose the effect of a protected attribute on resume scores into two paths: the business necessity path, mediated by job-relevant qualifications, and the redlining path, mediated by demographic proxies. This distinction allows auditors to separate legally permissible from impermissible sources of disparity. Evaluating four LLMs and four VLMs on PopResume's 60.8K resumes across five occupations, we identify five representative discrimination patterns that aggregate metrics fail to capture. Our results demonstrate that PSE-based evaluation reveals fairness issues masked by outcome-level measures, underscoring the need for causally-grounded auditing frameworks in AI-assisted hiring.", "date": "2026-03-24", "categories": []}
{"id": "2603.22837", "title": "Analysing LLM Persona Generation and Fairness Interpretation in Polarised Geopolitical Contexts", "abstract": "Large language models (LLMs) are increasingly utilised for social simulation and persona generation, necessitating an understanding of how they represent geopolitical identities. In this paper, we analyse personas generated for Palestinian and Israeli identities by five popular LLMs across 640 experimental conditions, varying context (war vs non-war) and assigned roles. We observe significant distributional patterns in the generated attributes: Palestinian profiles in war contexts are frequently associated with lower socioeconomic status and survival-oriented roles, whereas Israeli profiles predominantly retain middle-class status and specialised professional attributes. When prompted with explicit instructions to avoid harmful assumptions, models exhibit diverse distributional changes, e.g., marked increases in non-binary gender inferences or a convergence toward generic occupational roles (e.g., \"student\"), while the underlying socioeconomic distinctions often remain. Furthermore, analysis of reasoning traces reveals an interesting dynamics between model reasoning and generation: while rationales consistently mention fairness-related concepts, the final generated personas follow the aforementioned diverse distributional changes. These findings illustrate a picture of how models interpret geopolitical contexts, while suggesting that they process fairness and adjust in varied ways; there is no consistent, direct translation of fairness concepts into representative outcomes.", "date": "2026-03-24", "categories": []}
{"id": "2512.19735", "title": "Improving Fairness of Large Language Model-Based ICU Mortality Prediction via Case-Based Prompting", "abstract": "Accurately predicting mortality risk in intensive care unit (ICU) patients is essential for clinical decision-making. Although large language models (LLMs) show strong potential in structured medical prediction tasks, their outputs may exhibit biases related to demographic attributes such as sex, age, and race, limiting their reliability in fairness-critical clinical settings. Existing debiasing methods often degrade predictive performance, making it difficult to balance fairness and accuracy.\n  In this study, we systematically analyze fairness issues in LLM-based ICU mortality prediction and propose a clinically adaptive prompting framework that improves both performance and fairness without model retraining. We first design a multi-dimensional bias assessment scheme to identify subgroup disparities. Based on this, we introduce CAse Prompting (CAP), a training-free framework that integrates existing debiasing strategies and further guides models using similar historical misprediction cases paired with correct outcomes to correct biased reasoning.\n  We evaluate CAP on the MIMIC-IV dataset. Results show that AUROC improves from 0.806 to 0.873 and AUPRC from 0.497 to 0.694. Meanwhile, prediction disparities are substantially reduced across demographic groups, with reductions exceeding 90% in sex and certain White-Black comparisons. Feature reliance analysis further reveals highly consistent attention patterns across groups, with similarity above 0.98.\n  These findings demonstrate that fairness and performance in LLM-based clinical prediction can be jointly optimized through carefully designed prompting, offering a practical paradigm for developing reliable and equitable clinical decision-support systems.", "date": "2025-12-17", "categories": []}
{"id": "2601.18486", "title": "Different Demographic Cues Yield Inconsistent Conclusions About LLM Personalization and Bias", "abstract": "Demographic cue-based evaluation is widely used to study how large language models (LLMs) adapt their responses to signaled demographic attributes within and across groups. This approach typically relies on a single cue (e.g., names) as a proxy for group membership, implicitly treating different cues as interchangeable operationalizations of the same identity-conditioned behavior. We test this assumption in realistic advice-seeking interactions spanning 14.8 million prompts, focusing on race and gender in a U.S. context. We find that cues for the same group induce only partially overlapping changes in model responses, yielding inconsistent conclusions about personalization, while bias conclusions are unstable, with both magnitude and direction of group differences varying across cues. We further show that these inconsistencies reflect differences in cue-group association strength and linguistic features bundled within cues that shape model responses. Together, our findings suggest that demographic conditioning in LLMs is not a cue-invariant category-level parameter but depends fundamentally on how identity is cued, reflecting responses to linguistic signals rather than stable demographic categories. We therefore advocate multi-cue, mechanism-aware evaluations for robust and interpretable claims about demographic variation in LLM responses.", "date": "2026-01-26", "categories": []}
{"id": "2603.20254", "title": "AI Detectors Fail Diverse Student Populations: A Mathematical Framing of Structural Detection Limits", "abstract": "Student experiences and empirical studies report that \"black box\" AI text detectors produce high false positive rates with disproportionate errors against certain student populations, yet typically theoretical analyses model detection as a test between two known distributions for human and AI prose. This framing omits the structural feature of university assessment whereby an assessor generally does not know the individual student's writing distribution, making the null hypothesis composite. Standard application of the variational characterisation of total variation distance to this composite null shows trade-off bounds that any text-only, one-shot detector with useful power must produce false accusations at a rate governed by the distributional overlap between student writing and AI output. This is a constraint arising from population diversity that is logically independent of AI model quality and cannot be overcome by better detector engineering or technology. A subgroup mixture bound connects these quantities to observable demographic groups, providing a theoretical basis for the disparate impact patterns documented empirically. We propose suggestions to improve policy and practice, and argue that detection scores should not serve as sole evidence in misconduct proceedings.", "date": "2026-03-11", "categories": []}
{"id": "2603.21393", "title": "A Generalised Exponentiated Gradient Approach to Enhance Fairness in Binary and Multi-class Classification Tasks", "abstract": "The widespread use of AI and ML models in sensitive areas raises significant concerns about fairness. While the research community has introduced various methods for bias mitigation in binary classification tasks, the issue remains under-explored in multi-class classification settings. To address this limitation, in this paper, we first formulate the problem of fair learning in multi-class classification as a multi-objective problem between effectiveness (i.e., prediction correctness) and multiple linear fairness constraints. Next, we propose a Generalised Exponentiated Gradient (GEG) algorithm to solve this task. GEG is an in-processing algorithm that enhances fairness in binary and multi-class classification settings under multiple fairness definitions. We conduct an extensive empirical evaluation of GEG against six baselines across seven multi-class and three binary datasets, using four widely adopted effectiveness metrics and three fairness definitions. GEG overcomes existing baselines, with fairness improvements up to 92% and a decrease in accuracy up to 14%.", "date": "2026-03-22", "categories": []}
{"id": "2603.21524", "title": "CatRAG: Functor-Guided Structural Debiasing with Retrieval Augmentation for Fair LLMs", "abstract": "Large Language Models (LLMs) are deployed in high-stakes settings but can show demographic, gender, and geographic biases that undermine fairness and trust. Prior debiasing methods, including embedding-space projections, prompt-based steering, and causal interventions, often act at a single stage of the pipeline, resulting in incomplete mitigation and brittle utility trade-offs under distribution shifts. We propose CatRAG Debiasing, a dual-pronged framework that integrates functor with Retrieval-Augmented Generation (RAG) guided structural debiasing. The functor component leverages category-theoretic structure to induce a principled, structure-preserving projection that suppresses bias-associated directions in the embedding space while retaining task-relevant semantics. On the Bias Benchmark for Question Answering (BBQ) across three open-source LLMs (Meta Llama-3, OpenAI GPT-OSS, and Google Gemma-3), CatRAG achieves state-of-the-art results, improving accuracy by up to 40% over the corresponding base models and by more than 10% over prior debiasing methods, while reducing bias scores to near zero (from 60% for the base models) across gender, nationality, race, and intersectional subgroups.", "date": "2026-03-23", "categories": []}
{"id": "2603.18987", "title": "Unmasking Algorithmic Bias in Predictive Policing: A GAN-Based Simulation Framework with Multi-City Temporal Analysis", "abstract": "Predictive policing systems that direct patrol resources based on algorithmically generated crime forecasts have been widely deployed across US cities, yet their tendency to encode and amplify racial disparities remains poorly understood in quantitative terms. We present a reproducible simulation framework that couples a Generative Adversarial Network GAN with a Noisy OR patrol detection model to measure how racial bias propagates through the full enforcement pipeline from crime occurrence to police contact. Using 145000 plus Part 1 crime records from Baltimore 2017 to 2019 and 233000 plus records from Chicago 2022, augmented with US Census ACS demographic data, we compute four monthly bias metrics across 264 city year mode observations: the Disparate Impact Ratio DIR, Demographic Parity Gap, Gini Coefficient, and a composite Bias Amplification Score.\n  Our experiments reveal extreme and year variant bias in Baltimores detected mode, with mean annual DIR up to 15714 in 2019, moderate under detection of Black residents in Chicago DIR equals 0.22, and persistent Gini coefficients of 0.43 to 0.62 across all conditions. We further demonstrate that a Conditional Tabular GAN CTGAN debiasing approach partially redistributes detection rates but cannot eliminate structural disparity without accompanying policy intervention. Socioeconomic regression analysis confirms strong correlations between neighborhood racial composition and detection likelihood Pearson r equals 0.83 for percent White and r equals negative 0.81 for percent Black. A sensitivity analysis over patrol radius, officer count, and citizen reporting probability reveals that outcomes are most sensitive to officer deployment levels. The code and data are publicly available at this repository.", "date": "2026-03-19", "categories": []}
{"id": "2506.10586", "title": "Size-adaptive Hypothesis Testing for Fairness", "abstract": "Determining whether an algorithmic decision-making system discriminates against a specific demographic typically involves comparing a single point estimate of a fairness metric against a predefined threshold. This practice is statistically brittle: it ignores sampling error and treats small demographic subgroups the same as large ones. The problem intensifies in intersectional analyses, where multiple sensitive attributes are considered jointly, giving rise to a larger number of smaller groups. As these groups become more granular, the data representing them becomes too sparse for reliable estimation, and fairness metrics yield excessively wide confidence intervals, precluding meaningful conclusions about potential unfair treatments.\n  In this paper, we introduce a unified, size-adaptive, hypothesis-testing framework that turns fairness assessment into an evidence-based statistical decision. Our contribution is twofold. (i) For sufficiently large subgroups, we prove a Central-Limit result for the statistical parity difference, leading to analytic confidence intervals and a Wald test whose type-I (false positive) error is guaranteed at level $\\alpha$. (ii) For the long tail of small intersectional groups, we derive a fully Bayesian Dirichlet-multinomial estimator; Monte-Carlo credible intervals are calibrated for any sample size and naturally converge to Wald intervals as more data becomes available. We validate our approach empirically on benchmark datasets, demonstrating how our tests provide interpretable, statistically rigorous decisions under varying degrees of data availability and intersectionality.", "date": "2025-06-12", "categories": []}
{"id": "2511.20909", "title": "Evolved Sample Weights for Bias Mitigation: Effectiveness Depends on the Fairness Objective", "abstract": "Machine learning models trained on real-world data may inadvertently make biased predictions that negatively impact marginalized communities. Reweighting, which assigns a weight to each data point used during model training, can mitigate such bias, though sometimes at the cost of predictive accuracy. In this paper, we investigated this trade-off by comparing three methods for generating these weights: (1) evolving them using a Genetic Algorithm (GA), (2) computing them using only dataset characteristics, and (3) assigning equal weights to all data points. Model performance under each strategy was evaluated using paired predictive and fairness metrics. We used two predictive metrics (accuracy and area under the Receiver Operating Characteristic curve) and two fairness metrics (demographic parity and subgroup false negative fairness). By conducting experiments on eleven publicly available datasets (including two medical datasets), we show that evolved sample weights can produce models that achieve better trade-offs between fairness and predictive performance than alternative weighting methods. However, the magnitude of these benefits depends strongly on the choice of fairness objective. Our experiments reveal that the evolved weights were most effective when optimizing for demographic parity-independent of choice of the performance objective-yielding better performance than other weighting strategies on the largest number of datasets.", "date": "2025-11-25", "categories": []}
{"id": "2603.18530", "title": "When Names Change Verdicts: Intervention Consistency Reveals Systematic Bias in LLM Decision-Making", "abstract": "Large language models (LLMs) are increasingly used for high-stakes decisions, yet their susceptibility to spurious features remains poorly characterized. We introduce ICE-Guard, a framework applying intervention consistency testing to detect three types of spurious feature reliance: demographic (name/race swaps), authority (credential/prestige swaps), and framing (positive/negative restatements). Across 3,000 vignettes spanning 10 high-stakes domains, we evaluate 11 LLMs from 8 families and find that (1) authority bias (mean 5.8%) and framing bias (5.0%) substantially exceed demographic bias (2.2%), challenging the field's narrow focus on demographics; (2) bias concentrates in specific domains -- finance shows 22.6% authority bias while criminal justice shows only 2.8%; (3) structured decomposition, where the LLM extracts features and a deterministic rubric decides, reduces flip rates by up to 100% (median 49% across 9 models). We demonstrate an ICE-guided detect-diagnose-mitigate-verify loop achieving cumulative 78% bias reduction via iterative prompt patching. Validation against real COMPAS recidivism data shows COMPAS-derived flip rates exceed pooled synthetic rates, suggesting our benchmark provides a conservative estimate of real-world bias. Code and data are publicly available.", "date": "2026-03-19", "categories": []}
{"id": "2603.18729", "title": "Analysis Of Linguistic Stereotypes in Single and Multi-Agent Generative AI Architectures", "abstract": "Many works in the literature show that LLM outputs exhibit discriminatory behaviour, triggering stereotype-based inferences based on the dialect in which the inputs are written. This bias has been shown to be particularly pronounced when the same inputs are provided to LLMs in Standard American English (SAE) and African-American English (AAE). In this paper, we replicate existing analyses of dialect-sensitive stereotype generation in LLM outputs and investigate the effects of mitigation strategies, including prompt engineering (role-based and Chain-Of-Thought prompting) and multi-agent architectures composed of generate-critique-revise models. We define eight prompt templates to analyse different ways in which dialect bias can manifest, such as suggested names, jobs, and adjectives for SAE or AAE speakers. We use an LLM-as-judge approach to evaluate the bias in the results. Our results show that stereotype-bearing differences emerge between SAE- and AAE-related outputs across all template categories, with the strongest effects observed in adjective and job attribution. Baseline disparities vary substantially by model, with the largest SAE-AAE differential observed in Claude Haiku and the smallest in Phi-4 Mini. Chain-Of-Thought prompting proved to be an effective mitigation strategy for Claude Haiku, whereas the use of a multi-agent architecture ensured consistent mitigation across all the models. These findings suggest that for intersectionality-informed software engineering, fairness evaluation should include model-specific validation of mitigation strategies, and workflow-level controls (e.g., agentic architectures involving critique models) in high-impact LLM deployments. The current results are exploratory in nature and limited in scope, but can lead to extensions and replications by increasing the dataset size and applying the procedure to different languages or dialects.", "date": "2026-03-19", "categories": []}
{"id": "2603.18838", "title": "A Model Ensemble-Based Post-Processing Framework for Fairness-Aware Prediction", "abstract": "Striking an optimal balance between predictive performance and fairness continues to be a fundamental challenge in machine learning. In this work, we propose a post-processing framework that facilitates fairness-aware prediction by leveraging model ensembling. Designed to operate independently of any specific model internals, our approach is widely applicable across various learning tasks, model architectures, and fairness definitions. Through extensive experiments spanning classification, regression, and survival analysis, we demonstrate that the framework effectively enhances fairness while maintaining, or only minimally affecting, predictive accuracy.", "date": "2026-03-19", "categories": []}
{"id": "2404.19725", "title": "CurvFed: Curvature-Aligned Federated Learning for Fairness without Demographics", "abstract": "Modern human sensing applications often rely on data distributed across users and devices, where privacy concerns prevent centralized training. Federated Learning (FL) addresses this challenge by enabling collaborative model training without exposing raw data or attributes. However, achieving fairness in such settings remains difficult, as most human sensing datasets lack demographic labels, and FL's privacy guarantees limit the use of sensitive attributes. This paper introduces CurvFed: Curvature Aligned Federated Learning for Fairness without Demographics, a theoretically grounded framework that promotes fairness in FL without requiring any demographic or sensitive attribute information, a concept termed Fairness without Demographics (FWD), by optimizing the underlying loss landscape curvature. Building on the theory that equivalent loss landscape curvature corresponds to consistent model efficacy across sensitive attribute groups, CurvFed regularizes the top eigenvalue of the Fisher Information Matrix (FIM) as an efficient proxy for loss landscape curvature, both within and across clients. This alignment promotes uniform model behavior across diverse bias inducing factors, offering an attribute agnostic route to algorithmic fairness. CurvFed is especially suitable for real world human sensing FL scenarios involving single or multi user edge devices with unknown or multiple bias factors. We validated CurvFed through theoretical and empirical justifications, as well as comprehensive evaluations using three real world datasets and a deployment on a heterogeneous testbed of resource constrained devices. Additionally, we conduct sensitivity analyses on local training data volume, client sampling, communication overhead, resource costs, and runtime performance to demonstrate its feasibility for practical FL edge device deployment.", "date": "2024-04-30", "categories": []}
{"id": "2601.17054", "title": "Failing on Bias Mitigation: A Case Study on the Challenges of Fairness in Government Data", "abstract": "The potential for bias and unfairness in AI-supporting government services raises ethical and legal concerns. Using crime rate prediction with the Bristol City Council data as a case study, we examine how these issues persist. Rather than auditing real-world deployed systems, our goal is to understand why widely adopted bias mitigation techniques often fail when applied to government data. Our findings reveal that bias mitigation approaches applied to government data are not always effective -- not because of flaws in model architecture or metric selection, but due to the inherent properties of the data itself. Through comparing a set of comprehensive models and fairness methods, our experiments consistently show that the mitigation efforts cannot overcome the embedded unfairness in the data -- further reinforcing that the origin of bias lies in the structure and history of government datasets. We then explore the reasons for the mitigation failures in predictive models on government data and highlight the potential sources of unfairness posed by data distribution shifts, the accumulation of historical bias, and delays in data release. We also discover the limitations of the blind spots in fairness analysis and bias mitigation methods when only targeting a single sensitive feature through a set of intersectional fairness experiments. Although this study is limited to one city, the findings are highly suggestive, which can contribute to an early warning that biases in government data may persist even with standard mitigation methods.", "date": "2026-01-21", "categories": []}
{"id": "2305.13883", "title": "Leveraging Imperfect Sources to Detect Fairwashing in Black-Box Auditing", "abstract": "Algorithmic auditing has become central to platform accountability under frameworks such as the AI Act and the Digital Services Act. In practice, this obligation is discharged through dedicated Audit APIs. This architecture creates a paradox: the entity under scrutiny controls the evaluation interface. A platform facing legal sanctions can serve a compliant surrogate model on its Audit API, while running a discriminatory production system. This deceptive practice is known as fairwashing. Manipulation is undetectable if the auditor relies on only one source. To address this limitation, we introduce the Two-Source Audit Model (2SAM). This model cross-references the Audit API with an independent trusted stream. The key insight is that the trusted stream does not need to be perfectly aligned with the Audit API. We introduce a consistency proxy, a probabilistic mapping that can reconcile discrepancies between sources. This approach yields three results. First, we quantify the rate of manipulation above which a single-source auditor is blind. Second, we show how proxy quality governs detection power. Third, we provide a closed-form budget condition guaranteeing detection at any target confidence level, closing the blind spot mentioned above. We validate 2SAM on the UCI Adult dataset, achieving $70\\%$ detection power with as few as $127$ cross-verification queries out of a total budget of $750$, using a name-based gender proxy with $94.2\\%$ accuracy.", "date": "2023-05-23", "categories": []}
{"id": "2601.16926", "title": "Nishpaksh: TEC Standard-Compliant Framework for Fairness Auditing and Certification of AI Models", "abstract": "The growing reliance on Artificial Intelligence (AI) models in high-stakes decision-making systems, particularly within emerging telecom and 6G applications, underscores the urgent need for transparent and standardized fairness assessment frameworks. While global toolkits such as IBM AI Fairness 360 and Microsoft Fairlearn have advanced bias detection, they often lack alignment with region-specific regulatory requirements and national priorities. To address this gap, we propose Nishpaksh, an indigenous fairness evaluation tool that operationalizes the Telecommunication Engineering Centre (TEC) Standard for the Evaluation and Rating of Artificial Intelligence Systems. Nishpaksh integrates survey-based risk quantification, contextual threshold determination, and quantitative fairness evaluation into a unified, web-based dashboard. The tool employs vectorized computation, reactive state management, and certification-ready reporting to enable reproducible, audit-grade assessments, thereby addressing a critical post-standardization implementation need. Experimental validation on the COMPAS dataset demonstrates Nishpaksh's effectiveness in identifying attribute-specific bias and generating standardized fairness scores compliant with the TEC framework. The system bridges the gap between research-oriented fairness methodologies and regulatory AI governance in India, marking a significant step toward responsible and auditable AI deployment within critical infrastructure like telecommunications.", "date": "2026-01-23", "categories": []}
{"id": "2302.08018", "title": "Towards Fair Machine Learning Software: Understanding and Addressing Model Bias Through Counterfactual Thinking", "abstract": "The increasing use of Machine Learning (ML) software can lead to unfair and unethical decisions, thus fairness bugs in software are becoming a growing concern. Addressing these fairness bugs often involves sacrificing ML performance, such as accuracy. To address this issue, we present a novel counterfactual approach that uses counterfactual thinking to tackle the root causes of bias in ML software. In addition, our approach combines models optimized for both performance and fairness, resulting in an optimal solution in both aspects. We conducted a thorough evaluation of our approach on 10 benchmark tasks using a combination of 5 performance metrics, 3 fairness metrics, and 15 measurement scenarios, all applied to 8 real-world datasets. The conducted extensive evaluations show that the proposed method significantly improves the fairness of ML software while maintaining competitive performance, outperforming state-of-the-art solutions in 84.6% of overall cases based on a recent benchmarking tool.", "date": "2023-02-16", "categories": []}
{"id": "2501.15338", "title": "Fairness-aware Contextual Dynamic Pricing with Strategic Buyers", "abstract": "Contextual pricing strategies are prevalent in online retailing, where the seller adjusts prices based on products' attributes and buyers' characteristics. Although such strategies can enhance seller's profits, they raise concerns about fairness when significant price disparities emerge among specific groups, such as gender or race. These disparities can lead to adverse perceptions of fairness among buyers and may even violate the law and regulation. In contrast, price differences can incentivize disadvantaged buyers to strategically manipulate their group identity to obtain a lower price. In this paper, we investigate contextual dynamic pricing with fairness constraints, taking into account buyers' strategic behaviors when their group status is private and unobservable from the seller. We propose a dynamic pricing policy that simultaneously achieves price fairness and discourages strategic behaviors. Our policy achieves an upper bound of $O(\\sqrt{T}+H(T))$ regret over $T$ time horizons, where the term $H(T)$ arises from buyers' assessment of the fairness of the pricing policy based on their learned price difference. When buyers are able to learn the fairness of the price policy, this upper bound reduces to $O(\\sqrt{T})$. We also prove an $\\Omega(\\sqrt{T})$ regret lower bound of any pricing policy under our problem setting. We support our findings with extensive experimental evidence, showcasing our policy's effectiveness. In our real data analysis, we observe the existence of price discrimination against race in the loan application even after accounting for other contextual information. Our proposed pricing policy demonstrates a significant improvement, achieving 35.06% reduction in regret compared to the benchmark policy.", "date": "2025-01-25", "categories": []}
{"id": "2603.11750", "title": "Mitigating the Multiplicity Burden: The Role of Calibration in Reducing Predictive Multiplicity of Classifiers", "abstract": "As machine learning models are increasingly deployed in high-stakes environments, ensuring both probabilistic reliability and prediction stability has become critical. This paper examines the interplay between classification calibration and predictive multiplicity - the phenomenon in which multiple near-optimal models within the Rashomon set yield conflicting credit outcomes for the same applicant. Using nine diverse credit risk benchmark datasets, we investigate whether predictive multiplicity concentrates in regions of low predictive confidence and how post-hoc calibration can mitigate algorithmic arbitrariness. Our empirical analysis reveals that minority class observations bear a disproportionate multiplicity burden, as confirmed by significant disparities in predictive multiplicity and prediction confidence. Furthermore, our empirical comparisons indicate that applying post-hoc calibration methods - specifically Platt Scaling, Isotonic Regression, and Temperature Scaling - is associated with lower obscurity across the Rashomon set. Among the tested techniques, Platt Scaling and Isotonic Regression provide the most robust reduction in predictive multiplicity. These findings suggest that calibration can function as a consensus-enforcing layer and may support procedural fairness by mitigating predictive multiplicity.", "date": "2026-03-12", "categories": []}
{"id": "2603.13253", "title": "A Counterfactual Approach for Addressing Individual User Unfairness in Collaborative Recommender System", "abstract": "Recommender Systems (RSs) are exploited by various business enterprises to suggest their products (items) to consumers (users). Collaborative filtering (CF) is a widely used variant of RSs which learns hidden patterns from user-item interactions for recommending items to users. Recommendations provided by the traditional CF models are often biased. Generally, such models learn and update embeddings for all the users, thereby overlooking the biases toward each under-served users individually. This leads to certain users receiving poorer recommendations than the rest. Such unfair treatment toward users incur loss to the business houses. There is limited research which addressed individual user unfairness problem (IUUP). Existing literature employed explicit exploration-based multi-armed bandits, individual user unfairness metric, and explanation score to address this issue. Although, these works elucidate and identify the underlying individual user unfairness, however, they do not provide solutions for it. In this paper, we propose a dual-step approach which identifies and mitigates IUUP in recommendations. In the proposed work, we counterfactually introduce new interactions to the candidate users (one at a time) and subsequently analyze the benefit from this perturbation. This improves the user engagement with other users and items. Thus, the model can learn effective embeddings across the users. To showcase the effectiveness of the proposed counterfactual methodology, we conducted experiments on MovieLens-100K, Amazon Beauty and MovieLens-1M datasets. The experimental results validate the superiority of the proposed approach over the existing techniques.", "date": "2026-02-24", "categories": []}
{"id": "2603.14631", "title": "Anterior's Approach to Fairness Evaluation of Automated Prior Authorization System", "abstract": "Increasing staffing constraints and turnaround-time pressures in Prior authorization (PA) have led to increasing automation of decision systems to support PA review. Evaluating fairness in such systems poses unique challenges because legitimate clinical guidelines and medical necessity criteria often differ across demographic groups, making parity in approval rates an inappropriate fairness metric. We propose a fairness evaluation framework for prior authorization models based on model error rates rather than approval outcomes. Using 7,166 human-reviewed cases spanning 27 medical necessity guidelines, we assessed consistency in sex, age, race/ethnicity, and socioeconomic status. Our evaluation combined error-rate comparisons, tolerance-band analysis with a predefined $\\pm$5 percentage-point margin, statistical power evaluation, and protocol-controlled logistic regression. Across most demographics, model error rates were consistent, and confidence intervals fell within the predefined tolerance band, indicating no meaningful performance differences. For race/ethnicity, point estimates remain small, but subgroup sample sizes were limited, resulting in wide confidence intervals and underpowered tests, with inconclusive evidence within the dataset we explored. These findings illustrate a rigorous and regulator-aligned approach to fairness evaluation in administrative healthcare AI systems.", "date": "2026-03-15", "categories": []}
{"id": "2603.14947", "title": "FairMed-XGB: A Bayesian-Optimised Multi-Metric Framework with Explainability for Demographic Equity in Critical Healthcare Data", "abstract": "Machine learning models deployed in critical care settings exhibit demographic biases, particularly gender disparities, that undermine clinical trust and equitable treatment. This paper introduces FairMed-XGB, a novel framework that systematically detects and mitigates gender-based prediction bias while preserving model performance and transparency. The framework integrates a fairness-aware loss function combining Statistical Parity Difference, Theil Index, and Wasserstein Distance, jointly optimised via Bayesian Search into an XGBoost classifier. Post-mitigation evaluation on seven clinically distinct cohorts derived from the MIMIC-IV-ED and eICU databases demonstrates substantial bias reduction: Statistical Parity Difference decreases by 40 to 51 percent on MIMIC-IV-ED and 10 to 19 percent on eICU; Theil Index collapses by four to five orders of magnitude to near-zero values; Wasserstein Distance is reduced by 20 to 72 percent. These gains are achieved with negligible degradation in predictive accuracy (AUC-ROC drop <0.02). SHAP-based explainability reveals that the framework diminishes reliance on gender-proxy features, providing clinicians with actionable insights into how and where bias is corrected. FairMed-XGB offers a robust, interpretable, and ethically aligned solution for equitable clinical decision-making, paving the way for trustworthy deployment of AI in high-stakes healthcare environments.", "date": "2026-03-16", "categories": []}
{"id": "2505.20343", "title": "Do LLMs have a Gender (Entropy) Bias?", "abstract": "We investigate the existence and persistence of a specific type of gender bias in some of the popular LLMs and contribute a new benchmark dataset, RealWorldQuestioning (released on HuggingFace ), developed from real-world questions across four key domains in business and health contexts: education, jobs, personal financial management, and general health. We define and study entropy bias, which we define as a discrepancy in the amount of information generated by an LLM in response to real questions users have asked. We tested this using four different LLMs and evaluated the generated responses both qualitatively and quantitatively by using ChatGPT-4o (as \"LLM-as-judge\"). Our analyses (metric-based comparisons and \"LLM-as-judge\" evaluation) suggest that there is no significant bias in LLM responses for men and women at a category level. However, at a finer granularity (the individual question level), there are substantial differences in LLM responses for men and women in the majority of cases, which \"cancel\" each other out often due to some responses being better for males and vice versa. This is still a concern since typical users of these tools often ask a specific question (only) as opposed to several varied ones in each of these common yet important areas of life. We suggest a simple debiasing approach that iteratively merges the responses for the two genders to produce a final result. Our approach demonstrates that a simple, prompt-based debiasing strategy can effectively debias LLM outputs, thus producing responses with higher information content than both gendered variants in 78% of the cases, and consistently achieving a balanced integration in the remaining cases.", "date": "2025-05-24", "categories": []}
{"id": "2510.08724", "title": "Counterfactually Fair Conformal Prediction", "abstract": "While counterfactual fairness of point predictors is well studied, its extension to prediction sets--central to fair decision-making under uncertainty--remains underexplored. On the other hand, conformal prediction (CP) provides efficient, distribution-free, finite-sample valid prediction sets, yet does not ensure counterfactual fairness. We close this gap by developing Counterfactually Fair Conformal Prediction (CF-CP) that produces counterfactually fair prediction sets. Through symmetrization of conformity scores across protected-attribute interventions, we prove that CF-CP results in counterfactually fair prediction sets while maintaining the marginal coverage property. Furthermore, we empirically demonstrate that on both synthetic and real datasets, across regression and classification tasks, CF-CP achieves the desired counterfactual fairness and meets the target coverage rate with minimal increase in prediction set size. CF-CP offers a simple, training-free route to counterfactually fair uncertainty quantification.", "date": "2025-10-09", "categories": []}
{"id": "2603.11140", "title": "Procedural Fairness via Group Counterfactual Explanation", "abstract": "Fairness in machine learning research has largely focused on outcome-oriented fairness criteria such as Equalized Odds, while comparatively less attention has been given to procedural-oriented fairness, which addresses how a model arrives at its predictions. Neglecting procedural fairness means it is possible for a model to generate different explanations for different protected groups, thereby eroding trust. In this work, we introduce Group Counterfactual Integrated Gradients (GCIG), an in-processing regularization framework that enforces explanation invariance across groups, conditioned on the true label. For each input, GCIG computes explanations relative to multiple Group Conditional baselines and penalizes cross-group variation in these attributions during training. GCIG formalizes procedural fairness as Group Counterfactual explanation stability and complements existing fairness objectives that constrain predictions alone. We compared GCIG empirically against six state-of-the-art methods, and the results show that GCIG substantially reduces cross-group explanation disparity while maintaining competitive predictive performance and accuracy-fairness trade-offs. Our results also show that aligning model reasoning across groups offers a principled and practical avenue for advancing fairness beyond outcome parity.", "date": "2026-03-11", "categories": []}
{"id": "2603.11736", "title": "Gender Bias in Generative AI-assisted Recruitment Processes", "abstract": "In recent years, generative artificial intelligence (GenAI) systems have assumed increasingly crucial roles in selection processes, personnel recruitment and analysis of candidates' profiles. However, the employment of large language models (LLMs) risks reproducing, and in some cases amplifying, gender stereotypes and bias already present in the labour market. The objective of this paper is to evaluate and measure this phenomenon, analysing how a state-of-the-art generative model (GPT-5) suggests occupations based on gender and work experience background, focusing on under-35-year-old Italian graduates. The model has been prompted to suggest jobs to 24 simulated candidate profiles, which are balanced in terms of gender, age, experience and professional field. Although no significant differences emerged in job titles and industry, gendered linguistic patterns emerged in the adjectives attributed to female and male candidates, indicating a tendency of the model to associate women with emotional and empathetic traits, while men with strategic and analytical ones. The research raises an ethical question regarding the use of these models in sensitive processes, highlighting the need for transparency and fairness in future digital labour markets.", "date": "2026-03-12", "categories": []}
{"id": "2603.11793", "title": "Locating Demographic Bias at the Attention-Head Level in CLIP's Vision Encoder", "abstract": "Standard fairness audits of foundation models quantify that a model is biased, but not where inside the network the bias resides. We propose a mechanistic fairness audit that combines projected residual-stream decomposition, zero-shot Concept Activation Vectors, and bias-augmented TextSpan analysis to locate demographic bias at the level of individual attention heads in vision transformers. As a feasibility case study, we apply this pipeline to the CLIP ViT-L-14 encoder on 42 profession classes of the FACET benchmark, auditing both gender and age bias. For gender, the pipeline identifies four terminal-layer heads whose ablation reduces global bias (Cramer's V: 0.381 -> 0.362) while marginally improving accuracy (+0.42%); a layer-matched random control confirms that this effect is specific to the identified heads. A single head in the final layer contributes to the majority of the reduction in the most stereotyped classes, and class-level analysis shows that corrected predictions shift toward the correct occupation. For age, the same pipeline identifies candidate heads, but ablation produces weaker and less consistent effects, suggesting that age bias is encoded more diffusely than gender bias in this model. These results provide preliminary evidence that head-level bias localisation is feasible for discriminative vision encoders and that the degree of localisability may vary across protected attributes.\n  keywords: Bias . CLIP . Mechanistic Interpretability . Vision Transformer . Fairness", "date": "2026-03-12", "categories": []}
{"id": "2603.11936", "title": "Fair Learning for Bias Mitigation and Quality Optimization in Paper Recommendation", "abstract": "Despite frequent double-blind review, demographic biases of authors still disadvantage the underrepresented groups. We present Fair-PaperRec, a MultiLayer Perceptron (MLP)-based model that addresses demographic disparities in post-review paper acceptance decisions while maintaining high-quality requirements. Our methodology penalizes demographic disparities while preserving quality through intersectional criteria (e.g., race, country) and a customized fairness loss, in contrast to heuristic approaches. Evaluations using conference data from ACM Special Interest Group on Computer-Human Interaction (SIGCHI), Designing Interactive Systems (DIS), and Intelligent User Interfaces (IUI) indicate a 42.03% increase in underrepresented group participation and a 3.16% improvement in overall utility, indicating that diversity promotion does not compromise academic rigor and supports equity-focused peer review solutions.", "date": "2026-03-12", "categories": []}
{"id": "2301.10813", "title": "Improving Fairness with Ensemble Combination: Margin-Dependent Bounds", "abstract": "The concern about hidden discrimination in machine learning models is growing, as their widespread real-world applications increasingly impact human lives. Various techniques, including commonly used group fairness measures and several fairness-aware ensemble-based methods, have been developed to enhance fairness. However, existing fairness measures typically focus on only one aspect -- either group or individual fairness, and the compatibility difficulty among these measures indicates a possibility of remaining biases even when one of them is satisfied. Moreover, existing mechanisms to boost fairness usually present empirical results to show validity, yet few of them discuss whether fairness can be boosted with certain theoretical guarantees. To address these issues, we propose a fairness quality measure named `discriminative risk' by only perturbing protected attributes in instances, to express both individual and group fairness aspects. Furthermore, we investigate its properties and establish the first- and second-order oracle bounds and their relaxations, which show that fairness is possibly improved via ensemble combination with margin-dependent bounds. The analysis is suitable for both binary and multi-class classification. A few ensemble pruning methods are also proposed to utilise our proposed measure and obtain both accurate and fair sub-ensembles; comprehensive experiments are conducted to evaluate the effectiveness of the proposed fairness measure and pruning methods.", "date": "2023-01-25", "categories": []}
{"id": "2509.22699", "title": "Are you sure? Measuring models bias in content moderation through uncertainty", "abstract": "Automatic content moderation is crucial to ensuring safety in social media. Language Model-based classifiers are being increasingly adopted for this task, but it has been shown that they perpetuate racial and social biases. Even if several resources and benchmark corpora have been developed to challenge this issue, measuring the fairness of models in content moderation remains an open issue. In this work, we present an unsupervised approach that benchmarks models on the basis of their uncertainty in classifying messages annotated by people belonging to vulnerable groups. We use uncertainty, computed by means of the conformal prediction technique, as a proxy to analyze the bias of 11 models against women and non-white annotators and observe to what extent it diverges from metrics based on performance, such as the $F_1$ score. The results show that some pre-trained models predict with high accuracy the labels coming from minority groups, even if the confidence in their prediction is low. Therefore, by measuring the confidence of models, we are able to see which groups of annotators are better represented in pre-trained models and lead the debiasing process of these models before their effective use.", "date": "2025-09-21", "categories": []}
{"id": "2603.10332", "title": "Does Reasoning Make Search More Fair? Comparing Fairness in Reasoning and Non-Reasoning Rerankers", "abstract": "While reasoning rerankers, such as Rank1, have demonstrated strong abilities in improving ranking relevance, it is unclear how they perform on other retrieval qualities such as fairness. We conduct the first systematic comparison of fairness between reasoning and non-reasoning rerankers. Using the TREC 2022 Fair Ranking Track dataset, we evaluate six reranking models across multiple retrieval settings and demographic attributes. Our findings demonstrate reasoning neither improve nor harm fairness compared to non-reasoning approaches. Our fairness metric, Attention-Weighted Rank Fairness (AWRF) remained stable (0.33-0.35) across all models, even as relevance varies substantially (nDCG 0.247-1.000). Demographic breakdown analysis revealed fairness gaps for geographic attributes regardless of model architecture. These results indicate that future work in specializing reasoning models to be aware of fairness attributes could lead to improvements, as current implementations preserve the fairness characteristics of their input ranking.", "date": "2026-03-11", "categories": []}
{"id": "2410.02840", "title": "Overcoming Representation Bias in Fairness-Aware data Repair using Optimal Transport", "abstract": "Optimal transport (OT) has an important role in transforming data distributions in a manner which engenders fairness. Typically, the OT operators are learnt from the unfair attribute-labelled data, and then used for their repair. Two significant limitations of this approach are as follows: (i) the OT operators for underrepresented subgroups are poorly learnt (i.e. they are susceptible to representation bias); and (ii) these OT repairs cannot be effected on identically distributed but out-of-sample (i.e.\\ archival) data. In this paper, we address both of these problems by adopting a Bayesian nonparametric stopping rule for learning each attribute-labelled component of the data distribution. The induced OT-optimal quantization operators can then be used to repair the archival data. We formulate a novel definition of the fair distributional target, along with quantifiers that allow us to trade fairness against damage in the transformed data. These are used to reveal excellent performance of our representation-bias-tolerant scheme in simulated and benchmark data sets.", "date": "2024-10-03", "categories": []}
{"id": "2603.09662", "title": "No evaluation without fair representation : Impact of label and selection bias on the evaluation, performance and mitigation of classification models", "abstract": "Bias can be introduced in diverse ways in machine learning datasets, for example via selection or label bias. Although these bias types in themselves have an influence on important aspects of fair machine learning, their different impact has been understudied. In this work, we empirically analyze the effect of label bias and several subtypes of selection bias on the evaluation of classification models, on their performance, and on the effectiveness of bias mitigation methods. We also introduce a biasing and evaluation framework that allows to model fair worlds and their biased counterparts through the introduction of controlled bias in real-life datasets with low discrimination. Using our framework, we empirically analyze the impact of each bias type independently, while obtaining a more representative evaluation of models and mitigation methods than with the traditional use of a subset of biased data as test set. Our results highlight different factors that influence how impactful bias is on model performance. They also show an absence of trade-off between fairness and accuracy, and between individual and group fairness, when models are evaluated on a test set that does not exhibit unwanted bias. They furthermore indicate that the performance of bias mitigation methods is influenced by the type of bias present in the data. Our findings call for future work to develop more accurate evaluations of prediction models and fairness interventions, but also to better understand other types of bias, more complex scenarios involving the combination of different bias types, and other factors that impact the efficiency of the mitigation methods, such as dataset characteristics.", "date": "2026-03-10", "categories": []}
{"id": "2502.01406", "title": "GRADIEND: Feature Learning within Neural Networks Exemplified through Biases", "abstract": "AI systems frequently exhibit and amplify social biases, leading to harmful consequences in critical areas. This study introduces a novel encoder-decoder approach that leverages model gradients to learn a feature neuron encoding societal bias information such as gender, race, and religion. We show that our method can not only identify which weights of a model need to be changed to modify a feature, but even demonstrate that this can be used to rewrite models to debias them while maintaining other capabilities. We demonstrate the effectiveness of our approach across various model architectures and highlight its potential for broader applications.", "date": "2025-02-03", "categories": []}
{"id": "2503.15904", "title": "More Women, Same Stereotypes: Unpacking the Gender Bias Paradox in Large Language Models", "abstract": "Large Language Models (LLMs) have revolutionized natural language processing, yet concerns persist regarding their tendency to reflect or amplify social biases. This study introduces a novel evaluation framework to uncover gender biases in LLMs: using free-form storytelling to surface biases embedded within the models. A systematic analysis of ten prominent LLMs shows a consistent pattern of overrepresenting female characters across occupations, likely due to supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF). Paradoxically, despite this overrepresentation, the occupational gender distributions produced by these LLMs align more closely with human stereotypes than with real-world labor data. This highlights the challenge and importance of implementing balanced mitigation measures to promote fairness and prevent the establishment of potentially new biases. We release the prompts and LLM-generated stories at GitHub.", "date": "2025-03-20", "categories": []}
{"id": "2603.03359", "title": "ACES: Accent Subspaces for Coupling, Explanations, and Stress-Testing in Automatic Speech Recognition", "abstract": "ASR systems exhibit persistent performance disparities across accents, but whether these gaps reflect superficial biases or deep structural vulnerabilities remains unclear. We introduce ACES, a three-stage audit that extracts accent-discriminative subspaces from ASR representations, constrains adversarial attacks to them, and tests whether removing them improves fairness. On Wav2Vec2-base with seven accents, imperceptible perturbations (~60 dB SNR) along the accent subspace amplify the WER disparity gap by nearly 50% (21.3->31.8 pp), exceeding random-subspace controls; a permuted-label test confirms specificity to genuine accent structure. Partially removing the subspace worsens both WER and disparity, revealing that accent-discriminative and recognition-critical features are deeply entangled. ACES thus positions accent subspaces as powerful fairness-auditing tools, not simple erasure levers.", "date": "2026-02-28", "categories": []}
{"id": "2603.06733", "title": "Calibrated Credit Intelligence: Shift-Robust and Fair Risk Scoring with Bayesian Uncertainty and Gradient Boosting", "abstract": "Credit risk scoring must support high-stakes lending decisions where data distributions change over time, probability estimates must be reliable, and group-level fairness is required. While modern machine learning models improve default prediction accuracy, they often produce poorly calibrated scores under distribution shift and may create unfair outcomes when trained without explicit constraints. This paper proposes Calibrated Credit Intelligence (CCI), a deployment-oriented framework that combines (i) a Bayesian neural risk scorer to capture epistemic uncertainty and reduce overconfident errors, (ii) a fairnessconstrained gradient boosting model to control group disparities while preserving strong tabular performance, and (iii) a shiftaware fusion strategy followed by post-hoc probability calibration to stabilize decision thresholds in later time periods. We evaluate CCI on the Home Credit Credit Risk Model Stability benchmark using a time-consistent split to reflect real-world drift. Compared with strong baselines (LightGBM, XGBoost, CatBoost, TabNet, and a standalone Bayesian neural model), CCI achieves the best overall trade-off between discrimination, calibration, stability, and fairness. In particular, CCI reaches an AUC-ROC of 0.912 and an AUC-PR of 0.438, improves operational performance with Recall@1%FPR = 0.509, and reduces calibration error (Brier score 0.087, ECE 0.015). Under temporal shift, CCI shows a smaller AUC-PR drop from early to late periods (0.017), and it lowers group disparities (demographic parity gap 0.046, equal opportunity gap 0.037) compared to unconstrained boosting. These results indicate that CCI produces risk scores that are accurate, reliable, and more equitable under realistic deployment conditions.", "date": "2026-03-06", "categories": []}
{"id": "2603.06901", "title": "Fairness May Backfire: When Leveling-Down Occurs in Fair Machine Learning", "abstract": "As machine learning (ML) systems increasingly shape access to credit, jobs, and other opportunities, the fairness of algorithmic decisions has become a central concern. Yet it remains unclear when enforcing fairness constraints in these systems genuinely improves outcomes for affected groups or instead leads to \"leveling down,\" making one or both groups worse off. We address this question in a unified, population-level (Bayes) framework for binary classification under prevalent group fairness notions. Our Bayes approach is distribution-free and algorithm-agnostic, isolating the intrinsic effect of fairness requirements from finite-sample noise and from training and intervention specifics. We analyze two deployment regimes for ML classifiers under common legal and governance constraints: attribute-aware decision-making (sensitive attributes available at decision time) and attribute-blind decision-making (sensitive attributes excluded from prediction). We show that, in the attribute-aware regime, fair ML necessarily (weakly) improves outcomes for the disadvantaged group and (weakly) worsens outcomes for the advantaged group. In contrast, in the attribute-blind regime, the impact of fairness is distribution-dependent: fairness can benefit or harm either group and may shift both groups' outcomes in the same direction, leading to either leveling up or leveling down. We characterize the conditions under which these patterns arise and highlight the role of \"masked\" candidates in driving them. Overall, our results provide structural guidance on when pursuing algorithmic fairness is likely to improve group outcomes and when it risks systemic leveling down, informing fair ML design and deployment choices.", "date": "2026-03-06", "categories": []}
{"id": "2603.07368", "title": "Position: LLMs Must Use Functor-Based and RAG-Driven Bias Mitigation for Fairness", "abstract": "Biases in large language models (LLMs) often manifest as systematic distortions in associations between demographic attributes and professional or social roles, reinforcing harmful stereotypes across gender, ethnicity, and geography. This position paper advocates for addressing demographic and gender biases in LLMs through a dual-pronged methodology, integrating category-theoretic transformations and retrieval-augmented generation (RAG). Category theory provides a rigorous, structure-preserving mathematical framework that maps biased semantic domains to unbiased canonical forms via functors, ensuring bias elimination while preserving semantic integrity. Complementing this, RAG dynamically injects diverse, up-to-date external knowledge during inference, directly countering ingrained biases within model parameters. By combining structural debiasing through functor-based mappings and contextual grounding via RAG, we outline a comprehensive framework capable of delivering equitable and fair model outputs. Our synthesis of the current literature validates the efficacy of each approach individually, while addressing potential critiques demonstrates the robustness of this integrated strategy. Ensuring fairness in LLMs, therefore, demands both the mathematical rigor of category-theoretic transformations and the adaptability of retrieval augmentation.", "date": "2026-03-07", "categories": []}
{"id": "2603.08267", "title": "Towards a more efficient bias detection in financial language models", "abstract": "Bias in financial language models constitutes a major obstacle to their adoption in real-world applications. Detecting such bias is challenging, as it requires identifying inputs whose predictions change when varying properties unrelated to the decision, such as demographic attributes. Existing approaches typically rely on exhaustive mutation and pairwise prediction analysis over large corpora, which is effective but computationally expensive-particularly for large language models and can become impractical in continuous retraining and releasing processes. Aiming at reducing this cost, we conduct a large-scale study of bias in five financial language models, examining similarities in their bias tendencies across protected attributes and exploring cross-model-guided bias detection to identify bias-revealing inputs earlier. Our study uses approximately 17k real financial news sentences, mutated to construct over 125k original-mutant pairs. Results show that all models exhibit bias under both atomic (0.58\\%-6.05\\%) and intersectional (0.75\\%-5.97\\%) settings. Moreover, we observe consistent patterns in bias-revealing inputs across models, enabling substantial reuse and cost reduction in bias detection. For example, up to 73\\% of FinMA's biased behaviours can be uncovered using only 20\\% of the input pairs when guided by properties derived from DistilRoBERTa outputs.", "date": "2026-03-09", "categories": []}
{"id": "2408.01285", "title": "Do Prevalent Bias Metrics Capture Allocational Harms from LLMs?", "abstract": "Allocational harms occur when resources or opportunities are unfairly withheld from specific groups. Many proposed bias measures ignore the discrepancy between predictions, which are what the proposed methods consider, and decisions that are made as a result of those predictions. Our work examines the reliability of current bias metrics in assessing allocational harms arising from predictions of large language models (LLMs). We evaluate their predictive validity and utility for model selection across ten LLMs and two allocation tasks. Our results reveal that commonly-used bias metrics based on average performance gap and distribution distance fail to reliably capture group disparities in allocation outcomes. Our work highlights the need to account for how model predictions are used in decisions, in particular in contexts where they are influenced by how limited resources are allocated.", "date": "2024-08-02", "categories": []}
{"id": "2412.01711", "title": "Towards Resource Efficient and Interpretable Bias Mitigation in Large Language Models", "abstract": "Although large language models (LLMs) have demonstrated their effectiveness in a wide range of applications, they have also been observed to perpetuate unwanted biases present in the training data, potentially leading to harm for marginalized communities. In this paper, we mitigate bias by leveraging small biased and anti-biased expert models to obtain a debiasing signal that is added to the LLM output at decoding-time. This approach combines computational efficiency - fine-tuning a small model versus re-training a large model and interpretability - one can examine the probability shift from debiasing. The framework can also be tailored to specific contexts by switching the choice of the fine-tuning dataset. Experiments on mitigating gender, race, and religion biases on different architectures show a reduction in bias on several local and global bias metrics while preserving language model performance.", "date": "2024-12-02", "categories": []}
{"id": "2411.09847", "title": "Towards a Fairer Non-negative Matrix Factorization", "abstract": "There has been a recent critical need to study fairness and bias in machine learning (ML) algorithms. Since there is clearly no one-size-fits-all solution to fairness, ML methods should be developed alongside bias mitigation strategies that are practical and approachable to the practitioner. Motivated by recent work on ``fair\" PCA, here we consider the more challenging method of non-negative matrix factorization (NMF) as both a showcasing example and a method that is important in its own right for both topic modeling tasks and feature extraction for other ML tasks. We demonstrate that a modification of the objective function, by using a min-max formulation, may \\textit{sometimes} be able to offer an improvement in fairness for groups in the population. We derive two methods for the objective minimization, a multiplicative update rule as well as an alternating minimization scheme, and discuss implementation practicalities. We include a suite of synthetic and real experiments that show how the method may improve fairness while also highlighting the important fact that this may sometime increase error for some individuals and fairness is not a rigid definition and method choice should strongly depend on the application at hand.", "date": "2024-11-14", "categories": []}
{"id": "2603.05157", "title": "The Impact of Preprocessing Methods on Racial Encoding and Model Robustness in CXR Diagnosis", "abstract": "Deep learning models can identify racial identity with high accuracy from chest X-ray (CXR) recordings. Thus, there is widespread concern about the potential for racial shortcut learning, where a model inadvertently learns to systematically bias its diagnostic predictions as a function of racial identity. Such racial biases threaten healthcare equity and model reliability, as models may systematically misdiagnose certain demographic groups. Since racial shortcuts are diffuse - non-localized and distributed throughout the whole CXR recording - image preprocessing methods may influence racial shortcut learning, yet the potential of such methods for reducing biases remains underexplored. Here, we investigate the effects of image preprocessing methods including lung masking, lung cropping, and Contrast Limited Adaptive Histogram Equalization (CLAHE). These approaches aim to suppress spurious cues encoding racial information while preserving diagnostic accuracy. Our experiments reveal that simple bounding box-based lung cropping can be an effective strategy for reducing racial shortcut learning while maintaining diagnostic model performance, bypassing frequently postulated fairness-accuracy trade-offs.", "date": "2026-03-05", "categories": []}
{"id": "2603.05226", "title": "Learning Optimal Individualized Decision Rules with Conditional Demographic Parity", "abstract": "Individualized decision rules (IDRs) have become increasingly prevalent in societal applications such as personalized marketing, healthcare, and public policy design. However, a critical ethical concern arises from the potential discriminatory effects of IDRs trained on biased data. These algorithms may disproportionately harm individuals from minority subgroups defined by sensitive attributes like gender, race, or language. To address this issue, we propose a novel framework that incorporates demographic parity (DP) and conditional demographic parity (CDP) constraints into the estimation of optimal IDRs. We show that the theoretically optimal IDRs under DP and CDP constraints can be obtained by applying perturbations to the unconstrained optimal IDRs, enabling a computationally efficient solution. Theoretically, we derive convergence rates for both policy value and the fairness constraint term. The effectiveness of our methods is illustrated through comprehensive simulation studies and an empirical application to the Oregon Health Insurance Experiment.", "date": "2026-03-05", "categories": []}
{"id": "2603.05327", "title": "FairFinGAN: Fairness-aware Synthetic Financial Data Generation", "abstract": "Financial datasets often suffer from bias that can lead to unfair decision-making in automated systems. In this work, we propose FairFinGAN, a WGAN-based framework designed to generate synthetic financial data while mitigating bias with respect to the protected attribute. Our approach incorporates fairness constraints directly into the training process through a classifier, ensuring that the synthetic data is both fair and preserves utility for downstream predictive tasks. We evaluate our proposed model on five real-world financial datasets and compare it with existing GAN-based data generation methods. Experimental results show that our approach achieves superior fairness metrics without significant loss in data utility, demonstrating its potential as a tool for bias-aware data generation in financial applications.", "date": "2026-03-05", "categories": []}
{"id": "2603.03341", "title": "Ethical and Explainable AI in Reusable MLOps Pipelines", "abstract": "This paper introduces a unified machine learning operations (MLOps) framework that brings ethical artificial intelligence principles into practical use by enforcing fairness, explainability, and governance throughout the machine learning lifecycle. The proposed method reduces bias by lowering the demographic parity difference (DPD) from 0.31 to 0.04 without model retuning, and cross-dataset validation achieves an area under the curve (AUC) of 0.89 on the Statlog Heart dataset.\n  The framework maintains fairness metrics within operational limits across all deployments. Model deployment is blocked if the DPD exceeds 0.05 or if equalized odds (EO) exceeds 0.05 on the validation set. After deployment, retraining is automatically triggered if the 30-day Kolmogorov-Smirnov drift statistic exceeds 0.20. In production, the system consistently achieved DPD <= 0.05 and EO <= 0.03, while the KS statistic remained <= 0.20.\n  Decision-curve analysis indicates a positive net benefit in the 10 to 20 percent operating range, showing that the mitigated model preserves predictive utility while satisfying fairness constraints. These results demonstrate that automated fairness gates and explainability artefacts can be successfully deployed in production without disrupting operational flow, providing organizations with a practical and credible approach to implementing ethical, transparent, and trustworthy AI across diverse datasets and operational settings.", "date": "2026-02-15", "categories": []}
{"id": "2603.03820", "title": "Fairness Begins with State: Purifying Latent Preferences for Hierarchical Reinforcement Learning in Interactive Recommendation", "abstract": "Interactive recommender systems (IRS) are increasingly optimized with Reinforcement Learning (RL) to capture the sequential nature of user-system dynamics. However, existing fairness-aware methods often suffer from a fundamental oversight: they assume the observed user state is a faithful representation of true preferences. In reality, implicit feedback is contaminated by popularity-driven noise and exposure bias, creating a distorted state that misleads the RL agent. We argue that the persistent conflict between accuracy and fairness is not merely a reward-shaping issue, but a state estimation failure. In this work, we propose \\textbf{DSRM-HRL}, a framework that reformulates fairness-aware recommendation as a latent state purification problem followed by decoupled hierarchical decision-making. We introduce a Denoising State Representation Module (DSRM) based on diffusion models to recover the low-entropy latent preference manifold from high-entropy, noisy interaction histories. Built upon this purified state, a Hierarchical Reinforcement Learning (HRL) agent is employed to decouple conflicting objectives: a high-level policy regulates long-term fairness trajectories, while a low-level policy optimizes short-term engagement under these dynamic constraints. Extensive experiments on high-fidelity simulators (KuaiRec, KuaiRand) demonstrate that DSRM-HRL effectively breaks the \"rich-get-richer\" feedback loop, achieving a superior Pareto frontier between recommendation utility and exposure equity.", "date": "2026-03-04", "categories": []}
{"id": "2603.03867", "title": "k-hop Fairness: Addressing Disparities in Graph Link Prediction Beyond First-Order Neighborhoods", "abstract": "Link prediction (LP) plays a central role in graph-based applications, particularly in social recommendation. However, real-world graphs often reflect structural biases, most notably homophily, the tendency of nodes with similar attributes to connect. While this property can improve predictive performance, it also risks reinforcing existing social disparities. In response, fairness-aware LP methods have emerged, often seeking to mitigate these effects by promoting inter-group connections, that is, links between nodes with differing sensitive attributes (e.g., gender), following the principle of dyadic fairness. However, dyadic fairness overlooks potential disparities within the sensitive groups themselves. To overcome this issue, we propose $k$-hop fairness, a structural notion of fairness for LP, that assesses disparities conditioned on the distance between nodes in the graph. We formalize this notion through predictive fairness and structural bias metrics, and propose pre- and post-processing mitigation strategies. Experiments across standard LP benchmarks reveal: (1) a strong tendency of models to reproduce structural biases at different $k$-hops; (2) interdependence between structural biases at different hops when rewiring graphs; and (3) that our post-processing method achieves favorable $k$-hop performance-fairness trade-offs compared to existing fair LP baselines.", "date": "2026-03-04", "categories": []}
{"id": "2303.15585", "title": "(Un)fair devices: Moving beyond AI accuracy in personal sensing", "abstract": "Personal devices are omnipresent in our lives, seamlessly monitoring our activities, from smart rings tracking sleep patterns to smartwatches keeping an eye on missed heartbeats. The rich data streams from such devices fuel advanced Artificial Intelligence (AI) applications. Instead of solely relying on direct sensor measurements, these applications are increasingly leveraging Machine Learning (ML) model estimates to derive insights. But are these estimates biased or not? This literature review delivers compelling evidence about the impact of hidden biases that creep into ML models deployed on personal devices. We discuss critical bias issues drawn from prior work such as racial bias in pulse oximeters, weight bias in optical heart rate sensors, and sex bias in audio-based diagnostics. In response to these challenges, we advocate for a shift from prioritizing performance-oriented evaluations of personal devices to adopting assessments grounded in a human-centered approach. To facilitate this transition, we provide guidelines for the design, development, evaluation, and use of unbiased AI in personal devices, recognizing their potential impact on improving our health, lifestyle, and productivity -- more than any other technology.", "date": "2023-03-27", "categories": []}
{"id": "2603.02937", "title": "Bias and Fairness in Self-Supervised Acoustic Representations for Cognitive Impairment Detection", "abstract": "Speech-based detection of cognitive impairment (CI) offers a promising non-invasive approach for early diagnosis, yet performance disparities across demographic and clinical subgroups remain underexplored, raising concerns around fairness and generalizability. This study presents a systematic bias analysis of acoustic-based CI and depression classification using the DementiaBank Pitt Corpus. We compare traditional acoustic features (MFCCs, eGeMAPS) with contextualized speech embeddings from Wav2Vec 2.0 (W2V2), and evaluate classification performance across gender, age, and depression-status subgroups. For CI detection, higher-layer W2V2 embeddings outperform baseline features (UAR up to 80.6\\%), but exhibit performance disparities; specifically, females and younger participants demonstrate lower discriminative power (\\(AUC\\): 0.769 and 0.746, respectively) and substantial specificity disparities (\\(\\Delta_{spec}\\) up to 18\\% and 15\\%, respectively), leading to a higher risk of misclassifications than their counterparts. These disparities reflect representational biases, defined as systematic differences in model performance across demographic or clinical subgroups. Depression detection within CI subjects yields lower overall performance, with mild improvements from low and mid-level W2V2 layers. Cross-task generalization between CI and depression classification is limited, indicating that each task depends on distinct representations. These findings emphasize the need for fairness-aware model evaluation and subgroup-specific analysis in clinical speech applications, particularly in light of demographic and clinical heterogeneity in real-world applications.", "date": "2026-03-03", "categories": []}
{"id": "2412.20377", "title": "On Demographic Group Fairness Guarantees in Deep Learning", "abstract": "We present a theoretical framework analyzing the relationship between data distributions and fairness guarantees in equitable deep learning. We establish novel bounds that account for distribution heterogeneity across demographic groups, deriving fairness error and convergence rate bounds that characterize how distributional differences affect the fairness-accuracy trade-off. Extensive experiments across diverse modalities, including FairVision, CheXpert, HAM10000, FairFace, ACS Income, and CivilComments-WILDS, validate our theoretical findings, demonstrating that feature distribution differences across demographic groups significantly impact model fairness, with disparities particularly pronounced in racial categories. Motivated by these insights, we propose Fairness-Aware Regularization (FAR), a practical training objective that minimizes inter-group discrepancies in feature centroids and covariances. FAR consistently improves overall AUC, ES-AUC, and subgroup performance across all datasets. Our work advances the theoretical understanding of fairness in AI systems and provides a foundation for developing more equitable algorithms. The code for analysis is publicly available at https://github.com/Harvard-AI-and-Robotics-Lab/FairnessGuarantee.", "date": "2024-12-29", "categories": []}
{"id": "2503.11120", "title": "A Multi-Objective Evaluation Framework for Analyzing Utility-Fairness Trade-Offs in Machine Learning Systems", "abstract": "The evaluation of fairness models in Machine Learning involves complex challenges, such as defining appropriate metrics, balancing trade-offs between utility and fairness, and there are still gaps in this stage. This work presents a novel multi-objective evaluation framework that enables the analysis of utility-fairness trade-offs in Machine Learning systems. The framework was developed using criteria from Multi-Objective Optimization that collect comprehensive information regarding this complex evaluation task. The assessment of multiple Machine Learning systems is summarized, both quantitatively and qualitatively, in a straightforward manner through a radar chart and a measurement table encompassing various aspects such as convergence, system capacity, and diversity. The framework's compact representation of performance facilitates the comparative analysis of different Machine Learning strategies for decision-makers, in real-world applications, with single or multiple fairness requirements. In particular, this study focuses on the medical imaging domain, where fairness considerations are crucial due to the potential impact of biased diagnostic systems on patient outcomes. The proposed framework enables a systematic evaluation of multiple fairness constraints helping to identify and mitigate disparities among demographic groups while maintaining diagnostic performance. The framework is model-agnostic and flexible to be adapted to any kind of Machine Learning systems, that is, black- or white-box, any kind and quantity of evaluation metrics, including multidimensional fairness criteria. The functionality and effectiveness of the proposed framework is shown with different simulations, and an empirical study conducted on three real-world medical imaging datasets with various Machine Learning systems. Our evaluation framework is publicly available at https://pypi.org/project/fairical.", "date": "2025-03-14", "categories": []}
{"id": "2506.13104", "title": "Equitable Electronic Health Record Prediction with FAME: Fairness-Aware Multimodal Embedding", "abstract": "Electronic Health Record (EHR) data encompass diverse modalities -- text, images, and medical codes -- that are vital for clinical decision-making. To process these complex data, multimodal AI (MAI) has emerged as a powerful approach for fusing such information. However, most existing MAI models optimize for better prediction performance, potentially reinforcing biases across patient subgroups. Although bias-reduction techniques for multimodal models have been proposed, the individual strengths of each modality and their interplay in both reducing bias and optimizing performance remain underexplored. In this work, we introduce FAME (Fairness-Aware Multimodal Embeddings), a framework that explicitly weights each modality according to its fairness contribution. FAME optimizes both performance and fairness by incorporating a combined loss function. We leverage the Error Distribution Disparity Index (EDDI) to measure fairness across subgroups and propose a sign-agnostic aggregation method to balance fairness across subgroups, ensuring equitable model outcomes. We evaluate FAME with BEHRT and BioClinicalBERT, combining structured and unstructured EHR data, and demonstrate its effectiveness in terms of performance and fairness compared with other baselines across multiple EHR prediction tasks.", "date": "2025-06-16", "categories": []}
{"id": "2511.00177", "title": "Can SAEs reveal and mitigate racial biases of LLMs in healthcare?", "abstract": "LLMs are increasingly being used in healthcare. This promises to free physicians from drudgery, enabling better care to be delivered at scale. But the use of LLMs in this space also brings risks; for example, such models may worsen existing biases. How can we spot when LLMs are (spuriously) relying on patient race to inform predictions? In this work we assess the degree to which Sparse Autoencoders (SAEs) can reveal (and control) associations the model has made between race and stigmatizing concepts. We first identify SAE latents in Gemma-2 models which appear to correlate with Black individuals. We find that this latent activates on reasonable input sequences (e.g., \"African American\") but also problematic words like \"incarceration\". We then show that we can use this latent to steer models to generate outputs about Black patients, and further that this can induce problematic associations in model outputs as a result. For example, activating the Black latent increases the risk assigned to the probability that a patient will become \"belligerent\". We evaluate the degree to which such steering via latents might be useful for mitigating bias. We find that this offers improvements in simple settings, but is less successful for more realistic and complex clinical tasks. Overall, our results suggest that: SAEs may offer a useful tool in clinical applications of LLMs to identify problematic reliance on demographics but mitigating bias via SAE steering appears to be of marginal utility for realistic tasks.", "date": "2025-10-31", "categories": []}
{"id": "2602.22438", "title": "From Bias to Balance: Fairness-Aware Paper Recommendation for Equitable Peer Review", "abstract": "Despite frequent double-blind review, systemic biases related to author demographics still disadvantage underrepresented groups. We start from a simple hypothesis: if a post-review recommender is trained with an explicit fairness regularizer, it should increase inclusion without degrading quality. To test this, we introduce Fair-PaperRec, a Multi-Layer Perceptron (MLP) with a differentiable fairness loss over intersectional attributes (e.g., race, country) that re-ranks papers after double-blind review. We first probe the hypothesis on synthetic datasets spanning high, moderate, and near-fair biases. Across multiple randomized runs, these controlled studies map where increasing the fairness weight strengthens macro/micro diversity while keeping utility approximately stable, demonstrating robustness and adaptability under varying disparity levels. We then carry the hypothesis into the original setting, conference data from ACM Special Interest Group on Computer-Human Interaction (SIGCHI), Designing Interactive Systems (DIS), and Intelligent User Interfaces (IUI). In this real-world scenario, an appropriately tuned configuration of Fair-PaperRec achieves up to a 42.03% increase in underrepresented-group participation with at most a 3.16% change in overall utility relative to the historical selection. Taken together, the synthetic-to-original progression shows that fairness regularization can act as both an equity mechanism and a mild quality regularizer, especially in highly biased regimes. By first analyzing the behavior of the fairness parameters under controlled conditions and then validating them on real submissions, Fair-PaperRec offers a practical, equity-focused framework for post-review paper selection that preserves, and in some settings can even enhance, measured scholarly quality.", "date": "2026-02-25", "categories": []}
{"id": "2603.00590", "title": "Fair in Mind, Fair in Action? A Synchronous Benchmark for Understanding and Generation in UMLLMs", "abstract": "As artificial intelligence (AI) is increasingly deployed across domains, ensuring fairness has become a core challenge. However, the field faces a \"Tower of Babel'' dilemma: fairness metrics abound, yet their underlying philosophical assumptions often conflict, hindering unified paradigms-particularly in unified Multimodal Large Language Models (UMLLMs), where biases propagate systemically across tasks. To address this, we introduce the IRIS Benchmark, to our knowledge the first benchmark designed to synchronously evaluate the fairness of both understanding and generation tasks in UMLLMs. Enabled by our demographic classifier, ARES, and four supporting large-scale datasets, the benchmark is designed to normalize and aggregate arbitrary metrics into a high-dimensional \"fairness space'', integrating 60 granular metrics across three dimensions-Ideal Fairness, Real-world Fidelity, and Bias Inertia & Steerability (IRIS). Through this benchmark, our evaluation of leading UMLLMs uncovers systemic phenomena such as the \"generation gap'', individual inconsistencies like \"personality splits'', and the \"counter-stereotype reward'', while offering diagnostics to guide the optimization of their fairness capabilities. With its novel and extensible framework, the IRIS benchmark is capable of integrating evolving fairness metrics, ultimately helping to resolve the \"Tower of Babel'' impasse. Project Page: https://iris-benchmark-web.vercel.app/", "date": "2026-02-28", "categories": []}
{"id": "2505.19441", "title": "Fairness-in-the-Workflow: How Machine Learning Practitioners at Big Tech Companies Approach Fairness in Recommender Systems", "abstract": "Recommender systems (RS), which are widely deployed across high-stakes domains, are susceptible to biases that can cause large-scale societal impacts. Researchers have proposed methods to measure and mitigate such biases - but translating academic theory into practice is inherently challenging. Through a semi-structured interview study (N=11), we map the RS practitioner workflow within large technology companies, focusing on how technical teams consider fairness internally and in collaboration with legal, data, and fairness teams. We identify key challenges to incorporating fairness into existing RS workflows: defining fairness in RS contexts, balancing multi-stakeholder interests, and navigating dynamic environments. We also identify key organization-wide challenges: making time for fairness work and facilitating cross-team communication. Finally, we offer actionable recommendations for the RS community, including practitioners and HCI researchers.", "date": "2025-05-26", "categories": []}
{"id": "2602.23611", "title": "Fairness under Graph Uncertainty: Achieving Interventional Fairness with Partially Known Causal Graphs over Clusters of Variables", "abstract": "Algorithmic decisions about individuals require predictions that are not only accurate but also fair with respect to sensitive attributes such as gender and race. Causal notions of fairness align with legal requirements, yet many methods assume access to detailed knowledge of the underlying causal graph, which is a demanding assumption in practice. We propose a learning framework that achieves interventional fairness by leveraging a causal graph over \\textit{clusters of variables}, which is substantially easier to estimate than a variable-level graph. With possible \\textit{adjustment cluster sets} identified from such a cluster causal graph, our framework trains a prediction model by reducing the worst-case discrepancy between interventional distributions across these sets. To this end, we develop a computationally efficient barycenter kernel maximum mean discrepancy (MMD) that scales favorably with the number of sensitive attribute values. Extensive experiments show that our framework strikes a better balance between fairness and accuracy than existing approaches, highlighting its effectiveness under limited causal graph knowledge.", "date": "2026-02-27", "categories": []}
{"id": "2602.24014", "title": "Interpretable Debiasing of Vision-Language Models for Social Fairness", "abstract": "The rapid advancement of Vision-Language models (VLMs) has raised growing concerns that their black-box reasoning processes could lead to unintended forms of social bias. Current debiasing approaches focus on mitigating surface-level bias signals through post-hoc learning or test-time algorithms, while leaving the internal dynamics of the model largely unexplored. In this work, we introduce an interpretable, model-agnostic bias mitigation framework, DeBiasLens, that localizes social attribute neurons in VLMs through sparse autoencoders (SAEs) applied to multimodal encoders. Building upon the disentanglement ability of SAEs, we train them on facial image or caption datasets without corresponding social attribute labels to uncover neurons highly responsive to specific demographics, including those that are underrepresented. By selectively deactivating the social neurons most strongly tied to bias for each group, we effectively mitigate socially biased behaviors of VLMs without degrading their semantic knowledge. Our research lays the groundwork for future auditing tools, prioritizing social fairness in emerging real-world AI systems.", "date": "2026-02-27", "categories": []}
{"id": "2404.01877", "title": "Procedural Fairness in Machine Learning", "abstract": "Fairness in machine learning (ML) has garnered significant attention. However, current research has mainly concentrated on the distributive fairness of ML models, with limited focus on another dimension of fairness, i.e., procedural fairness. In this paper, we first define the procedural fairness of ML models by drawing from the established understanding of procedural fairness in philosophy and psychology fields, and then give formal definitions of individual and group procedural fairness. Based on the proposed definition, we further propose a novel metric to evaluate the group procedural fairness of ML models, called $GPF_{FAE}$, which utilizes a widely used explainable artificial intelligence technique, namely feature attribution explanation (FAE), to capture the decision process of ML models. We validate the effectiveness of $GPF_{FAE}$ on a synthetic dataset and eight real-world datasets. Our experimental studies have revealed the relationship between procedural and distributive fairness of ML models. After validating the proposed metric for assessing the procedural fairness of ML models, we then propose a method for identifying the features that lead to the procedural unfairness of the model and propose two methods to improve procedural fairness based on the identified unfair features. Our experimental results demonstrate that we can accurately identify the features that lead to procedural unfairness in the ML model, and both of our proposed methods can significantly improve procedural fairness while also improving distributive fairness, with a slight sacrifice on the model performance.", "date": "2024-04-02", "categories": []}
{"id": "2506.16782", "title": "What Is the Point of Equality in Machine Learning Fairness? Beyond Equality of Opportunity", "abstract": "Fairness in machine learning (ML) has become a rapidly growing area of research. But why, in the first place, is unfairness in ML wrong? And why should we care about improving fairness? Most fair-ML research implicitly appeals to distributive equality: the idea that desirable benefits and goods, such as opportunities (e.g., Barocas et al., 2023), should be equally distributed across society. Unfair ML models, then, are seen as wrong because they unequally distribute such benefits. This paper argues that this exclusive focus on distributive equality offers an incomplete and potentially misleading ethical foundation. Grounding ML fairness in egalitarianism--the view that equality is a fundamental moral and social ideal--requires challenging structural inequality: systematic, institutional, and durable arrangements that privilege some groups while disadvantaging others. Structural inequality manifests through ML systems in two primary forms: allocative harms (e.g., economic loss) and representational harms (e.g., stereotypes, erasure). While distributive equality helps address allocative harms, it fails to explain why representational harms are wrong--why it is wrong for ML systems to reinforce social hierarchies that stratify people into superior and inferior groups--and why ML systems should aim to foster a society where people relate as equals (i.e., relational equality). To address these limitations, the paper proposes a multifaceted egalitarian framework for ML fairness that integrates both distributive and relational equality. Drawing on critical social and political philosophy, this framework offers a more comprehensive ethical foundation for tackling the full spectrum of harms perpetuated by ML systems. The paper also outlines practical pathways for implementing the framework across the entire ML pipeline.", "date": "2025-06-20", "categories": []}
{"id": "2509.17956", "title": "\"I think this is fair\": Uncovering the Complexities of Stakeholder Decision-Making in AI Fairness Assessment", "abstract": "Assessing fairness in artificial intelligence (AI) typically involves AI experts who select protected features, fairness metrics, and set fairness thresholds to assess outcome fairness. However, little is known about how stakeholders, particularly those affected by AI outcomes but lacking AI expertise, assess fairness. To address this gap, we conducted a qualitative study with 26 stakeholders without AI expertise, representing potential decision subjects in a credit rating scenario, to examine how they assess fairness when placed in the role of deciding on features with priority, metrics, and thresholds. We reveal that stakeholders' fairness decisions are more complex than typical AI expert practices: they considered features far beyond legally protected features, tailored metrics for specific contexts, set diverse yet stricter fairness thresholds, and even preferred designing customized fairness. Our results extend the understanding of how stakeholders can meaningfully contribute to AI fairness governance and mitigation, underscoring the importance of incorporating stakeholders' nuanced fairness judgments.", "date": "2025-09-22", "categories": []}
{"id": "2602.22560", "title": "Operationalizing Fairness: Post-Hoc Threshold Optimization Under Hard Resource Limits", "abstract": "The deployment of machine learning in high-stakes domains requires a balance between predictive safety and algorithmic fairness. However, existing fairness interventions often as- sume unconstrained resources and employ group-specific decision thresholds that violate anti- discrimination regulations. We introduce a post-hoc, model-agnostic threshold optimization framework that jointly balances safety, efficiency, and equity under strict and hard capacity constraints. To ensure legal compliance, the framework enforces a single, global decision thresh- old. We formulated a parameterized ethical loss function coupled with a bounded decision rule that mathematically prevents intervention volumes from exceeding the available resources. An- alytically, we prove the key properties of the deployed threshold, including local monotonicity with respect to ethical weighting and the formal identification of critical capacity regimes. We conducted extensive experimental evaluations on diverse high-stakes datasets. The principal re- sults demonstrate that capacity constraints dominate ethical priorities; the strict resource limit determines the final deployed threshold in over 80% of the tested configurations. Furthermore, under a restrictive 25% capacity limit, the proposed framework successfully maintains high risk identification (recall ranging from 0.409 to 0.702), whereas standard unconstrained fairness heuristics collapse to a near-zero utility. We conclude that theoretical fairness objectives must be explicitly subordinated to operational capacity limits to remain in deployment. By decou- pling predictive scoring from policy evaluation and strictly bounding intervention rates, this framework provides a practical and legally compliant mechanism for stakeholders to navigate unavoidable ethical trade-offs in resource-constrained environments.", "date": "2026-02-26", "categories": []}
{"id": "2602.21509", "title": "Fair Model-based Clustering", "abstract": "The goal of fair clustering is to find clusters such that the proportion of sensitive attributes (e.g., gender, race, etc.) in each cluster is similar to that of the entire dataset. Various fair clustering algorithms have been proposed that modify standard K-means clustering to satisfy a given fairness constraint. A critical limitation of several existing fair clustering algorithms is that the number of parameters to be learned is proportional to the sample size because the cluster assignment of each datum should be optimized simultaneously with the cluster center, and thus scaling up the algorithms is difficult. In this paper, we propose a new fair clustering algorithm based on a finite mixture model, called Fair Model-based Clustering (FMC). A main advantage of FMC is that the number of learnable parameters is independent of the sample size and thus can be scaled up easily. In particular, mini-batch learning is possible to obtain clusters that are approximately fair. Moreover, FMC can be applied to non-metric data (e.g., categorical data) as long as the likelihood is well-defined. Theoretical and empirical justifications for the superiority of the proposed algorithm are provided.", "date": "2026-02-25", "categories": []}
{"id": "2602.00044", "title": "When LLMs Imagine People: A Human-Centered Persona Brainstorm Audit for Bias and Fairness in Creative Applications", "abstract": "Large Language Models (LLMs) used in creative workflows can reinforce stereotypes and perpetuate inequities, making fairness auditing essential. Existing methods rely on constrained tasks and fixed benchmarks, leaving open-ended creative outputs unexamined. We introduce the Persona Brainstorm Audit (PBA), a scalable and easy to extend auditing method for bias detection across multiple intersecting identity and social roles in open-ended persona generation. PBA quantifies bias using degree-of-freedom-aware normalized Cram\\'er's V, producing interpretable severity labels that enable fair comparison across models and dimensions. Applying PBA to 12 LLMs (120,000 personas, 16 bias dimensions), we find that bias evolves nonlinearly across model generations: larger and newer models are not consistently fairer, and biases that initially decrease can resurface in later releases. Intersectional analysis reveals disparities hidden by single-axis metrics, where dimensions appearing fair individually can exhibit high bias in combination. Robustness analyses show PBA remains stable under varying sample sizes, role-playing prompts, and debiasing prompts, establishing its reliability for fairness auditing in LLMs.", "date": "2026-01-19", "categories": []}
{"id": "2602.20624", "title": "Physics-based phenomenological characterization of cross-modal bias in multimodal models", "abstract": "The term 'algorithmic fairness' is used to evaluate whether AI models operate fairly in both comparative (where fairness is understood as formal equality, such as \"treat like cases as like\") and non-comparative (where unfairness arises from the model's inaccuracy, arbitrariness, or inscrutability) contexts. Recent advances in multimodal large language models (MLLMs) are breaking new ground in multimodal understanding, reasoning, and generation; however, we argue that inconspicuous distortions arising from complex multimodal interaction dynamics can lead to systematic bias. The purpose of this position paper is twofold: first, it is intended to acquaint AI researchers with phenomenological explainable approaches that rely on the physical entities that the machine experiences during training/inference, as opposed to the traditional cognitivist symbolic account or metaphysical approaches; second, it is to state that this phenomenological doctrine will be practically useful for tackling algorithmic fairness issues in MLLMs. We develop a surrogate physics-based model that describes transformer dynamics (i.e., semantic network structure and self-/cross-attention) to analyze the dynamics of cross-modal bias in MLLM, which are not fully captured by conventional embedding- or representation-level analyses. We support this position through multi-input diagnostic experiments: 1) perturbation-based analyses of emotion classification using Qwen2.5-Omni and Gemma 3n, and 2) dynamical analysis of Lorenz chaotic time-series prediction through the physical surrogate. Across two architecturally distinct MLLMs, we show that multimodal inputs can reinforce modality dominance rather than mitigate it, as revealed by structured error-attractor patterns under systematic label perturbation, complemented by dynamical analysis.", "date": "2026-02-24", "categories": []}
{"id": "2411.01685", "title": "Reducing Biases in Record Matching Through Scores Calibration", "abstract": "Record matching models typically output a real-valued matching score that is later consumed through thresholding, ranking, or human review. While fairness in record matching has mostly been assessed using binary decisions at a fixed threshold, such evaluations can miss systematic disparities in the entire score distribution and can yield conclusions that change with the chosen threshold. We introduce a threshold-independent notion of score bias that extends standard group-fairness criteria-demographic parity (DP), equal opportunity (EO), and equalized odds (EOD)-from binary outputs to score functions by integrating group-wise metric gaps over all thresholds. Using this metric, we empirically show that several state-of-the-art deep matchers can exhibit substantial score bias even when appearing fair at commonly used thresholds. To mitigate these disparities without retraining the underlying matcher, we propose two model-agnostic post-processing methods that only require score evaluations on an (unlabeled) calibration set. Calib targets DP by aligning minority/majority score distributions to a common Wasserstein barycenter via a quantile-based optimal-transport map, with finite-sample guarantees on both residual DP bias and score distortion. C-Calib extends this idea to label-dependent notions (EO/EOD) by performing barycenter alignment conditionally on an estimated label, and we characterize how its guarantees depend on both sample size and label-estimation error. Experiments on standard record-matching benchmarks and multiple neural matchers confirm that Calib and C-Calib substantially reduce score bias with minimal loss in accuracy.", "date": "2024-11-03", "categories": []}
{"id": "2505.11111", "title": "FairSHAP: Preprocessing for Fairness Through Attribution-Based Data Augmentation", "abstract": "Ensuring fairness in machine learning models is critical, particularly in high-stakes domains where biased decisions can lead to serious societal consequences. Existing preprocessing approaches generally lack transparent mechanisms for identifying which features or instances are responsible for unfairness. This obscures the rationale behind data modifications. We introduce FairSHAP, a novel pre-processing framework that leverages Shapley value attribution to improve both individual and group fairness. FairSHAP identifies fairness-critical instances in the training data using an interpretable measure of feature importance, and systematically modifies them through instance-level matching across sensitive groups. This process reduces discriminative risk - an individual fairness metric - while preserving data integrity and model accuracy. We demonstrate that FairSHAP significantly improves demographic parity and equality of opportunity across diverse tabular datasets, achieving fairness gains with minimal data perturbation and, in some cases, improved predictive performance. As a model-agnostic and transparent method, FairSHAP integrates seamlessly into existing machine learning pipelines and provides actionable insights into the sources of bias.Our code is on https://github.com/ZhuMuMu0216/FairSHAP.", "date": "2025-05-16", "categories": []}
{"id": "2510.03734", "title": "Cost Efficient Fairness Audit Under Partial Feedback", "abstract": "We study the problem of auditing the fairness of a given classifier under partial feedback, where true labels are available only for positively classified individuals, (e.g., loan repayment outcomes are observed only for approved applicants). We introduce a novel cost model for acquiring additional labeled data, designed to more accurately reflect real-world costs such as credit assessment, loan processing, and potential defaults. Our goal is to find optimal fairness audit algorithms that are more cost-effective than random exploration and natural baselines.\n  In our work, we consider two audit settings: a black-box model with no assumptions on the data distribution, and a mixture model, where features and true labels follow a mixture of exponential family distributions. In the black-box setting, we propose a near-optimal auditing algorithm under mild assumptions and show that a natural baseline can be strictly suboptimal. In the mixture model setting, we design a novel algorithm that achieves significantly lower audit cost than the black-box case. Our approach leverages prior work on learning from truncated samples and maximum-a-posteriori oracles, and extends known results on spherical Gaussian mixtures to handle exponential family mixtures, which may be of independent interest. Moreover, our algorithms apply to popular fairness metrics including demographic parity, equal opportunity, and equalized odds. Empirically, we demonstrate strong performance of our algorithms on real-world fair classification datasets like Adult Income and Law School, consistently outperforming natural baselines by around 50% in terms of audit cost.", "date": "2025-10-04", "categories": []}
{"id": "2602.18870", "title": "Federated Measurement of Demographic Disparities from Quantile Sketches", "abstract": "Many fairness goals are defined at a population level that misaligns with siloed data collection, which remains unsharable due to privacy regulations. Horizontal federated learning (FL) enables collaborative modeling across clients with aligned features without sharing raw data. We study federated auditing of demographic parity through score distributions, measuring disparity as a Wasserstein--Frechet variance between sensitive-group score laws, and expressing the population metric in federated form that makes explicit how silo-specific selection drives local-global mismatch. For the squared Wasserstein distance, we prove an ANOVA-style decomposition that separates (i) selection-induced mixture effects from (ii) cross-silo heterogeneity, yielding tight bounds linking local and global metrics. We then propose a one-shot, communication-efficient protocol in which each silo shares only group counts and a quantile summary of its local score distributions, enabling the server to estimate global disparity and its decomposition, with $O(1/k)$ discretization bias ($k$ quantiles) and finite-sample guarantees. Experiments on synthetic data and COMPAS show that a few dozen quantiles suffice to recover global disparity and diagnose its sources.", "date": "2026-02-21", "categories": []}
{"id": "2503.07313", "title": "The influence of missing data mechanisms and simple missing data handling techniques on fairness", "abstract": "Machine learning algorithms permeate the day-to-day aspects of our lives and therefore studying the fairness of these algorithms before implementation is crucial. One way in which bias can manifest in a dataset is through missing values. Missing data are often assumed to be missing completely randomly; in reality the propensity of data being missing is often tied to the demographic characteristics of individuals. There is limited research into how missing values and the handling thereof can impact the fairness of an algorithm. Most researchers either apply listwise deletion or tend to use simpler methods of imputation (e.g. mean or mode) compared to more advanced approaches (e.g. multiple imputation). This study considers the fairness of various classification algorithms after a range of missing data handling strategies is applied. Missing values are generated (i.e. amputed) in three popular datasets for classification fairness, by creating a high percentage of missing values using three missing data mechanisms. The results show that the missing data mechanism does not significantly impact fairness; across the missing data handling techniques listwise deletion gives the highest fairness on average and amongst the classification algorithms random forests leads to the highest fairness on average. The interaction effect of the missing data handling technique and the classification algorithm is also often significant.", "date": "2025-03-10", "categories": []}
{"id": "2602.18201", "title": "SOMtime the World Ain$'$t Fair: Violating Fairness Using Self-Organizing Maps", "abstract": "Unsupervised representations are widely assumed to be neutral with respect to sensitive attributes when those attributes are withheld from training. We show that this assumption is false. Using SOMtime, a topology-preserving representation method based on high-capacity Self-Organizing Maps, we demonstrate that sensitive attributes such as age and income emerge as dominant latent axes in purely unsupervised embeddings, even when explicitly excluded from the input. On two large-scale real-world datasets (the World Values Survey across five countries and the Census-Income dataset), SOMtime recovers monotonic orderings aligned with withheld sensitive attributes, achieving Spearman correlations of up to 0.85, whereas PCA and UMAP typically remain below 0.23 (with a single exception reaching 0.31), and against t-SNE and autoencoders which achieve at most 0.34. Furthermore, unsupervised segmentation of SOMtime embeddings produces demographically skewed clusters, demonstrating downstream fairness risks without any supervised task. These findings establish that \\textit{fairness through unawareness} fails at the representation level for ordinal sensitive attributes and that fairness auditing must extend to unsupervised components of machine learning pipelines. We have made the code available at~ https://github.com/JosephBingham/SOMtime", "date": "2026-02-20", "categories": []}
{"id": "2411.06624", "title": "A Review of Fairness and A Practical Guide to Selecting Context-Appropriate Fairness Metrics in Machine Learning", "abstract": "Recent regulatory proposals for artificial intelligence emphasize fairness requirements for machine learning models. However, precisely defining the appropriate measure of fairness is challenging due to philosophical, cultural and political contexts. Biases can infiltrate machine learning models in complex ways depending on the model's context, rendering a single common metric of fairness insufficient. This ambiguity highlights the need for criteria to guide the selection of context-aware measures, an issue of increasing importance given the proliferation of ever tighter regulatory requirements. To address this, we developed a flowchart to guide the selection of contextually appropriate fairness measures. Twelve criteria were used to formulate the flowchart. This included consideration of model assessment criteria, model selection criteria, and data bias. We also review fairness literature in the context of machine learning and link it to core regulatory instruments to assist policymakers, AI developers, researchers, and other stakeholders in appropriately addressing fairness concerns and complying with relevant regulatory requirements.", "date": "2024-11-10", "categories": []}
{"id": "2512.23769", "title": "Uncovering Discrimination Clusters: Quantifying and Explaining Systematic Fairness Violations", "abstract": "Fairness in algorithmic decision-making is often framed in terms of individual fairness, which requires that similar individuals receive similar outcomes. A system violates individual fairness if there exists a pair of inputs differing only in protected attributes (such as race or gender) that lead to significantly different outcomes-for example, one favorable and the other unfavorable. While this notion highlights isolated instances of unfairness, it fails to capture broader patterns of systematic or clustered discrimination that may affect entire subgroups. We introduce and motivate the concept of discrimination clustering, a generalization of individual fairness violations. Rather than detecting single counterfactual disparities, we seek to uncover regions of the input space where small perturbations in protected features lead to k-significantly distinct clusters of outcomes. That is, for a given input, we identify a local neighborhood-differing only in protected attributes-whose members' outputs separate into many distinct clusters. These clusters reveal significant arbitrariness in treatment solely based on protected attributes that help expose patterns of algorithmic bias that elude pairwise fairness checks. We present HyFair, a hybrid technique that combines formal symbolic analysis (via SMT and MILP solvers) to certify individual fairness with randomized search to discover discriminatory clusters. This combination enables both formal guarantees-when no counterexamples exist-and the detection of severe violations that are computationally challenging for symbolic methods alone. Given a set of inputs exhibiting high k-unfairness, we introduce a novel explanation method to generate interpretable, decision-tree-style artifacts. Our experiments demonstrate that HyFair outperforms state-of-the-art fairness verification and local explanation methods.", "date": "2025-12-29", "categories": []}
{"id": "2602.16438", "title": "Intra-Fairness Dynamics: The Bias Spillover Effect in Targeted LLM Alignment", "abstract": "Conventional large language model (LLM) fairness alignment largely focuses on mitigating bias along single sensitive attributes, overlooking fairness as an inherently multidimensional and context-specific value. This approach risks creating systems that achieve narrow fairness metrics while exacerbating disparities along untargeted attributes, a phenomenon known as bias spillover. While extensively studied in machine learning, bias spillover remains critically underexplored in LLM alignment. In this work, we investigate how targeted gender alignment affects fairness across nine sensitive attributes in three state-of-the-art LLMs (Mistral 7B, Llama 3.1 8B, Qwen 2.5 7B). Using Direct Preference Optimization and the BBQ benchmark, we evaluate fairness under ambiguous and disambiguous contexts. Our findings reveal noticeable bias spillover: while aggregate results show improvements, context-aware analysis exposes significant degradations in ambiguous contexts, particularly for physical appearance ($p< 0.001$ across all models), sexual orientation, and disability status. We demonstrate that improving fairness along one attribute can inadvertently worsen disparities in others under uncertainty, highlighting the necessity of context-aware, multi-attribute fairness evaluation frameworks.", "date": "2026-02-18", "categories": []}
{"id": "2504.15206", "title": "How Global Calibration Strengthens Multiaccuracy", "abstract": "Multiaccuracy and multicalibration are multigroup fairness notions for prediction that have found numerous applications in learning and computational complexity. They can be achieved from a single learning primitive: weak agnostic learning. Here we investigate the power of multiaccuracy as a learning primitive, both with and without the additional assumption of calibration. We find that multiaccuracy in itself is rather weak, but that the addition of global calibration (this notion is called calibrated multiaccuracy) boosts its power substantially, enough to recover implications that were previously known only assuming the stronger notion of multicalibration.\n  We give evidence that multiaccuracy might not be as powerful as standard weak agnostic learning, by showing that there is no way to post-process a multiaccurate predictor to get a weak learner, even assuming the best hypothesis has correlation $1/2$. Rather, we show that it yields a restricted form of weak agnostic learning, which requires some concept in the class to have correlation greater than $1/2$ with the labels. However, by also requiring the predictor to be calibrated, we recover not just weak, but strong agnostic learning.\n  A similar picture emerges when we consider the derivation of hardcore measures from predictors satisfying multigroup fairness notions. On the one hand, while multiaccuracy only yields hardcore measures of density half the optimal, we show that (a weighted version of) calibrated multiaccuracy achieves optimal density.\n  Our results yield new insights into the complementary roles played by multiaccuracy and calibration in each setting. They shed light on why multiaccuracy and global calibration, although not particularly powerful by themselves, together yield considerably stronger notions.", "date": "2025-04-21", "categories": []}
{"id": "2502.01713", "title": "Auditing a Dutch Public Sector Risk Profiling Algorithm Using an Unsupervised Bias Detection Tool", "abstract": "Algorithms are increasingly used to automate or aid human decisions, yet recent research shows that these algorithms may exhibit bias across legally protected demographic groups. However, data on these groups may be unavailable to organizations or external auditors due to privacy legislation. This paper studies bias detection using an unsupervised bias detection tool when data on demographic groups are unavailable. We collaborated with the Dutch Executive Agency for Education to audit an algorithm that was used to assign risk scores to college students at the national level in the Netherlands between 2012-2023. Our audit covers more than 250,000 students across the country. The unsupervised bias detection tool highlights known disparities between students with a non-European migration background and students with a Dutch or European-migration background. Our contributions are two-fold: (1) we assess bias in a real-world, large-scale, and high-stakes decision-making process by a governmental organization; (2) we provide the unsupervised bias detection tool in an open-source library for others to use to complete bias audits. Our work serves as a starting point for a deliberative assessment by human experts to evaluate potential discrimination in algorithmic decision-making.", "date": "2025-02-03", "categories": []}
{"id": "2508.17622", "title": "The Statistical Fairness-Accuracy Frontier", "abstract": "We study fairness-accuracy tradeoffs when a single predictive model must serve multiple demographic groups. A useful tool for understanding this tradeoff is the fairness-accuracy (FA) Pareto frontier, which characterizes the set of models that cannot be improved in either fairness or accuracy without worsening the other. While characterizing the FA frontier requires full knowledge of the data distribution, we focus on the finite-sample regime, quantifying how well a designer can approximate any point on the frontier from limited data and bounding the worst-case gap. In particular, we derive worst-case-optimal estimators that depend on the designer's knowledge of the covariate distribution. For each estimator, we characterize how finite-sample effects asymmetrically impact each group's welfare and identify optimal sample allocation strategies. Finally, we provide uniform finite-sample bounds for the entire FA frontier, yielding confidence bands that quantify the reliability of welfare comparisons across alternative fairness-accuracy tradeoffs.", "date": "2025-08-25", "categories": []}
{"id": "2602.13246", "title": "Global AI Bias Audit for Technical Governance", "abstract": "This paper presents the outputs of the exploratory phase of a global audit of Large Language Models (LLMs) project. In this exploratory phase, I used the Global AI Dataset (GAID) Project as a framework to stress-test the Llama-3 8B model and evaluate geographic and socioeconomic biases in technical AI governance awareness. By stress-testing the model with 1,704 queries across 213 countries and eight technical metrics, I identified a significant digital barrier and gap separating the Global North and South. The results indicate that the model was only able to provide number/fact responses in 11.4% of its query answers, where the empirical validity of such responses was yet to be verified. The findings reveal that AI's technical knowledge is heavily concentrated in higher-income regions, while lower-income countries from the Global South are subject to disproportionate systemic information gaps. This disparity between the Global North and South poses concerning risks for global AI safety and inclusive governance, as policymakers in underserved regions may lack reliable data-driven insights or be misled by hallucinated facts. This paper concludes that current AI alignment and training processes reinforce existing geoeconomic and geopolitical asymmetries, and urges the need for more inclusive data representation to ensure AI serves as a truly global resource.", "date": "2026-02-01", "categories": []}
{"id": "2602.13651", "title": "Cumulative Utility Parity for Fair Federated Learning under Intermittent Client Participation", "abstract": "In real-world federated learning (FL) systems, client participation is intermittent, heterogeneous, and often correlated with data characteristics or resource constraints. Existing fairness approaches in FL primarily focus on equalizing loss or accuracy conditional on participation, implicitly assuming that clients have comparable opportunities to contribute over time. However, when participation itself is uneven, these objectives can lead to systematic under-representation of intermittently available clients, even if per-round performance appears fair. We propose cumulative utility parity, a fairness principle that evaluates whether clients receive comparable long-term benefit per participation opportunity, rather than per training round. To operationalize this notion, we introduce availability-normalized cumulative utility, which disentangles unavoidable physical constraints from avoidable algorithmic bias arising from scheduling and aggregation. Experiments on temporally skewed, non-IID federated benchmarks demonstrate that our approach substantially improves long-term representation parity, while maintaining near-perfect performance.", "date": "2026-02-14", "categories": []}
{"id": "2602.14970", "title": "Counterfactual Fairness Evaluation of LLM-Based Contact Center Agent Quality Assurance System", "abstract": "Large Language Models (LLMs) are increasingly deployed in contact-center Quality Assurance (QA) to automate agent performance evaluation and coaching feedback. While LLMs offer unprecedented scalability and speed, their reliance on web-scale training data raises concerns regarding demographic and behavioral biases that may distort workforce assessment. We present a counterfactual fairness evaluation of LLM-based QA systems across 13 dimensions spanning three categories: Identity, Context, and Behavioral Style. Fairness is quantified using the Counterfactual Flip Rate (CFR), the frequency of binary judgment reversals, and the Mean Absolute Score Difference (MASD), the average shift in coaching or confidence scores across counterfactual pairs. Evaluating 18 LLMs on 3,000 real-world contact center transcripts, we find systematic disparities, with CFR ranging from 5.4% to 13.0% and consistent MASD shifts across confidence, positive, and improvement scores. Larger, more strongly aligned models show lower unfairness, though fairness does not track accuracy. Contextual priming of historical performance induces the most severe degradations (CFR up to 16.4%), while implicit linguistic identity cues remain a persistent bias source. Finally, we analyze the efficacy of fairness-aware prompting, finding that explicit instructions yield only modest improvements in evaluative consistency. Our findings underscore the need for standardized fairness auditing pipelines prior to deploying LLMs in high-stakes workforce evaluation.", "date": "2026-02-16", "categories": []}
{"id": "2505.24262", "title": "On Fairness of Task Arithmetic: The Role of Task Vectors", "abstract": "Model editing techniques, particularly task arithmetic with task vectors, offer an efficient alternative to full fine-tuning by enabling direct parameter updates through simple arithmetic operations. While this approach promises substantial computational savings, its impact on fairness has remained largely unexplored -- despite growing concern over biased outcomes in high-stakes applications such as hate speech detection. In this work, we present the first systematic study of group fairness in task arithmetic within this binary text and image classification regime, comparing it against full fine-tuning (FFT) and Low-Rank Adaptation (LoRA). We evaluate across multiple language models and datasets using standard group fairness metrics, including Demographic Parity and Equalized Odds. Our analysis shows that task vectors can be tuned to achieve competitive accuracy while reducing disparities, and that merging subgroup-specific task vectors provides a practical mechanism for steering fairness outcomes. We further provide a theoretical bound linking task vector scaling to fairness metrics, offering insight into the observed trade-offs. Together, these findings establish task arithmetic not only as a cost-efficient editing method but also as a fairness-aware alternative to existing adaptation techniques, within the standard group-fair classification setting, laying the groundwork for responsible deployment of large language models.", "date": "2025-05-30", "categories": []}
{"id": "2512.15926", "title": "DSO: Direct Steering Optimization for Bias Mitigation", "abstract": "Generative models are often deployed to make decisions on behalf of users, such as vision-language models (VLMs) identifying which person in a room is a doctor to help visually impaired individuals. Yet, VLM decisions are influenced by the perceived demographic attributes of people in the input, which can lead to biased outcomes like failing to identify women as doctors. Moreover, when reducing bias leads to performance loss, users may have varying needs for balancing bias mitigation with overall model capabilities, highlighting the demand for methods that enable controllable bias reduction during inference. Activation steering is a popular approach for inference-time controllability that has shown potential in inducing safer behavior in large language models (LLMs). However, we observe that current steering methods struggle to correct biases, where equiprobable outcomes across demographic groups are required. To address this, we propose Direct Steering Optimization (DSO) which uses reinforcement learning to find linear transformations for steering activations, tailored to mitigate bias while maintaining control over model performance. We demonstrate that DSO achieves state-of-the-art trade-off between fairness and capabilities on both VLMs and LLMs, while offering practitioners inference-time control over the trade-off. Overall, our work highlights the benefit of designing steering strategies that are directly optimized to control model behavior, providing more effective bias intervention than methods that rely on pre-defined heuristics for controllability.", "date": "2025-12-17", "categories": []}
{"id": "2601.20269", "title": "Empirical Likelihood-Based Fairness Auditing: Distribution-Free Certification and Flagging", "abstract": "Machine learning models in high-stakes applications, such as recidivism prediction and automated personnel selection, often exhibit systematic performance disparities across sensitive subpopulations, raising critical concerns regarding algorithmic bias. Fairness auditing addresses these risks through two primary functions: certification, which verifies adherence to fairness constraints; and flagging, which isolates specific demographic groups experiencing disparate treatment. However, existing auditing techniques are frequently limited by restrictive distributional assumptions or prohibitive computational overhead. We propose a novel empirical likelihood-based (EL) framework that constructs robust statistical measures for model performance disparities. Unlike traditional methods, our approach is non-parametric; the proposed disparity statistics follow asymptotically chi-square or mixed chi-square distributions, ensuring valid inference without assuming underlying data distributions. This framework uses a constrained optimization profile that admits stable numerical solutions, facilitating both large-scale certification and efficient subpopulation discovery. Empirically, the EL methods outperform bootstrap-based approaches, yielding coverage rates closer to nominal levels while reducing computational latency by several orders of magnitude. We demonstrate the practical utility of this framework on the COMPAS dataset, where it successfully flags intersectional biases, specifically identifying a significantly higher positive prediction rate for African-American males under 25 and a systemic under-prediction for Caucasian females relative to the population mean.", "date": "2026-01-28", "categories": []}
{"id": "2602.11500", "title": "A Generic Framework for Fair Consensus Clustering in Streams", "abstract": "Consensus clustering seeks to combine multiple clusterings of the same dataset, potentially derived by considering various non-sensitive attributes by different agents in a multi-agent environment, into a single partitioning that best reflects the overall structure of the underlying dataset. Recent work by Chakraborty et al, introduced a fair variant under proportionate fairness and obtained a constant-factor approximation by naively selecting the best closest fair input clustering; however, their offline approach requires storing all input clusterings, which is prohibitively expensive for most large-scale applications.\n  In this paper, we initiate the study of fair consensus clustering in the streaming model, where input clusterings arrive sequentially and memory is limited. We design the first constant-factor algorithm that processes the stream while storing only a logarithmic number of inputs. En route, we introduce a new generic algorithmic framework that integrates closest fair clustering with cluster fitting, yielding improved approximation guarantees not only in the streaming setting but also when revisited offline. Furthermore, the framework is fairness-agnostic: it applies to any fairness definition for which an approximately close fair clustering can be computed efficiently. Finally, we extend our methods to the more general k-median consensus clustering problem.", "date": "2026-02-12", "categories": []}
{"id": "2602.11722", "title": "PAC-Bayesian Generalization Guarantees for Fairness on Stochastic and Deterministic Classifiers", "abstract": "Classical PAC generalization bounds on the prediction risk of a classifier are insufficient to provide theoretical guarantees on fairness when the goal is to learn models balancing predictive risk and fairness constraints. We propose a PAC-Bayesian framework for deriving generalization bounds for fairness, covering both stochastic and deterministic classifiers. For stochastic classifiers, we derive a fairness bound using standard PAC-Bayes techniques. Whereas for deterministic classifiers, as usual PAC-Bayes arguments do not apply directly, we leverage a recent advance in PAC-Bayes to extend the fairness bound beyond the stochastic setting. Our framework has two advantages: (i) It applies to a broad class of fairness measures that can be expressed as a risk discrepancy, and (ii) it leads to a self-bounding algorithm in which the learning procedure directly optimizes a trade-off between generalization bounds on the prediction risk and on the fairness. We empirically evaluate our framework with three classical fairness measures, demonstrating not only its usefulness but also the tightness of our bounds.", "date": "2026-02-12", "categories": []}
{"id": "2602.11785", "title": "Safe Fairness Guarantees Without Demographics in Classification: Spectral Uncertainty Set Perspective", "abstract": "As automated classification systems become increasingly prevalent, concerns have emerged over their potential to reinforce and amplify existing societal biases. In the light of this issue, many methods have been proposed to enhance the fairness guarantees of classifiers. Most of the existing interventions assume access to group information for all instances, a requirement rarely met in practice. Fairness without access to demographic information has often been approached through robust optimization techniques,which target worst-case outcomes over a set of plausible distributions known as the uncertainty set. However, their effectiveness is strongly influenced by the chosen uncertainty set. In fact, existing approaches often overemphasize outliers or overly pessimistic scenarios, compromising both overall performance and fairness. To overcome these limitations, we introduce SPECTRE, a minimax-fair method that adjusts the spectrum of a simple Fourier feature mapping and constrains the extent to which the worst-case distribution can deviate from the empirical distribution. We perform extensive experiments on the American Community Survey datasets involving 20 states. The safeness of SPECTRE comes as it provides the highest average values on fairness guarantees together with the smallest interquartile range in comparison to state-of-the-art approaches, even compared to those with access to demographic group information. In addition, we provide a theoretical analysis that derives computable bounds on the worst-case error for both individual groups and the overall population, as well as characterizes the worst-case distributions responsible for these extremal performances", "date": "2026-02-12", "categories": []}
{"id": "2602.12133", "title": "Neutral Prompts, Non-Neutral People: Quantifying Gender and Skin-Tone Bias in Gemini Flash 2.5 Image and GPT Image 1.5", "abstract": "This study quantifies gender and skin-tone bias in two widely deployed commercial image generators - Gemini Flash 2.5 Image (NanoBanana) and GPT Image 1.5 - to test the assumption that neutral prompts yield demographically neutral outputs. We generated 3,200 photorealistic images using four semantically neutral prompts. The analysis employed a rigorous pipeline combining hybrid color normalization, facial landmark masking, and perceptually uniform skin tone quantification using the Monk (MST), PERLA, and Fitzpatrick scales. Neutral prompts produced highly polarized defaults. Both models exhibited a strong \"default white\" bias (>96% of outputs). However, they diverged sharply on gender: Gemini favored female-presenting subjects, while GPT favored male-presenting subjects with lighter skin tones. This research provides a large-scale, comparative audit of state-of-the-art models using an illumination-aware colorimetric methodology, distinguishing aesthetic rendering from underlying pigmentation in synthetic imagery. The study demonstrates that neutral prompts function as diagnostic probes rather than neutral instructions. It offers a robust framework for auditing algorithmic visual culture and challenges the sociolinguistic assumption that unmarked language results in inclusive representation.", "date": "2026-02-12", "categories": []}
{"id": "2505.23798", "title": "Unveiling the \"Fairness Seesaw\": Discovering and Mitigating Gender and Race Bias in Vision-Language Models", "abstract": "Although Vision-Language Models (VLMs) have achieved remarkable success, the knowledge mechanisms underlying their social biases remain a black box, where fairness- and ethics-related problems harm certain groups of people in society. It is unknown to what extent VLMs yield gender and race bias in generative responses. In this paper, we conduct a systematic discovery of gender and race bias in state-of-the-art VLMs, focusing not only on surface-level responses but also on the internal probability distributions and hidden state dynamics. Our empirical analysis reveals three critical findings: 1) The Fairness Paradox: Models often generate fair text labels while maintaining highly skewed confidence scores (mis-calibration) toward specific social groups. 2) Layer-wise Fluctuation: Fairness knowledge is not uniformly distributed; it peaks in intermediate layers and undergoes substantial knowledge erosion in the final layers. 3) Residual Discrepancy: Within a single hidden layer, different residual streams carry conflicting social knowledge - some reinforcing fairness while others amplifying bias. Leveraging these insights, we propose RES-FAIR (RESidual Flow Adjustment for Inference Recalibration), a post-hoc framework that mitigates bias by localizing and projecting hidden states away from biased residual directions while amplifying fair components. Evaluations on PAIRS and SocialCounterfactuals datasets demonstrate that our discovery-based approach significantly improves response fairness and confidence calibration without compromising general reasoning abilities. Our work provides a new lens for understanding how multi-modal models store and process sensitive social information.", "date": "2025-05-26", "categories": []}
{"id": "2509.22291", "title": "Bridging Fairness and Explainability: Can Input-Based Explanations Promote Fairness in Hate Speech Detection?", "abstract": "Natural language processing (NLP) models often replicate or amplify social bias from training data, raising concerns about fairness. At the same time, their black-box nature makes it difficult for users to recognize biased predictions and for developers to effectively mitigate them. While some studies suggest that input-based explanations can help detect and mitigate bias, others question their reliability in ensuring fairness. Existing research on explainability in fair NLP has been predominantly qualitative, with limited large-scale quantitative analysis. In this work, we conduct the first systematic study of the relationship between explainability and fairness in hate speech detection, focusing on both encoder- and decoder-only models. We examine three key dimensions: (1) identifying biased predictions, (2) selecting fair models, and (3) mitigating bias during model training. Our findings show that input-based explanations can effectively detect biased predictions and serve as useful supervision for reducing bias during training, but they are unreliable for selecting fair models among candidates.Our code is available at https://github.com/Ewanwong/fairness_x_explainability.", "date": "2025-09-26", "categories": []}
{"id": "2601.23001", "title": "Bias Beyond Borders: Political Ideology Evaluation and Steering in Multilingual LLMs", "abstract": "Large Language Models (LLMs) increasingly shape global discourse, making fairness and ideological neutrality essential for responsible AI deployment. Despite growing attention to political bias in LLMs, prior work largely focuses on high-resource, Western languages or narrow multilingual settings, leaving cross-lingual consistency and safe post-hoc mitigation underexplored. To address this gap, we present a large-scale multilingual evaluation of political bias spanning 50 countries and 33 languages. We introduce a complementary post-hoc mitigation framework, Cross-Lingual Alignment Steering (CLAS), designed to augment existing steering methods by aligning ideological representations across languages and dynamically regulating intervention strength. This method aligns latent ideological representations induced by political prompts into a shared ideological subspace, ensuring cross lingual consistency, with the adaptive mechanism prevents over correction and preserves coherence. Experiments demonstrate substantial bias reduction along both economic and social axes with minimal degradation in response quality. The proposed framework establishes a scalable and interpretable paradigm for fairness-aware multilingual LLM governance, balancing ideological neutrality with linguistic and cultural diversity.", "date": "2026-01-30", "categories": []}
{"id": "2602.10457", "title": "Analyzing Fairness of Neural Network Prediction via Counterfactual Dataset Generation", "abstract": "Interpreting the inference-time behavior of deep neural networks remains a challenging problem. Existing approaches to counterfactual explanation typically ask: What is the closest alternative input that would alter the model's prediction in a desired way? In contrast, we explore counterfactual datasets. Rather than perturbing the input, our method efficiently finds the closest alternative training dataset, one that differs from the original dataset by changing a few labels. Training a new model on this altered dataset can then lead to a different prediction of a given test instance. This perspective provides a new way to assess fairness by directly analyzing the influence of label bias on training and inference. Our approach can be characterized as probing whether a given prediction depends on biased labels. Since exhaustively enumerating all possible alternate datasets is infeasible, we develop analysis techniques that trace how bias in the training data may propagate through the learning algorithm to the trained network. Our method heuristically ranks and modifies the labels of a bounded number of training examples to construct a counterfactual dataset, retrains the model, and checks whether its prediction on a chosen test case changes. We evaluate our approach on feedforward neural networks across over 1100 test cases from 7 widely-used fairness datasets. Results show that it modifies only a small subset of training labels, highlighting its ability to pinpoint the critical training examples that drive prediction changes. Finally, we demonstrate how our counterfactual datasets reveal connections between training examples and test cases, offering an interpretable way to probe dataset bias.", "date": "2026-02-11", "categories": []}
{"id": "2510.04114", "title": "Wasserstein projection distance for fairness testing of regression models", "abstract": "Fairness testing evaluates whether a model satisfies a specified fairness criterion across different groups, yet most research has focused on classification models, leaving regression models underexplored. This paper introduces a framework for fairness testing in regression models, leveraging Wasserstein distance to project data distribution and focusing on expectation-based criteria. Upon categorizing fairness criteria for regression, we derive a Wasserstein projection test statistic from dual reformulation, and derive asymptotic bounds and limiting distributions, allowing us to formulate both a hypothesis-testing procedure and an optimal data perturbation method to improve fairness while balancing accuracy. Experiments on synthetic data demonstrate that the proposed hypothesis-testing approach offers higher specificity compared to permutation-based tests. To illustrate its potential applications, we apply our framework to two case studies on real data, showing (1) statistically significant gender disparities that appear on student performance data across multiple models, and (2) significant unfairness between pollution areas under multiple fairness criteria affecting housing price data, robust to different group divisions, with feature-level analysis identifying spatial and socioeconomic drivers.", "date": "2025-10-05", "categories": []}
{"id": "2601.19186", "title": "Double Fairness Policy Learning: Integrating Action Fairness and Outcome Fairness in Decision-making", "abstract": "Fairness is a central pillar of trustworthy machine learning, especially in domains where accuracy- or profit-driven optimization is insufficient. While most fairness research focuses on supervised learning, fairness in policy learning remains less explored. Because policy learning is interventional, it induces two distinct fairness targets: action fairness (equitable action assignments) and outcome fairness (equitable downstream consequences). Crucially, equalizing actions does not generally equalize outcomes when groups face different constraints or respond differently to the same action. We propose a novel double fairness learning (DFL) framework that explicitly manages the trade-off among three objectives: action fairness, outcome fairness, and value maximization. We integrate fairness directly into a multi-objective optimization problem for policy learning and employ a lexicographic weighted Tchebyshev method that recovers Pareto solutions beyond convex settings, with theoretical guarantees on the regret bounds. Our framework is flexible and accommodates various commonly used fairness notions. Extensive simulations demonstrate improved performance relative to competing methods. In applications to a motor third-party liability insurance dataset and an entrepreneurship training dataset, DFL substantially improves both action and outcome fairness while incurring only a modest reduction in overall value.", "date": "2026-01-27", "categories": []}
{"id": "2602.09196", "title": "Fair Feature Importance Scores via Feature Occlusion and Permutation", "abstract": "As machine learning models increasingly impact society, their opaque nature poses challenges to trust and accountability, particularly in fairness contexts. Understanding how individual features influence model outcomes is crucial for building interpretable and equitable models. While feature importance metrics for accuracy are well-established, methods for assessing feature contributions to fairness remain underexplored. We propose two model-agnostic approaches to measure fair feature importance. First, we propose to compare model fairness before and after permuting feature values. This simple intervention-based approach decouples a feature and model predictions to measure its contribution to training. Second, we evaluate the fairness of models trained with and without a given feature. This occlusion-based score enjoys dramatic computational simplification via minipatch learning. Our empirical results reflect the simplicity and effectiveness of our proposed metrics for multiple predictive tasks. Both methods offer simple, scalable, and interpretable solutions to quantify the influence of features on fairness, providing new tools for responsible machine learning development.", "date": "2026-02-09", "categories": []}
{"id": "2502.18534", "title": "MAFE: Enabling Equitable Algorithm Design in Multi-Agent Multi-Stage Decision-Making Systems", "abstract": "Algorithmic fairness is often studied in static or single-agent settings, yet many real-world decision-making systems involve multiple interacting entities whose multi-stage actions jointly influence long-term outcomes. Existing fairness methods applied at isolated decision points frequently fail to mitigate disparities that accumulate over time. Although recent work has modeled fairness as a sequential decision-making problem, it typically assumes centralized agents or simplified dynamics, limiting its applicability to complex social systems. We introduce MAFE, a suite of Multi-Agent Fair Environments designed to simulate realistic, modular, and dynamic systems in which fairness emerges from the interplay of multiple agents. We demonstrate MAFEs across three domains -- loan processing, healthcare, and higher education -- that support heterogeneous agents, configurable interventions, and fairness metrics. The environments are open-source and compatible with standard multi-agent reinforcement learning (MARL) libraries, enabling reproducible evaluation of fairness-aware policies. Through extensive experiments on cooperative use cases, we demonstrate how MAFE facilitates the design of equitable multi-agent algorithms and reveals critical trade-offs between fairness, performance, and coordination. MAFE provides a foundation for systematic progress in dynamic, multi-agent fairness research.", "date": "2025-02-25", "categories": []}
{"id": "2505.22984", "title": "A Computational Approach to Improving Fairness in K-means Clustering", "abstract": "The popular K-means clustering algorithm potentially suffers from a major weakness for further analysis or interpretation. Some cluster may have disproportionately more (or fewer) points from one of the subpopulations in terms of some sensitive variable, e.g., gender or race. Such a fairness issue may cause bias and unexpected social consequences. This work attempts to improve the fairness of K-means clustering with a two-stage optimization formulation--clustering first and then adjust cluster membership of a small subset of selected data points. Two computationally efficient algorithms are proposed in identifying those data points that are expensive for fairness, with one focusing on nearest data points outside of a cluster and the other on highly 'mixed' data points. Experiments on benchmark datasets show substantial improvement on fairness with a minimal impact to clustering quality. The proposed algorithms can be easily extended to a broad class of clustering algorithms or fairness metrics.", "date": "2025-05-29", "categories": []}
{"id": "2510.23935", "title": "Understanding Fairness and Prediction Error through Subspace Decomposition and Influence Analysis", "abstract": "Machine learning models have achieved widespread success but often inherit and amplify historical biases, resulting in unfair outcomes. Traditional fairness methods typically impose constraints at the prediction level, without addressing underlying biases in data representations. In this work, we propose a principled framework that adjusts data representations to balance predictive utility and fairness. Using sufficient dimension reduction, we decompose the feature space into target-relevant, sensitive, and shared components, and control the fairness-utility trade-off by selectively removing sensitive information. We provide a theoretical analysis of how prediction error and fairness gaps evolve as shared subspaces are added, and employ influence functions to quantify their effects on the asymptotic behavior of parameter estimates. Experiments on both synthetic and real-world datasets validate our theoretical insights and show that the proposed method effectively improves fairness while preserving predictive performance.", "date": "2025-10-27", "categories": []}
{"id": "2602.06984", "title": "Empowering Affected Individuals to Shape AI Fairness Assessments: Processes, Criteria, and Tools", "abstract": "AI systems are increasingly used in high-stakes domains such as credit rating, where fairness concerns are critical. Existing fairness assessments are typically conducted by AI experts or regulators using predefined protected attributes and metrics, which often fail to capture the diversity and nuance of fairness notions held by the individuals who are affected by these systems' decisions, such as decision subjects. Recent work has therefore called for involving affected individuals in fairness assessment, yet little empirical evidence exists on how they create their own fairness criteria or what kinds of criteria they produce - knowledge that could not only inform experts' fairness evaluation and mitigation, but also guide the design of AI assessment tools. We address this gap through a qualitative user study with 18 participants in a credit rating scenario. Participants first articulated their fairness notions in their own words. Then, participants turned them into concrete quantified and operationalized fairness criteria, through an interactive prototype we designed. Our findings provide empirical evidence of the process through which people's fairness notions emerge via grounding in model features, and uncover a diverse set of individuals' custom-defined criteria for both outcome and procedural fairness. We provide design implications for processes and tools that support more inclusive and value-sensitive AI fairness assessment.", "date": "2026-01-27", "categories": []}
{"id": "2602.07710", "title": "On Generation in Metric Spaces", "abstract": "We study generation in separable metric instance spaces. We extend the language generation framework from Kleinberg and Mullainathan [2024] beyond countable domains by defining novelty through metric separation and allowing asymmetric novelty parameters for the adversary and the generator. We introduce the $(\\varepsilon,\\varepsilon')$-closure dimension, a scale-sensitive analogue of closure dimension, which yields characterizations of uniform and non-uniform generatability and a sufficient condition for generation in the limit. Along the way, we identify a sharp geometric contrast. Namely, in doubling spaces, including all finite-dimensional normed spaces, generatability is stable across novelty scales and invariant under equivalent metrics. In general metric spaces, however, generatability can be highly scale-sensitive and metric-dependent; even in the natural infinite-dimensional Hilbert space $\\ell^2$, all notions of generation may fail abruptly as the novelty parameters vary.", "date": "2026-02-07", "categories": []}
{"id": "2602.07799", "title": "Fairness Aware Reward Optimization", "abstract": "Demographic skews in human preference data propagate systematic unfairness through reward models into aligned LLMs. We introduce Fairness Aware Reward Optimization (Faro), an in-processing framework that trains reward models under demographic parity, equalized odds, or counterfactual fairness constraints. We provide the first theoretical analysis of reward-level fairness in LLM alignment, establishing: (i) provable fairness certificates for Faro-trained rewards with controllable slack; a (ii) formal characterization of the accuracy-fairness trade-off induced by KL-regularized fine-tuning, proving fairness transfers from reward to policy; and the (iii) existence of a non-empty Pareto frontier. Unlike pre- and post-processing methods, Faro ensures reward models are simultaneously ordinal (ranking correctly), cardinal (calibrated), and fair. Across multiple LLMs and benchmarks, Faro significantly reduces bias and harmful generations while maintaining or improving model quality.", "date": "2026-02-08", "categories": []}
{"id": "2602.08124", "title": "Gender and Race Bias in Consumer Product Recommendations by Large Language Models", "abstract": "Large Language Models are increasingly employed in generating consumer product recommendations, yet their potential for embedding and amplifying gender and race biases remains underexplored. This paper serves as one of the first attempts to examine these biases within LLM-generated recommendations. We leverage prompt engineering to elicit product suggestions from LLMs for various race and gender groups and employ three analytical methods-Marked Words, Support Vector Machines, and Jensen-Shannon Divergence-to identify and quantify biases. Our findings reveal significant disparities in the recommendations for demographic groups, underscoring the need for more equitable LLM recommendation systems.", "date": "2026-02-08", "categories": []}
{"id": "2602.08589", "title": "FairRARI: A Plug and Play Framework for Fairness-Aware PageRank", "abstract": "PageRank (PR) is a fundamental algorithm in graph machine learning tasks. Owing to the increasing importance of algorithmic fairness, we consider the problem of computing PR vectors subject to various group-fairness criteria based on sensitive attributes of the vertices. At present, principled algorithms for this problem are lacking - some cannot guarantee that a target fairness level is achieved, while others do not feature optimality guarantees. In order to overcome these shortcomings, we put forth a unified in-processing convex optimization framework, termed FairRARI, for tackling different group-fairness criteria in a ``plug and play'' fashion. Leveraging a variational formulation of PR, the framework computes fair PR vectors by solving a strongly convex optimization problem with fairness constraints, thereby ensuring that a target fairness level is achieved. We further introduce three different fairness criteria which can be efficiently tackled using FairRARI to compute fair PR vectors with the same asymptotic time-complexity as the original PR algorithm. Extensive experiments on real-world datasets showcase that FairRARI outperforms existing methods in terms of utility, while achieving the desired fairness levels across multiple vertex groups; thereby highlighting its effectiveness.", "date": "2026-02-09", "categories": []}
{"id": "2602.08660", "title": "Equalized Generative Treatment: Matching f-divergences for Fairness in Generative Models", "abstract": "Fairness is a crucial concern for generative models, which not only reflect but can also amplify societal and cultural biases. Existing fairness notions for generative models are largely adapted from classification and focus on balancing the probability of generating samples from each sensitive group. We show that such criteria are brittle, as they can be met even when different sensitive groups are modeled with widely varying quality. To address this limitation, we introduce a new fairness definition for generative models, termed as equalized generative treatment (EGT), which requires comparable generation quality across all sensitive groups, with quality measured via a reference f-divergence. We further analyze the trade-offs induced by EGT, demonstrating that enforcing fairness constraints necessarily couples the overall model quality to that of the most challenging group to approximate. This indicates that a simple yet efficient min-max fine-tuning method should be able to balance f-divergences across sensitive groups to satisfy EGT. We validate this theoretical insight through a set of experiments on both image and text generation tasks. We demonstrate that min-max methods consistently achieve fairer outcomes compared to other approaches from the literature, while maintaining competitive overall performance for both tasks.", "date": "2026-02-09", "categories": []}
{"id": "2602.06181", "title": "Uncertainty Drives Social Bias Changes in Quantized Large Language Models", "abstract": "Post-training quantization reduces the computational cost of large language models but fundamentally alters their social biases in ways that aggregate metrics fail to capture. We present the first large-scale study of 50 quantized models evaluated on PostTrainingBiasBench, a unified benchmark of 13 closed- and open-ended bias datasets. We identify a phenomenon we term quantization-induced masked bias flipping, in which up to 21% of responses flip between biased and unbiased states after quantization, despite showing no change in aggregate bias scores. These flips are strongly driven by model uncertainty, where the responses with high uncertainty are 3-11x more likely to change than the confident ones. Quantization strength amplifies this effect, with 4-bit quantized models exhibiting 4-6x more behavioral changes than 8-bit quantized models. Critically, these changes create asymmetric impacts across demographic groups, where bias can worsen by up to 18.6% for some groups while improving by 14.1% for others, yielding misleadingly neutral aggregate outcomes. Larger models show no consistent robustness advantage, and group-specific shifts vary unpredictably across model families. Our findings demonstrate that compression fundamentally alters bias patterns, requiring crucial post-quantization evaluation and interventions to ensure reliability in practice.", "date": "2026-02-05", "categories": []}
{"id": "2509.21293", "title": "Optimal Robust Recourse with $L^p$-Bounded Model Change", "abstract": "Recourse provides individuals who received undesirable labels (e.g., denied a loan) from algorithmic decision-making systems with a minimum-cost improvement suggestion to achieve the desired outcome. However, in practice, models often get updated to reflect changes in the data distribution or environment, invalidating the recourse recommendations (i.e., following the recourse will not lead to the desirable outcome). The robust recourse literature addresses this issue by providing a framework for computing recourses whose validity is resilient to slight changes in the model. However, since the optimization problem of computing robust recourse is non-convex (even for linear models), most of the current approaches do not have any theoretical guarantee on the optimality of the recourse. Recent work by Kayastha et. al. provides the first provably optimal algorithm for robust recourse with respect to generalized linear models when the model changes are measured using the $L^{\\infty}$ norm. However, using the $L^{\\infty}$ norm can lead to recourse solutions with a high price. To address this shortcoming, we consider more constrained model changes defined by the $L^p$ norm, where $p\\geq 1$ but $p\\neq \\infty$, and provide a new algorithm that provably computes the optimal robust recourse for generalized linear models. Empirically, for both linear and non-linear models, we demonstrate that our algorithm achieves a significantly lower price of recourse (up to several orders of magnitude) compared to prior work and also exhibits a better trade-off between the implementation cost of recourse and its validity. Our empirical analysis also illustrates that our approach provides more sparse recourses compared to prior work and remains resilient to post-processing approaches that guarantee feasibility.", "date": "2025-09-25", "categories": []}
{"id": "2602.04408", "title": "Separation-Utility Pareto Frontier: An Information-Theoretic Characterization", "abstract": "We study the Pareto frontier (optimal trade-off) between utility and separation, a fairness criterion requiring predictive independence from sensitive attributes conditional on the true outcome. Through an information-theoretic lens, we prove a characterization of the utility-separation Pareto frontier, establish its concavity, and thereby prove the increasing marginal cost of separation in terms of utility. In addition, we characterize the conditions under which this trade-off becomes strict, providing a guide for trade-off selection in practice. Based on the theoretical characterization, we develop an empirical regularizer based on conditional mutual information (CMI) between predictions and sensitive attributes given the true outcome. The CMI regularizer is compatible with any deep model trained via gradient-based optimization and serves as a scalar monitor of residual separation violations, offering tractable guarantees during training. Finally, numerical experiments support our theoretical findings: across COMPAS, UCI Adult, UCI Bank, and CelebA, the proposed method substantially reduces separation violations while matching or exceeding the utility of established baseline methods. This study thus offers a provable, stable, and flexible approach to enforcing separation in deep learning.", "date": "2026-02-04", "categories": []}
{"id": "2602.05084", "title": "Individual Fairness In Strategic Classification", "abstract": "Strategic classification, where individuals modify their features to influence machine learning (ML) decisions, presents critical fairness challenges. While group fairness in this setting has been widely studied, individual fairness remains underexplored. We analyze threshold-based classifiers and prove that deterministic thresholds violate individual fairness. Then, we investigate the possibility of using a randomized classifier to achieve individual fairness. We introduce conditions under which a randomized classifier ensures individual fairness and leverage these conditions to find an optimal and individually fair randomized classifier through a linear programming problem. Additionally, we demonstrate that our approach can be extended to group fairness notions. Experiments on real-world datasets confirm that our method effectively mitigates unfairness and improves the fairness-accuracy trade-off.", "date": "2026-02-04", "categories": []}
{"id": "2602.05144", "title": "Fairness Under Group-Conditional Prior Probability Shift: Invariance, Drift, and Target-Aware Post-Processing", "abstract": "Machine learning systems are often trained and evaluated for fairness on historical data, yet deployed in environments where conditions have shifted. A particularly common form of shift occurs when the prevalence of positive outcomes changes differently across demographic groups--for example, when disease rates rise faster in one population than another, or when economic conditions affect loan default rates unequally. We study group-conditional prior probability shift (GPPS), where the label prevalence $P(Y=1\\mid A=a)$ may change between training and deployment while the feature-generation process $P(X\\mid Y,A)$ remains stable. Our analysis yields three main contributions. First, we prove a fundamental dichotomy: fairness criteria based on error rates (equalized odds) are structurally invariant under GPPS, while acceptance-rate criteria (demographic parity) can drift--and we prove this drift is unavoidable for non-trivial classifiers (shift-robust impossibility). Second, we show that target-domain risk and fairness metrics are identifiable without target labels: the invariance of ROC quantities under GPPS enables consistent estimation from source labels and unlabeled target data alone, with finite-sample guarantees. Third, we propose TAP-GPPS, a label-free post-processing algorithm that estimates prevalences from unlabeled data, corrects posteriors, and selects thresholds to satisfy demographic parity in the target domain. Experiments validate our theoretical predictions and demonstrate that TAP-GPPS achieves target fairness with minimal utility loss.", "date": "2026-02-05", "categories": []}
{"id": "2602.05162", "title": "SHaSaM: Submodular Hard Sample Mining for Fair Facial Attribute Recognition", "abstract": "Deep neural networks often inherit social and demographic biases from annotated data during model training, leading to unfair predictions, especially in the presence of sensitive attributes like race, age, gender etc. Existing methods fall prey to the inherent data imbalance between attribute groups and inadvertently emphasize on sensitive attributes, worsening unfairness and performance. To surmount these challenges, we propose SHaSaM (Submodular Hard Sample Mining), a novel combinatorial approach that models fairness-driven representation learning as a submodular hard-sample mining problem. Our two-stage approach comprises of SHaSaM-MINE, which introduces a submodular subset selection strategy to mine hard positives and negatives - effectively mitigating data imbalance, and SHaSaM-LEARN, which introduces a family of combinatorial loss functions based on Submodular Conditional Mutual Information to maximize the decision boundary between target classes while minimizing the influence of sensitive attributes. This unified formulation restricts the model from learning features tied to sensitive attributes, significantly enhancing fairness without sacrificing performance. Experiments on CelebA and UTKFace demonstrate that SHaSaM achieves state-of-the-art results, with up to 2.7 points improvement in model fairness (Equalized Odds) and a 3.5% gain in Accuracy, within fewer epochs as compared to existing methods.", "date": "2026-02-05", "categories": []}
{"id": "2602.05707", "title": "Fix Representation (Optimally) Before Fairness: Finite-Sample Shrinkage Population Correction and the True Price of Fairness Under Subpopulation Shift", "abstract": "Machine learning practitioners frequently observe tension between predictive accuracy and group fairness constraints -- yet sometimes fairness interventions appear to improve accuracy. We show that both phenomena can be artifacts of training data that misrepresents subgroup proportions. Under subpopulation shift (stable within-group distributions, shifted group proportions), we establish: (i) full importance-weighted correction is asymptotically unbiased but finite-sample suboptimal; (ii) the optimal finite-sample correction is a shrinkage reweighting that interpolates between target and training mixtures; (iii) apparent \"fairness helps accuracy\" can arise from comparing fairness methods to an improperly-weighted baseline. We provide an actionable evaluation protocol: fix representation (optimally) before fairness -- compare fairness interventions against a shrinkage-corrected baseline to isolate the true, irreducible price of fairness. Experiments on synthetic and real-world benchmarks (Adult, COMPAS) validate our theoretical predictions and demonstrate that this protocol eliminates spurious tradeoffs, revealing the genuine fairness-utility frontier.", "date": "2026-02-05", "categories": []}
{"id": "2602.05713", "title": "Projected Boosting with Fairness Constraints: Quantifying the Cost of Fair Training Distributions", "abstract": "Boosting algorithms enjoy strong theoretical guarantees: when weak learners maintain positive edge, AdaBoost achieves geometric decrease of exponential loss. We study how to incorporate group fairness constraints into boosting while preserving analyzable training dynamics. Our approach, FairBoost, projects the ensemble-induced exponential-weights distribution onto a convex set of distributions satisfying fairness constraints (as a reweighting surrogate), then trains weak learners on this fair distribution. The key theoretical insight is that projecting the training distribution reduces the effective edge of weak learners by a quantity controlled by the KL-divergence of the projection. We prove an exponential-loss bound where the convergence rate depends on weak learner edge minus a \"fairness cost\" term $\\delta_t = \\sqrt{\\mathrm{KL}(w^t \\| q^t)/2}$. This directly quantifies the accuracy-fairness tradeoff in boosting dynamics. Experiments on standard benchmarks validate the theoretical predictions and demonstrate competitive fairness-accuracy tradeoffs with stable training curves.", "date": "2026-02-05", "categories": []}
{"id": "2508.17944", "title": "Toward Substantive Intersectional Algorithmic Fairness: Desiderata for a Feminist Approach", "abstract": "People's experiences of discrimination are often shaped by multiple intersecting factors, yet algorithmic fairness research rarely reflects this complexity. While intersectionality offers tools for understanding how forms of oppression interact, current approaches to intersectional algorithmic fairness tend to focus on narrowly defined demographic subgroups. These methods contribute important insights but risk oversimplifying social reality and neglecting structural inequalities. In this paper, we outline how a substantive approach to intersectional algorithmic fairness can reorient this research and practice. In particular, we propose Substantive Intersectional Algorithmic Fairness, extending Green's (2022) notion of substantive algorithmic fairness with insights from intersectional feminist theory. Aiming to provide as actionable guidance as possible, our approach is articulated as ten desiderata to guide the design, assessment, and deployment of algorithmic systems that address systemic inequities while mitigating harms to intersectionally marginalized communities. Rather than prescribing fixed operationalizations, these desiderata invite AI practitioners and experts to reflect on assumptions of neutrality, the use of protected attributes, the inclusion of multiply marginalized groups, and the transformative potential of algorithmic systems. By bridging computational and social science perspectives, the approach emphasizes that fairness cannot be separated from social context, and that in some cases, principled non-deployment may be necessary.", "date": "2025-08-25", "categories": []}
{"id": "2602.03895", "title": "Benchmarking Bias Mitigation Toward Fairness Without Harm from Vision to LVLMs", "abstract": "Machine learning models trained on real-world data often inherit and amplify biases against certain social groups, raising urgent concerns about their deployment at scale. While numerous bias mitigation methods have been proposed, comparing the effectiveness of bias mitigation methods remains difficult due to heterogeneous datasets, inconsistent fairness metrics, isolated evaluation of vision versus multi-modal models, and insufficient hyperparameter tuning that undermines fair comparisons. We introduce NH-Fair, a unified benchmark for fairness without harm that spans both vision models and large vision-language models (LVLMs) under standardized data, metrics, and training protocols, covering supervised and zero-shot regimes. Our key contributions are: (1) a systematic ERM tuning study that identifies training choices with large influence on both utility and disparities, yielding empirically grounded guidelines to help practitioners reduce expensive hyperparameter tuning space in achieving strong fairness and accuracy; (2) evidence that many debiasing methods do not reliably outperform a well-tuned ERM baseline, whereas a composite data-augmentation method consistently delivers parity gains without sacrificing utility, emerging as a promising practical strategy. (3) an analysis showing that while LVLMs achieve higher average accuracy, they still exhibit subgroup disparities, and gains from scaling are typically smaller than those from architectural or training-protocol choices. NH-Fair provides a reproducible, tuning-aware pipeline for rigorous, harm-aware fairness evaluation.", "date": "2026-02-03", "categories": []}
{"id": "2602.04125", "title": "Attack-Resistant Uniform Fairness for Linear and Smooth Contextual Bandits", "abstract": "Modern systems, such as digital platforms and service systems, increasingly rely on contextual bandits for online decision-making; however, their deployment can inadvertently create unfair exposure among arms, undermining long-term platform sustainability and supplier trust. This paper studies the contextual bandit problem under a uniform $(1-\\delta)$-fairness constraint, and addresses its unique vulnerabilities to strategic manipulation. The fairness constraint ensures that preferential treatment is strictly justified by an arm's actual reward across all contexts and time horizons, using uniformity to prevent statistical loopholes. We develop novel algorithms that achieve (nearly) minimax-optimal regret for both linear and smooth reward functions, while maintaining strong $(1-\\tilde{O}(1/T))$-fairness guarantees, and further characterize the theoretically inherent yet asymptotically marginal \"price of fairness\". However, we reveal that such merit-based fairness becomes uniquely susceptible to signal manipulation. We show that an adversary with a minimal $\\tilde{O}(1)$ budget can not only degrade overall performance as in traditional attacks, but also selectively induce insidious fairness-specific failures while leaving conspicuous regret measures largely unaffected. To counter this, we design robust variants incorporating corruption-adaptive exploration and error-compensated thresholding. Our approach yields the first minimax-optimal regret bounds under $C$-budgeted attack while preserving $(1-\\tilde{O}(1/T))$-fairness. Numerical experiments and a real-world case demonstrate that our algorithms sustain both fairness and efficiency.", "date": "2026-02-04", "categories": []}
{"id": "2602.04339", "title": "RISE: Interactive Visual Diagnosis of Fairness in Machine Learning Models", "abstract": "Evaluating fairness under domain shift is challenging because scalar metrics often obscure exactly where and how disparities arise. We introduce \\textit{RISE} (Residual Inspection through Sorted Evaluation), an interactive visualization tool that converts sorted residuals into interpretable patterns. By connecting residual curve structures to formal fairness notions, RISE enables localized disparity diagnosis, subgroup comparison across environments, and the detection of hidden fairness issues. Through post-hoc analysis, RISE exposes accuracy-fairness trade-offs that aggregate statistics miss, supporting more informed model selection.", "date": "2026-02-04", "categories": []}
{"id": "2602.04398", "title": "Bi-directional Bias Attribution: Debiasing Large Language Models without Modifying Prompts", "abstract": "Large language models (LLMs) have demonstrated impressive capabilities across a wide range of natural language processing tasks. However, their outputs often exhibit social biases, raising fairness concerns. Existing debiasing methods, such as fine-tuning on additional datasets or prompt engineering, face scalability issues or compromise user experience in multi-turn interactions. To address these challenges, we propose a framework for detecting stereotype-inducing words and attributing neuron-level bias in LLMs, without the need for fine-tuning or prompt modification. Our framework first identifies stereotype-inducing adjectives and nouns via comparative analysis across demographic groups. We then attribute biased behavior to specific neurons using two attribution strategies based on integrated gradients. Finally, we mitigate bias by directly intervening on their activations at the projection layer. Experiments on three widely used LLMs demonstrate that our method effectively reduces bias while preserving overall model performance. Code is available at the github link: https://github.com/XMUDeepLIT/Bi-directional-Bias-Attribution.", "date": "2026-02-04", "categories": []}
{"id": "2503.17089", "title": "Understanding-informed Bias Mitigation for Fair CMR Segmentation", "abstract": "Artificial intelligence (AI) is increasingly being used for medical imaging tasks. However, there can be biases in AI models, particularly when they are trained using imbalanced training datasets. One such example has been the strong ethnicity bias effect in cardiac magnetic resonance (CMR) image segmentation models. Although this phenomenon has been reported in a number of publications, little is known about the effectiveness of bias mitigation algorithms in this domain. We aim to investigate the impact of common bias mitigation methods to address bias between Black and White subjects in AI-based CMR segmentation models. Specifically, we use oversampling, importance reweighing and Group DRO as well as combinations of these techniques to mitigate the ethnicity bias. Second, motivated by recent findings on the root causes of AI-based CMR segmentation bias, we evaluate the same methods using models trained and evaluated on cropped CMR images. We find that bias can be mitigated using oversampling, significantly improving performance for the underrepresented Black subjects whilst not significantly reducing the majority White subjects' performance. Using cropped images increases performance for both ethnicities and reduces the bias, whilst adding oversampling as a bias mitigation technique with cropped images reduces the bias further. When testing the models on an external clinical validation set, we find high segmentation performance and no statistically significant bias.", "date": "2025-03-21", "categories": []}
{"id": "2510.12036", "title": "On the Interplay between Human Label Variation and Model Fairness", "abstract": "The impact of human label variation (HLV) on model fairness is an unexplored topic. This paper examines the interplay by comparing training on majority-vote labels with a range of HLV methods. Our experiments show that without explicit debiasing, HLV training methods have a positive impact on fairness under certain configurations.", "date": "2025-10-14", "categories": []}
{"id": "2512.01834", "title": "Mitigating Gender Bias in Depression Detection via Counterfactual Inference", "abstract": "Audio-based depression detection models have demonstrated promising performance but often suffer from gender bias due to imbalanced training data. Epidemiological statistics show a higher prevalence of depression in females, leading models to learn spurious correlations between gender and depression. Consequently, models tend to over-diagnose female patients while underperforming on male patients, raising significant fairness concerns. To address this, we propose a novel Counterfactual Debiasing Framework grounded in causal inference. We construct a causal graph to model the decision-making process and identify gender bias as the direct causal effect of gender on the prediction. During inference, we employ counterfactual inference to estimate and subtract this direct effect, ensuring the model relies primarily on authentic acoustic pathological features. Extensive experiments on the DAIC-WOZ dataset using two advanced acoustic backbones demonstrate that our framework not only significantly reduces gender bias but also improves overall detection performance compared to existing debiasing strategies.", "date": "2025-12-01", "categories": []}
{"id": "2602.00032", "title": "Happy Young Women, Grumpy Old Men? Emotion-Driven Demographic Biases in Synthetic Face Generation", "abstract": "Synthetic face generation has rapidly advanced with the emergence of text-to-image (T2I) and of multimodal large language models, enabling high-fidelity image production from natural-language prompts. Despite the widespread adoption of these tools, the biases, representational quality, and cross-cultural consistency of these models remain poorly understood. Prior research on biases in the synthetic generation of human faces has examined demographic biases, yet there is little research on how emotional prompts influence demographic representation and how models trained in different cultural and linguistic contexts vary in their output distributions. We present a systematic audit of eight state-of-the-art T2I models comprising four models developed by Western organizations and four developed by Chinese institutions, all prompted identically. Using state-of-the-art facial analysis algorithms, we estimate the gender, race, age, and attractiveness levels in the generated faces. To measure the deviations from global population statistics, we apply information-theoretic bias metrics including Kullback-Leibler and Jensen-Shannon divergences. Our findings reveal persistent demographic and emotion-conditioned biases in all models regardless of their country of origin. We discuss implications for fairness, socio-technical harms, governance, and the development of transparent generative systems.", "date": "2026-01-18", "categories": []}
{"id": "2602.02512", "title": "Efficient Edge Rewiring Strategies for Enhancing PageRank Fairness", "abstract": "We study the notion of unfairness in social networks, where a group such as females in a male-dominated industry are disadvantaged in access to important information, e.g. job posts, due to their less favorable positions in the network. We investigate a well-established network-based formulation of fairness called PageRank fairness, which refers to a fair allocation of the PageRank weights among distinct groups. Our goal is to enhance the PageRank fairness by modifying the underlying network structure. More precisely, we study the problem of maximizing PageRank fairness with respect to a disadvantaged group, when we are permitted to rewire a fixed number of edges in the network. Building on a greedy approach, we leverage techniques from fast sampling of rooted spanning forests to devise an effective linear-time algorithm for this problem. To evaluate the accuracy and performance of our proposed algorithm, we conduct a large set of experiments on various real-world network data. Our experiments demonstrate that the proposed algorithm significantly outperforms the existing ones. Our algorithm is capable of generating accurate solutions for networks of million nodes in just a few minutes.", "date": "2026-01-23", "categories": []}
{"id": "2602.02516", "title": "Measuring Individual User Fairness with User Similarity and Effectiveness Disparity", "abstract": "Individual user fairness is commonly understood as treating similar users similarly. In Recommender Systems (RSs), several evaluation measures exist for quantifying individual user fairness. These measures evaluate fairness via either: (i) the disparity in RS effectiveness scores regardless of user similarity, or (ii) the disparity in items recommended to similar users regardless of item relevance. Both disparity in recommendation effectiveness and user similarity are very important in fairness, yet no existing individual user fairness measure simultaneously accounts for both. In brief, current user fairness evaluation measures implement a largely incomplete definition of fairness. To fill this gap, we present Pairwise User unFairness (PUF), a novel evaluation measure of individual user fairness that considers both effectiveness disparity and user similarity. PUF is the only measure that can express this important distinction. We empirically validate that PUF does this consistently across 4 datasets and 7 rankers, and robustly when varying user similarity or effectiveness. In contrast, all other measures are either almost insensitive to effectiveness disparity or completely insensitive to user similarity. We contribute the first RS evaluation measure to reliably capture both user similarity and effectiveness in individual user fairness. Our code: https://github.com/theresiavr/PUF-individual-user-fairness-recsys.", "date": "2026-01-23", "categories": []}
{"id": "2602.02566", "title": "A Comparative Simulation Study of the Fairness and Accuracy of Predictive Policing Systems in Baltimore City", "abstract": "There are ongoing discussions about predictive policing systems, such as those deployed in Los Angeles, California and Baltimore, Maryland, being unfair, for example, by exhibiting racial bias. Studies found that unfairness may be due to feedback loops and being trained on historically biased recorded data. However, comparative studies on predictive policing systems are few and are not sufficiently comprehensive. In this work, we perform a comprehensive comparative simulation study on the fairness and accuracy of predictive policing technologies in Baltimore. Our results suggest that the situation around bias in predictive policing is more complex than was previously assumed. While predictive policing exhibited bias due to feedback loops as was previously reported, we found that the traditional alternative, hot spots policing, had similar issues. Predictive policing was found to be more fair and accurate than hot spots policing in the short term, although it amplified bias faster, suggesting the potential for worse long-run behavior. In Baltimore, in some cases the bias in these systems tended toward over-policing in White neighborhoods, unlike in previous studies. Overall, this work demonstrates a methodology for city-specific evaluation and behavioral-tendency comparison of predictive policing systems, showing how such simulations can reveal inequities and long-term tendencies.", "date": "2026-01-30", "categories": []}
{"id": "2602.02582", "title": "Uncertainty and Fairness Awareness in LLM-Based Recommendation Systems", "abstract": "Large language models (LLMs) enable powerful zero-shot recommendations by leveraging broad contextual knowledge, yet predictive uncertainty and embedded biases threaten reliability and fairness. This paper studies how uncertainty and fairness evaluations affect the accuracy, consistency, and trustworthiness of LLM-generated recommendations. We introduce a benchmark of curated metrics and a dataset annotated for eight demographic attributes (31 categorical values) across two domains: movies and music. Through in-depth case studies, we quantify predictive uncertainty (via entropy) and demonstrate that Google DeepMind's Gemini 1.5 Flash exhibits systematic unfairness for certain sensitive attributes; measured similarity-based gaps are SNSR at 0.1363 and SNSV at 0.0507. These disparities persist under prompt perturbations such as typographical errors and multilingual inputs. We further integrate personality-aware fairness into the RecLLM evaluation pipeline to reveal personality-linked bias patterns and expose trade-offs between personalization and group fairness. We propose a novel uncertainty-aware evaluation methodology for RecLLMs, present empirical insights from deep uncertainty case studies, and introduce a personality profile-informed fairness benchmark that advances explainability and equity in LLM recommendations. Together, these contributions establish a foundation for safer, more interpretable RecLLMs and motivate future work on multi-model benchmarks and adaptive calibration for trustworthy deployment.", "date": "2026-01-31", "categories": []}
{"id": "2504.01154", "title": "Past-Discounting is Key for Learning Markovian Fairness with Long Horizons", "abstract": "Fairness is an important consideration for dynamic resource allocation in multi-agent systems. Many existing methods treat fairness as a one-shot problem without considering temporal dynamics, which misses the nuances of accumulating inequalities over time. Recent approaches overcome this limitation by tracking allocations over time, assuming perfect recall of all past utilities. While the former neglects long-term equity, the latter introduces a critical challenge: the augmented state space required to track cumulative utilities grows unboundedly with time, hindering the scalability and convergence of learning algorithms. Motivated by behavioral insights that human fairness judgments discount distant events, we introduce a framework for temporal fairness that incorporates past-discounting into the learning problem. This approach offers a principled interpolation between instantaneous and perfect-recall fairness. Our central contribution is a past-discounted framework for memory tracking and a theoretical analysis of fairness memories, showing past-discounting guarantees a bounded, horizon-independent state space, a property that we prove perfect-recall methods lack. This result unlocks the ability to learn fair policies tractably over arbitrarily long horizons. We formalize this framework, demonstrate its necessity with experiments showing that perfect recall fails where past-discounting succeeds, and provide a clear path toward building scalable and equitable resource allocation systems.", "date": "2025-04-01", "categories": []}
{"id": "2508.03199", "title": "Beyond Content: How Grammatical Gender Shapes Visual Representation in Text-to-Image Models", "abstract": "Research on bias in Text-to-Image (T2I) models has primarily focused on demographic representation and stereotypical attributes, overlooking a fundamental question: how does grammatical gender influence visual representation across languages? We introduce a cross-linguistic benchmark examining words where grammatical gender contradicts stereotypical gender associations (e.g., ``une sentinelle'' - grammatically feminine in French but referring to the stereotypically masculine concept ``guard''). Our dataset spans five gendered languages (French, Spanish, German, Italian, Russian) and two gender-neutral control languages (English, Chinese), comprising 800 unique prompts that generated 28,800 images across three state-of-the-art T2I models. Our analysis reveals that grammatical gender dramatically influences image generation: masculine grammatical markers increase male representation to 73% on average (compared to 22% with gender-neutral English), while feminine grammatical markers increase female representation to 38% (compared to 28% in English). These effects vary systematically by language resource availability and model architecture, with high-resource languages showing stronger effects. Our findings establish that language structure itself, not just content, shapes AI-generated visual outputs, introducing a new dimension for understanding bias and fairness in multilingual, multimodal systems.", "date": "2025-08-05", "categories": []}
{"id": "2509.15206", "title": "Fair-GPTQ: Bias-Aware Quantization for Large Language Models", "abstract": "High memory demands of generative language models have drawn attention to quantization, which reduces computational cost, memory usage, and latency by mapping model weights to lower-precision integers. Approaches such as GPTQ effectively minimize input-weight product errors during quantization; however, recent empirical studies show that they can increase biased outputs and degrade performance on fairness benchmarks, and it remains unclear which specific weights cause this issue. In this work, we draw new links between quantization and model fairness by adding explicit group-fairness constraints to the quantization objective and introduce Fair-GPTQ, the first quantization method explicitly designed to reduce unfairness in large language models. The added constraints guide the learning of the rounding operation toward less-biased text generation for protected groups. Specifically, we focus on stereotype generation involving occupational bias and discriminatory language spanning gender, race, and religion. Fair-GPTQ has minimal impact on performance, preserving at least 90% of baseline accuracy on zero-shot benchmarks, reduces unfairness relative to a half-precision model, and retains the memory and speed benefits of 4-bit quantization. We also compare the performance of Fair-GPTQ with existing debiasing methods and find that it achieves performance on par with the iterative null-space projection debiasing approach on racial-stereotype benchmarks. Overall, the results validate our theoretical solution to the quantization problem with a group-bias term, highlight its applicability for reducing group bias at quantization time in generative models, and demonstrate that our approach can further be used to analyze channel- and weight-level contributions to fairness during quantization.", "date": "2025-09-18", "categories": []}
{"id": "2510.01254", "title": "Do Bias Benchmarks Generalise? Evidence from Voice-based Evaluation of Gender Bias in SpeechLLMs", "abstract": "Recent work in benchmarking bias and fairness in speech large language models (SpeechLLMs) has relied heavily on multiple-choice question answering (MCQA) formats. The model is tasked to choose between stereotypical, anti-stereotypical, or neutral/irrelevant answers given an input speech prompt and an optional text prompt. Such MCQA benchmarks implicitly assume that model performance is consistent across other MCQA tasks, voices, and other task formats such as more realistic, long-form evaluations. In this paper, we probe that assumption. We fine-tune three SpeechLLMs using LoRA adapters to induce specific MCQA behaviours: preference for stereotypical, anti-stereotypical, or neutral/uncertain answers. We then evaluate whether these behaviours generalise to another, distinct MCQA benchmark, and more critically to long-form, creative generation tasks. Our results show that performance on MCQA bias benchmarks fails to reliably predict performances across other MCQA benchmarks, and more importantly across long-form tasks. We conclude that current MCQA bias benchmarks show limited evidence of cross-task generalisation in the speech domain, and also propose an evaluation suite for measuring behaviour transferability in future models and benchmarks.", "date": "2025-09-24", "categories": []}
{"id": "2601.11651", "title": "Aesthetics as Structural Harm: Algorithmic Lookism Across Text-to-Image Generation and Classification", "abstract": "This paper examines algorithmic lookism-the systematic preferential treatment based on physical appearance-in text-to-image (T2I) generative AI and a downstream gender classification task. Through the analysis of 26,400 synthetic faces created with Stable Diffusion 2.1 and 3.5 Medium, we demonstrate how generative AI models systematically associate facial attractiveness with positive attributes and vice-versa, mirroring socially constructed biases rather than evidence-based correlations. Furthermore, we find significant gender bias in three gender classification algorithms depending on the attributes of the input faces. Our findings reveal three critical harms: (1) the systematic encoding of attractiveness-positive attribute associations in T2I models; (2) gender disparities in classification systems, where women's faces, particularly those generated with negative attributes, suffer substantially higher misclassification rates than men's; and (3) intensifying aesthetic constraints in newer models through age homogenization, gendered exposure patterns, and geographic reductionism. These convergent patterns reveal algorithmic lookism as systematic infrastructure operating across AI vision systems, compounding existing inequalities through both representation and recognition.\n  Disclaimer: This work includes visual and textual content that reflects stereotypical associations between physical appearance and socially constructed attributes, including gender, race, and traits associated with social desirability. Any such associations found in this study emerge from the biases embedded in generative AI systems-not from empirical truths or the authors' views.", "date": "2026-01-15", "categories": []}
{"id": "2602.00094", "title": "Trade-offs Between Individual and Group Fairness in Machine Learning: A Comprehensive Review", "abstract": "Algorithmic fairness has become a central concern in computational decision-making systems, where ensuring equitable outcomes is essential for both ethical and legal reasons. Two dominant notions of fairness have emerged in the literature: Group Fairness (GF), which focuses on mitigating disparities across demographic subpopulations, and Individual Fairness (IF), which emphasizes consistent treatment of similar individuals. These notions have traditionally been studied in isolation. In contrast, this survey examines methods that jointly address GF and IF, integrating both perspectives within unified frameworks and explicitly characterizing the trade-offs between them. We provide a systematic and critical review of hybrid fairness approaches, organizing existing methods according to the fairness mechanisms they employ and the algorithmic and mathematical strategies used to reconcile multiple fairness criteria. For each class of methods, we examine their theoretical foundations, optimization mechanisms, and empirical evaluation practices, and discuss their limitations. Additionally, we discuss the challenges and identify open research directions for developing principled, context-aware hybrid fairness methods. By synthesizing insights across the literature, this survey aims to serve as a comprehensive resource for researchers and practitioners seeking to design hybrid algorithms that provide reliable fairness guarantees at both the individual and group levels.", "date": "2026-01-23", "categories": []}
{"id": "2602.00405", "title": "RobustDebias: Debiasing Language Models using Distributionally Robust Optimization", "abstract": "Pretrained language models have been shown to exhibit biases and social stereotypes. Prior work on debiasing these models has largely focused on modifying embedding spaces during pretraining, which is not scalable for large models. Fine-tuning pretrained models on task-specific datasets can both degrade model performance and amplify biases present in the fine-tuning data. We address bias amplification during fine-tuning rather than costly pretraining, focusing on BERT models due to their widespread use in language understanding tasks. While Empirical Risk Minimization effectively optimizes downstream performance, it often amplifies social biases during fine-tuning. To counter this, we propose \\textit{RobustDebias}, a novel mechanism which adapts Distributionally Robust Optimization (DRO) to debias language models during fine-tuning. Our approach debiases models across multiple demographics during MLM fine-tuning and generalizes to any dataset or task. Extensive experiments on various language models show significant bias mitigation with minimal performance impact.", "date": "2026-01-30", "categories": []}
{"id": "2602.00647", "title": "CoRe-Fed: Bridging Collaborative and Representation Fairness via Federated Embedding Distillation", "abstract": "With the proliferation of distributed data sources, Federated Learning (FL) has emerged as a key approach to enable collaborative intelligence through decentralized model training while preserving data privacy. However, conventional FL algorithms often suffer from performance disparities across clients caused by heterogeneous data distributions and unequal participation, which leads to unfair outcomes. Specifically, we focus on two core fairness challenges, i.e., representation bias, arising from misaligned client representations, and collaborative bias, stemming from inequitable contribution during aggregation, both of which degrade model performance and generalizability. To mitigate these disparities, we propose CoRe-Fed, a unified optimization framework that bridges collaborative and representation fairness via embedding-level regularization and fairness-aware aggregation. Initially, an alignment-driven mechanism promotes semantic consistency between local and global embeddings to reduce representational divergence. Subsequently, a dynamic reward-penalty-based aggregation strategy adjusts each client's weight based on participation history and embedding alignment to ensure contribution-aware aggregation. Extensive experiments across diverse models and datasets demonstrate that CoRe-Fed improves both fairness and model performance over the state-of-the-art baseline algorithms.", "date": "2026-01-31", "categories": []}
{"id": "2602.01030", "title": "Bias in the Ear of the Listener: Assessing Sensitivity in Audio Language Models Across Linguistic, Demographic, and Positional Variations", "abstract": "This work presents the first systematic investigation of speech bias in multilingual MLLMs. We construct and release the BiasInEar dataset, a speech-augmented benchmark based on Global MMLU Lite, spanning English, Chinese, and Korean, balanced by gender and accent, and totaling 70.8 hours ($\\approx$4,249 minutes) of speech with 11,200 questions. Using four complementary metrics (accuracy, entropy, APES, and Fleiss' $\\kappa$), we evaluate nine representative models under linguistic (language and accent), demographic (gender), and structural (option order) perturbations. Our findings reveal that MLLMs are relatively robust to demographic factors but highly sensitive to language and option order, suggesting that speech can amplify existing structural biases. Moreover, architectural design and reasoning strategy substantially affect robustness across languages. Overall, this study establishes a unified framework for assessing fairness and robustness in speech-integrated LLMs, bridging the gap between text- and speech-based evaluation. The resources can be found at https://github.com/ntunlplab/BiasInEar.", "date": "2026-02-01", "categories": []}
{"id": "2602.01558", "title": "How Implicit Bias Accumulates and Propagates in LLM Long-term Memory", "abstract": "Long-term memory mechanisms enable Large Language Models (LLMs) to maintain continuity and personalization across extended interaction lifecycles, but they also introduce new and underexplored risks related to fairness. In this work, we study how implicit bias, defined as subtle statistical prejudice, accumulates and propagates within LLMs equipped with long-term memory. To support systematic analysis, we introduce the Decision-based Implicit Bias (DIB) Benchmark, a large-scale dataset comprising 3,776 decision-making scenarios across nine social domains, designed to quantify implicit bias in long-term decision processes. Using a realistic long-horizon simulation framework, we evaluate six state-of-the-art LLMs integrated with three representative memory architectures on DIB and demonstrate that LLMs' implicit bias does not remain static but intensifies over time and propagates across unrelated domains. We further analyze mitigation strategies and show that a static system-level prompting baseline provides limited and short-lived debiasing effects. To address this limitation, we propose Dynamic Memory Tagging (DMT), an agentic intervention that enforces fairness constraints at memory write time. Extensive experimental results show that DMT substantially reduces bias accumulation and effectively curtails cross-domain bias propagation.", "date": "2026-02-02", "categories": []}
{"id": "2409.10825", "title": "Unveiling and Mitigating Bias in Large Language Model Recommendations: A Path to Fairness", "abstract": "Large Language Model (LLM)-based recommendation systems excel in delivering comprehensive suggestions by deeply analyzing content and user behavior. However, they often inherit biases from skewed training data, favoring mainstream content while underrepresenting diverse or non-traditional options. This study explores the interplay between bias and LLM-based recommendation systems, focusing on music, song, and book recommendations across diverse demographic and cultural groups. This paper analyzes bias in LLM-based recommendation systems across multiple models (GPT, LLaMA, and Gemini), revealing its deep and pervasive impact on outcomes. Intersecting identities and contextual factors, like socioeconomic status, further amplify biases, complicating fair recommendations across diverse groups. Our findings reveal that bias in these systems is deeply ingrained, yet even simple interventions like prompt engineering can significantly reduce it. We further propose a retrieval-augmented generation strategy to mitigate bias more effectively. Numerical experiments validate these strategies, demonstrating both the pervasive nature of bias and the impact of the proposed solutions.", "date": "2024-09-17", "categories": []}
{"id": "2502.09626", "title": "Evidence for Phenotype-Driven Disparities in Freezing of Gait Detection and Approaches to Bias Mitigation", "abstract": "Freezing of gait (FOG) is a debilitating symptom of Parkinson's disease (PD) and a common cause of injurious falls. Recent advances in wearable-based human activity recognition (HAR) enable FOG detection, but bias and fairness in these models remain understudied. Bias refers to systematic errors leading to unequal outcomes, while fairness refers to consistent performance across subject groups. Biased models could systematically underserve patients with specific FOG phenotypes or demographics, potentially widening care disparities. We systematically evaluated bias and fairness of state-of-the-art HAR models for FOG detection across phenotypes and demographics using multi-site datasets. We assessed four mitigation approaches: conventional methods (threshold optimization and adversarial debiasing) and transfer learning approaches (multi-site transfer and fine-tuning large pretrained models). Fairness was quantified using demographic parity ratio (DPR) and equalized odds ratio (EOR). HAR models exhibited substantial bias (DPR & EOR < 0.8) across age, sex, disease duration, and critically, FOG phenotype. Phenotype-specific bias is particularly concerning as tremulous and akinetic FOG require different clinical management. Conventional bias mitigation methods failed: threshold optimization (DPR=-0.126, EOR=+0.063) and adversarial debiasing (DPR=-0.008, EOR=-0.001) showed minimal improvement. In contrast, transfer learning from multi-site datasets significantly improved fairness (DPR=+0.037, p<0.01; EOR=+0.045, p<0.01) and performance (F1-score=+0.020, p<0.05). Transfer learning across diverse datasets is essential for developing equitable HAR models that reliably detect FOG across all patient phenotypes, ensuring wearable-based monitoring benefits all individuals with PD.", "date": "2025-01-29", "categories": []}
{"id": "2509.12396", "title": "Information Loss and Disparate Effects in Network Embeddings", "abstract": "An extensive line of work studies fairness interventions for network embeddings, but less is known about their baseline behavior. In this work, we ask: how do baseline embeddings (without fairness interventions) produce disparate effects at the representation level? We analyze the asymptotic behavior of low-dimensional embeddings on stochastic block model (SBM) graphs, which encode both homophily and group structure. We characterize exact conditions under which embeddings cause information loss, showing that the amount of information loss depends directly on the graph's density and assortativity. Notably, very different graphs can produce identical embeddings in the limit, and this non-invertibility disproportionately affects smaller and sparser communities. As a result, simple downstream tasks, such as link prediction, introduce higher error rates for these communities, helping explain disparities widely observed in practice.", "date": "2025-09-15", "categories": []}
{"id": "2511.16377", "title": "Optimal Fairness under Local Differential Privacy", "abstract": "We investigate how to optimally design local differential privacy (LDP) mechanisms that reduce data unfairness and thereby improve fairness in downstream classification. We first derive a closed-form optimal mechanism for binary sensitive attributes and then develop a tractable optimization framework that yields the corresponding optimal mechanism for multi-valued attributes. As a theoretical contribution, we establish that for discrimination-accuracy optimal classifiers, reducing data unfairness necessarily leads to lower classification unfairness, thus providing a direct link between privacy-aware pre-processing and classification fairness. Empirically, we demonstrate that our approach consistently outperforms existing LDP mechanisms in reducing data unfairness across diverse datasets and fairness metrics, while maintaining accuracy close to that of non-private models. Moreover, compared with leading pre-processing and post-processing fairness methods, our mechanism achieves a more favorable accuracy-fairness trade-off while simultaneously preserving the privacy of sensitive attributes. Taken together, these results highlight LDP as a principled and effective pre-processing fairness intervention technique.", "date": "2025-11-20", "categories": []}
{"id": "2311.01591", "title": "Fair Graph Machine Learning under Adversarial Missingness Processes", "abstract": "Graph Neural Networks (GNNs) have achieved state-of-the-art results in many relevant tasks where decisions might disproportionately impact specific communities. However, existing work on fair GNNs often assumes that either sensitive attributes are fully observed or they are missing completely at random. We show that an adversarial missingness process can inadvertently disguise a fair model through the imputation, leading the model to overestimate the fairness of its predictions. We address this challenge by proposing Better Fair than Sorry (BFtS), a fair missing data imputation model for sensitive attributes. The key principle behind BFtS is that imputations should approximate the worst-case scenario for fairness -- i.e. when optimizing fairness is the hardest. We implement this idea using a 3-player adversarial scheme where two adversaries collaborate against a GNN classifier, and the classifier minimizes the maximum bias. Experiments using synthetic and real datasets show that BFtS often achieves a better fairness x accuracy trade-off than existing alternatives under an adversarial missingness process.", "date": "2023-11-02", "categories": []}
{"id": "2601.21132", "title": "Large Language Models Naively Recover Ethnicity from Individual Records", "abstract": "I demonstrate that large language models can infer ethnicity from names with accuracy exceeding that of Bayesian Improved Surname Geocoding (BISG) without additional training data, enabling inference outside the United States and to contextually appropriate classification categories. Using stratified samples from Florida and North Carolina voter files with self-reported race, LLM-based classification achieves up to 84.7% accuracy, outperforming BISG (68.2%) on balanced samples. I test six models including Gemini 3 Flash, GPT-4o, and open-source alternatives such as DeepSeek v3.2 and GLM-4.7. Enabling extended reasoning can improve accuracy by 1-3 percentage points, though effects vary across contexts; including metadata such as party registration reaches 86.7%. LLM classification also reduces the income bias inherent in BISG, where minorities in wealthier neighborhoods are systematically misclassified as White. I further validate using Lebanese voter registration with religious sect (64.3% accuracy), Indian MPs from reserved constituencies (99.2%), and Indian land records with caste classification (74.0%). Aggregate validation across India, Uganda, Nepal, Armenia, Chile, and Costa Rica using original full-count voter rolls demonstrates that the method recovers known population distributions where naming conventions are distinctive. For large-scale applications, small transformer models fine-tuned on LLM labels exceed BISG accuracy while enabling local deployment at no cost.", "date": "2026-01-29", "categories": []}
{"id": "2601.22093", "title": "Investigating Associational Biases in Inter-Model Communication of Large Generative Models", "abstract": "Social bias in generative AI can manifest not only as performance disparities but also as associational bias, whereby models learn and reproduce stereotypical associations between concepts and demographic groups, even in the absence of explicit demographic information (e.g., associating doctors with men). These associations can persist, propagate, and potentially amplify across repeated exchanges in inter-model communication pipelines, where one generative model's output becomes another's input. This is especially salient for human-centred perception tasks, such as human activity recognition and affect prediction, where inferences about behaviour and internal states can lead to errors or stereotypical associations that propagate into unequal treatment. In this work, focusing on human activity and affective expression, we study how such associations evolve within an inter-model communication pipeline that alternates between image generation and image description. Using the RAF-DB and PHASE datasets, we quantify demographic distribution drift induced by model-to-model information exchange and assess whether these drifts are systematic using an explainability pipeline. Our results reveal demographic drifts toward younger representations for both actions and emotions, as well as toward more female-presenting representations, primarily for emotions. We further find evidence that some predictions are supported by spurious visual regions (e.g., background or hair) rather than concept-relevant cues (e.g., body or face). We also examine whether these demographic drifts translate into measurable differences in downstream behaviour, i.e., while predicting activity and emotion labels. Finally, we outline mitigation strategies spanning data-centric, training and deployment interventions, and emphasise the need for careful safeguards when deploying interconnected models in human-centred AI systems.", "date": "2026-01-29", "categories": []}
{"id": "2507.01063", "title": "FAIR-MATCH: A Multi-Objective Framework for Bias Mitigation in Reciprocal Dating Recommendations", "abstract": "Online dating platforms have fundamentally transformed the formation of romantic relationships, with millions of users worldwide relying on algorithmic matching systems to find compatible partners. However, current recommendation systems in dating applications suffer from significant algorithmic deficiencies, including but not limited to popularity bias, filter bubble effects, and inadequate reciprocity modeling that limit effectiveness and introduce harmful biases. This research integrates foundational work with recent empirical findings to deliver a detailed analysis of dating app recommendation systems, highlighting key issues and suggesting research-backed solutions. Through analysis of reciprocal recommendation frameworks, fairness evaluation metrics, and industry implementations, we demonstrate that current systems achieve modest performance with collaborative filtering reaching 25.1\\% while reciprocal methods achieve 28.7\\%. Our proposed mathematical framework addresses these limitations through enhanced similarity measures, multi-objective optimization, and fairness-aware algorithms that maintain competitive accuracy while improving demographic representation to reduce algorithmic bias.", "date": "2025-06-30", "categories": []}
{"id": "2601.20217", "title": "An Accounting Identity for Algorithmic Fairness", "abstract": "We derive an accounting identity for predictive models that links accuracy with common fairness criteria. The identity shows that for globally calibrated models, the weighted sums of miscalibration within groups and error imbalance across groups is equal to a \"total unfairness budget.\" For binary outcomes, this budget is the model's mean-squared error times the difference in group prevalence across outcome classes. The identity nests standard impossibility results as special cases, while also describing inherent tradeoffs when one or more fairness measures are not perfectly satisfied. The results suggest that accuracy and fairness are best viewed as complements in binary prediction tasks: increasing accuracy necessarily shrinks the total unfairness budget and vice-versa. Experiments on benchmark data confirm the theory and show that many fairness interventions largely substitute between fairness violations, and when they reduce accuracy they tend to expand the total unfairness budget. The results extend naturally to prediction tasks with non-binary outcomes, illustrating how additional outcome information can relax fairness incompatibilities and identifying conditions under which the binary-style impossibility does and does not extend to regression tasks.", "date": "2026-01-28", "categories": []}
{"id": "2601.20848", "title": "Post-Training Fairness Control: A Single-Train Framework for Dynamic Fairness in Recommendation", "abstract": "Despite growing efforts to mitigate unfairness in recommender systems, existing fairness-aware methods typically fix the fairness requirement at training time and provide limited post-training flexibility. However, in real-world scenarios, diverse stakeholders may demand differing fairness requirements over time, so retraining for different fairness requirements becomes prohibitive. To address this limitation, we propose Cofair, a single-train framework that enables post-training fairness control in recommendation. Specifically, Cofair introduces a shared representation layer with fairness-conditioned adapter modules to produce user embeddings specialized for varied fairness levels, along with a user-level regularization term that guarantees user-wise monotonic fairness improvements across these levels. We theoretically establish that the adversarial objective of Cofair upper bounds demographic parity and the regularization term enforces progressive fairness at user level. Comprehensive experiments on multiple datasets and backbone models demonstrate that our framework provides dynamic fairness at different levels, delivering comparable or better fairness-accuracy curves than state-of-the-art baselines, without the need to retrain for each new fairness requirement. Our code is publicly available at https://github.com/weixinchen98/Cofair.", "date": "2026-01-28", "categories": []}
{"id": "2506.14988", "title": "Fair Algorithms with Probing for Multi-Agent Multi-Armed Bandits", "abstract": "We propose a multi-agent multi-armed bandit (MA-MAB) framework aimed at ensuring fair outcomes across agents while maximizing overall system performance. A key challenge in this setting is decision-making under limited information about arm rewards. To address this, we introduce a novel probing framework that strategically gathers information about selected arms before allocation. In the offline setting, where reward distributions are known, we leverage submodular properties to design a greedy probing algorithm with a provable performance bound. For the more complex online setting, we develop an algorithm that achieves sublinear regret while maintaining fairness. Extensive experiments on synthetic and real-world datasets show that our approach outperforms baseline methods, achieving better fairness and efficiency.", "date": "2025-06-17", "categories": []}
{"id": "2509.25795", "title": "Assessing Algorithmic Bias in Language-Based Depression Detection: A Comparison of DNN and LLM Approaches", "abstract": "This paper investigates algorithmic bias in language-based models for automated depression detection, focusing on socio-demographic disparities related to gender and race/ethnicity. Models trained using deep neural networks (DNN) based embeddings are compared to few-shot learning approaches with large language models (LLMs), evaluating both performance and fairness on clinical interview transcripts from the Distress Analysis Interview Corpus/Wizard-of-Oz (DAIC-WOZ). To mitigate bias, fairness-aware loss functions are applied to DNN-based models, while in-context learning with varied prompt framing and shot counts is explored for LLMs. Results indicate that LLMs outperform DNN-based models in depression classification, particularly for underrepresented groups such as Hispanic participants. LLMs also exhibit reduced gender bias compared to DNN-based embeddings, though racial disparities persist. Among fairness-aware techniques for mitigating bias in DNN-based embeddings, the worst-group loss, which is designed to minimize loss for the worst-performing demographic group, achieves a better balance between performance and fairness. In contrast, the fairness-regularized loss minimizes loss across all groups but performs less effectively. In LLMs, guided prompting with ethical framing helps mitigate gender bias in the 1-shot setting. However, increasing the number of shots does not lead to further reductions in disparities. For race/ethnicity, neither prompting strategy nor increasing $N$ in $N$-shot learning effectively reduces disparities.", "date": "2025-09-30", "categories": []}
{"id": "2510.21091", "title": "Doubly-Regressing Approach for Subgroup Fairness", "abstract": "Algorithmic fairness is a socially crucial topic in real-world applications of AI.\n  Among many notions of fairness, subgroup fairness is widely studied when multiple sensitive attributes (e.g., gender, race, age) are present.\n  However, as the number of sensitive attributes grows, the number of subgroups increases accordingly, creating heavy computational burdens and data sparsity problem (subgroups with too small sizes).\n  In this paper, we develop a novel learning algorithm for subgroup fairness which resolves these issues by focusing on subgroups with sufficient sample sizes as well as marginal fairness (fairness for each sensitive attribute).\n  To this end, we formalize a notion of subgroup-subset fairness and introduce a corresponding distributional fairness measure called the supremum Integral Probability Metric (supIPM).\n  Building on this formulation, we propose the Doubly Regressing Adversarial learning for subgroup Fairness (DRAF) algorithm, which reduces a surrogate fairness gap for supIPM with much less computation than directly reducing supIPM.\n  Theoretically, we prove that the proposed surrogate fairness gap is an upper bound of supIPM.\n  Empirically, we show that the DRAF algorithm outperforms baseline methods in benchmark datasets, specifically when the number of sensitive attributes is large so that many subgroups are very small.", "date": "2025-10-24", "categories": []}
{"id": "2601.19035", "title": "Unravelling the (In)compatibility of Statistical-Parity and Equalized-Odds", "abstract": "A key challenge in employing data, algorithms and data-driven systems is to adhere to the principle of fairness and justice. Statistical fairness measures belong to an important category of technical/formal mechanisms for detecting fairness issues in data and algorithms. In this contribution we study the relations between two types of statistical fairness measures namely Statistical-Parity and Equalized-Odds. The Statistical-Parity measure does not rely on having ground truth, i.e., (objectively) labeled target attributes. This makes Statistical-Parity a suitable measure in practice for assessing fairness in data and data classification algorithms. Therefore, Statistical-Parity is adopted in many legal and professional frameworks for assessing algorithmic fairness. The Equalized-Odds measure, on the contrary, relies on having (reliable) ground-truth, which is not always feasible in practice. Nevertheless, there are several situations where the Equalized-Odds definition should be satisfied to enforce false prediction parity among sensitive social groups. We present a novel analyze of the relation between Statistical-Parity and Equalized-Odds, depending on the base-rates of sensitive groups. The analysis intuitively shows how and when base-rate imbalance causes incompatibility between Statistical-Parity and Equalized-Odds measures. As such, our approach provides insight in (how to make design) trade-offs between these measures in practice. Further, based on our results, we plea for examining base-rate (im)balance and investigating the possibility of such an incompatibility before enforcing or relying on the Statistical-Parity criterion. The insights provided, we foresee, may trigger initiatives to improve or adjust the current practice and/or the existing legal frameworks.", "date": "2026-01-26", "categories": []}
{"id": "2601.19595", "title": "Intersectional Fairness via Mixed-Integer Optimization", "abstract": "The deployment of Artificial Intelligence in high-risk domains, such as finance and healthcare, necessitates models that are both fair and transparent. While regulatory frameworks, including the EU's AI Act, mandate bias mitigation, they are deliberately vague about the definition of bias. In line with existing research, we argue that true fairness requires addressing bias at the intersections of protected groups. We propose a unified framework that leverages Mixed-Integer Optimization (MIO) to train intersectionally fair and intrinsically interpretable classifiers. We prove the equivalence of two measures of intersectional fairness (MSD and SPSF) in detecting the most unfair subgroup and empirically demonstrate that our MIO-based algorithm improves performance in finding bias. We train high-performing, interpretable classifiers that bound intersectional bias below an acceptable threshold, offering a robust solution for regulated industries and beyond.", "date": "2026-01-27", "categories": []}
{"id": "2407.03473", "title": "Exploring LGBTQ+ Bias in Generative AI Answers across Different Country and Religious Contexts", "abstract": "Previous discussions have highlighted the need for generative AI tools to become more culturally sensitive, yet often neglect the complexities of handling content about minorities, who are perceived differently across cultures and religions. Our study examined how two generative AI systems respond to homophobic statements with varying cultural and religious context information. Findings showed ChatGPT 3.5's replies exhibited cultural relativism, in contrast to Bard's, which stressed human rights and provided more support for LGBTQ+ issues. Both demonstrated significant change in responses based on contextual information provided in the prompts, suggesting that AI systems may adjust in their responses the degree and forms of support for LGBTQ+ people according to information they receive about the user's background. The study contributes to understanding the social and ethical implications of AI responses and argues that any work to make generative AI outputs more culturally diverse requires a grounding in fundamental human rights. A revised edition of this preprint is available open access at Big Data & Society at https://doi.org/10.1177/20539517251396069", "date": "2024-07-03", "categories": []}
{"id": "2506.12371", "title": "Path-specific effects for pulse-oximetry guided decisions in critical care", "abstract": "Identifying and measuring biases associated with sensitive attributes is a crucial consideration in healthcare to prevent treatment disparities. One prominent issue is inaccurate pulse oximeter readings, which tend to overestimate oxygen saturation for dark-skinned patients and misrepresent supplemental oxygen needs. Most existing research has revealed statistical disparities linking device measurement errors to patient outcomes in intensive care units (ICUs) without causal formalization. This study causally investigates how racial discrepancies in oximetry measurements affect invasive ventilation in ICU settings. We employ a causal inference-based approach using path-specific effects to isolate the impact of bias by race on clinical decision-making. To estimate these effects, we leverage a doubly robust estimator, propose its self-normalized variant for improved sample efficiency, and provide novel finite-sample guarantees. Our methodology is validated on semi-synthetic data and applied to two large real-world health datasets: MIMIC-IV and eICU. Contrary to prior work, our analysis reveals minimal impact of racial discrepancies on invasive ventilation rates. However, path-specific effects mediated by oxygen saturation disparity are more pronounced on ventilation duration, and the severity differs across datasets. Our work provides a novel pipeline for investigating potential disparities in clinical decision-making and, more importantly, highlights the necessity of causal methods to robustly assess fairness in healthcare.", "date": "2025-06-14", "categories": []}
{"id": "2508.18989", "title": "Ask Me Again Differently: GRAS for Measuring Bias in Vision Language Models on Gender, Race, Age, and Skin Tone", "abstract": "As Vision Language Models (VLMs) become integral to real-world applications, understanding their demographic biases is critical. We introduce GRAS, a benchmark for uncovering demographic biases in VLMs across gender, race, age, and skin tone, offering the most diverse coverage to date. We further propose the GRAS Bias Score, an interpretable metric for quantifying bias. We benchmark five state-of-the-art VLMs and reveal concerning bias levels, with the least biased model attaining a GRAS Bias Score of only 2 out of 100. Our findings also reveal a methodological insight: evaluating bias in VLMs with visual question answering (VQA) requires considering multiple formulations of a question. Our code, data, and evaluation results are publicly available.", "date": "2025-08-26", "categories": []}
{"id": "2601.16989", "title": "The Voice of Equity: A Systematic Evaluation of Bias Mitigation Techniques for Speech-Based Cognitive Impairment Detection Across Architectures and Demographics", "abstract": "Speech-based detection of cognitive impairment offers a scalable, non-invasive screening, yet algorithmic bias across demographic and linguistic subgroups remains critically underexplored. We present the first comprehensive fairness analysis framework for speech-based multi-class cognitive impairment detection, systematically evaluating bias mitigation across architectures, and demographic subgroups. We developed two transformer-based architectures, SpeechCARE-AGF and Whisper-LWF-LoRA, on the multilingual NIA PREPARE Challenge dataset. Unlike prior work that typically examines single mitigation techniques, we compared pre-processing, in-processing, and post-processing approaches, assessing fairness via Equality of Opportunity and Equalized Odds across gender, age, education, and language. Both models achieved strong performance (F1: SpeechCARE-AGF 70.87, Whisper-LWF-LoRA 71.46) but exhibited substantial fairness disparities. Adults >=80 showed lower sensitivity versus younger groups; Spanish speakers demonstrated reduced TPR versus English speakers. Mitigation effectiveness varied by architecture: oversampling improved SpeechCARE-AGF for older adults (80+ TPR: 46.19%=>49.97%) but minimally affected Whisper-LWF-LoRA. This study addresses a critical healthcare AI gap by demonstrating that architectural design fundamentally shapes bias patterns and mitigation effectiveness. Adaptive fusion mechanisms enable flexible responses to data interventions, while frequency reweighting offers robust improvements across architectures. Our findings establish that fairness interventions must be tailored to both model architecture and demographic characteristics, providing a systematic framework for developing equitable speech-based screening tools essential for reducing diagnostic disparities in cognitive healthcare.", "date": "2026-01-07", "categories": []}
{"id": "2601.17146", "title": "Falsifying Predictive Algorithm", "abstract": "Empirical investigations into unintended model behavior often show that the algorithm is predicting another outcome than what was intended. These exposes highlight the need to identify when algorithms predict unintended quantities - ideally before deploying them into consequential settings. We propose a falsification framework that provides a principled statistical test for discriminant validity: the requirement that an algorithm predict intended outcomes better than impermissible ones. Drawing on falsification practices from causal inference, econometrics, and psychometrics, our framework compares calibrated prediction losses across outcomes to assess whether the algorithm exhibits discriminant validity with respect to a specified impermissible proxy. In settings where the target outcome is difficult to observe, multiple permissible proxy outcomes may be available; our framework accommodates both this setting and the case with a single permissible proxy. Throughout we use nonparametric hypothesis testing methods that make minimal assumptions on the data-generating process. We illustrate the method in an admissions setting, where the framework establishes discriminant validity with respect to gender but fails to establish discriminant validity with respect to race. This demonstrates how falsification can serve as an early validity check, prior to fairness or robustness analyses. We also provide analysis in a criminal justice setting, where we highlight the limitations of our framework and emphasize the need for complementary approaches to assess other aspects of construct validity and external validity.", "date": "2026-01-23", "categories": []}
{"id": "2601.17944", "title": "Credit Fairness: Online Fairness In Shared Resource Pools", "abstract": "We consider a setting in which a group of agents share resources that must be allocated among them in each discrete time period. Agents have time-varying demands and derive constant marginal utility from each unit of resource received up to their demand, with zero utility for any additional resources. In this setting, it is known that independently maximizing the minimum utility in each round satisfies sharing incentives (agents weakly prefer participating in the mechanism to not participating), strategyproofness (agents have no incentive to misreport their demands), and Pareto efficiency (Freeman et al. 2018). However, recent work (Vuppalapati et al. 2023) has shown that this max-min mechanism can lead to large disparities in the total resources received by agents, even when they have the same average demand. In this paper, we introduce credit fairness, a strengthening of sharing incentives that ensures agents who lend resources in early rounds are able to recoup them in later rounds. Credit fairness can be achieved in conjunction with either Pareto efficiency or strategyproofness, but not both. We propose a mechanism that is credit fair and Pareto efficient, and we evaluate its performance in a computational resource-sharing setting.", "date": "2026-01-25", "categories": []}
{"id": "2601.18342", "title": "Structural Gender Bias in Credit Scoring: Proxy Leakage", "abstract": "As financial institutions increasingly adopt machine learning for credit risk assessment, the persistence of algorithmic bias remains a critical barrier to equitable financial inclusion. This study provides a comprehensive audit of structural gender bias within the Taiwan Credit Default dataset, specifically challenging the prevailing doctrine of \"fairness through blindness.\" Despite the removal of explicit protected attributes and the application of industry standard fairness interventions, our results demonstrate that gendered predictive signals remain deeply embedded within non-sensitive features. Utilizing SHAP (SHapley Additive exPlanations), we identify that variables such as Marital Status, Age, and Credit Limit function as potent proxies for gender, allowing models to maintain discriminatory pathways while appearing statistically fair. To mathematically quantify this leakage, we employ an adversarial inverse modeling framework. Our findings reveal that the protected gender attribute can be reconstructed from purely non-sensitive financial features with an ROC AUC score of 0.65, demonstrating that traditional fairness audits are insufficient for detecting implicit structural bias. These results advocate for a shift from surface-level statistical parity toward causal-aware modeling and structural accountability in financial AI.", "date": "2026-01-26", "categories": []}
{"id": "2501.13223", "title": "Data Matters Most: Auditing Social Bias in Contrastive Vision Language Models", "abstract": "Vision-language models (VLMs) deliver strong zero-shot recognition but frequently inherit social biases from their training data. We systematically disentangle three design factors -- model size, training-data scale, and training-data source -- by comparing CLIP and OpenCLIP, two models that share an identical contrastive objective yet differ in encoder width and in the image-text corpora on which they are pre-trained (400M proprietary pairs vs. 400M/2B LAION). Across balanced face-analysis benchmarks, enlarging the encoder reduces gender skew in CLIP but amplifies both gender and racial skew in OpenCLIP; increasing the LAION corpus from 400M to 2B further increases OpenCLIP bias. At matched model and data budgets, substituting proprietary data with LAION improves gender fairness while increasing racial skew, underscoring data source as the primary driver of bias patterns. We also evaluate three post-hoc, test-time debiasing strategies -- Bias Prompts, Prompt Array, and SANER. Debiasing reduces but does not eliminate harm, and its effectiveness is source- and size-dependent: Bias Prompts most effectively reduce gender skew in CLIP at smaller model sizes, whereas Prompt Array and SANER more reliably reduce racial skew in OpenCLIP; scaling LAION reconfigures which method is most fair. Taken together, these findings challenge the assumption that bigger models or datasets are automatically fairer and foreground training data source as the key determinant of both bias and mitigation efficacy. We release code and evaluation scripts to enable transparent, reproducible auditing of future VLMs.", "date": "2025-01-22", "categories": []}
{"id": "2601.16349", "title": "Regional Bias in Large Language Models", "abstract": "This study investigates regional bias in large language models (LLMs), an emerging concern in AI fairness and global representation. We evaluate ten prominent LLMs: GPT-3.5, GPT-4o, Gemini 1.5 Flash, Gemini 1.0 Pro, Claude 3 Opus, Claude 3.5 Sonnet, Llama 3, Gemma 7B, Mistral 7B, and Vicuna-13B using a dataset of 100 carefully designed prompts that probe forced-choice decisions between regions under contextually neutral scenarios. We introduce FAZE, a prompt-based evaluation framework that measures regional bias on a 10-point scale, where higher scores indicate a stronger tendency to favor specific regions. Experimental results reveal substantial variation in bias levels across models, with GPT-3.5 exhibiting the highest bias score (9.5) and Claude 3.5 Sonnet scoring the lowest (2.5). These findings indicate that regional bias can meaningfully undermine the reliability, fairness, and inclusivity of LLM outputs in real-world, cross-cultural applications. This work contributes to AI fairness research by highlighting the importance of inclusive evaluation frameworks and systematic approaches for identifying and mitigating geographic biases in language models.", "date": "2026-01-22", "categories": []}
{"id": "2601.16724", "title": "Mitigating Bias in Automated Grading Systems for ESL Learners: A Contrastive Learning Approach", "abstract": "As Automated Essay Scoring (AES) systems are increasingly used in high-stakes educational settings, concerns regarding algorithmic bias against English as a Second Language (ESL) learners have increased. Current Transformer-based regression models trained primarily on native-speaker corpora often learn spurious correlations between surface-level L2 linguistic features and essay quality. In this study, we conduct a bias study of a fine-tuned DeBERTa-v3 model using the ASAP 2.0 and ELLIPSE datasets, revealing a constrained score scaling for high-proficiency ESL writing where high-proficiency ESL essays receive scores 10.3% lower than Native speaker essays of identical human-rated quality. To mitigate this, we propose applying contrastive learning with a triplet construction strategy: Contrastive Learning with Matched Essay Pairs. We constructed a dataset of 17,161 matched essay pairs and fine-tuned the model using Triplet Margin Loss to align the latent representations of ESL and Native writing. Our approach reduced the high-proficiency scoring disparity by 39.9% (to a 6.2% gap) while maintaining a Quadratic Weighted Kappa (QWK) of 0.76. Post-hoc linguistic analysis suggests the model successfully disentangled sentence complexity from grammatical error, preventing the penalization of valid L2 syntactic structures.", "date": "2026-01-23", "categories": []}
{"id": "2501.12962", "title": "It's complicated. The relationship of algorithmic fairness and non-discrimination provisions for high-risk systems in the EU AI Act", "abstract": "What constitutes a fair decision? This question is not only difficult for humans but becomes more challenging when Artificial Intelligence (AI) models are used. In light of discriminatory algorithmic behaviors, the EU has recently passed the AI Act, which mandates specific rules for high-risk systems, incorporating both traditional legal non-discrimination regulations and machine learning based algorithmic fairness concepts. This paper aims to bridge these two different concepts in the AI Act through: First, a necessary high-level introduction of both concepts targeting legal and computer science-oriented scholars, and second, an in-depth analysis of the AI Act's relationship between legal non-discrimination regulations and algorithmic fairness. Our analysis reveals three key findings: (1.) Most non-discrimination regulations target only high-risk AI systems. (2.) The regulation of high-risk systems encompasses both data input requirements and output monitoring, though these regulations are partly inconsistent and raise questions of computational feasibility. (3.) Finally, we consider the possible (future) interaction of classical EU non-discrimination law and the AI Act regulations. We recommend developing more specific auditing and testing methodologies for AI systems. This paper aims to serve as a foundation for future interdisciplinary collaboration between legal scholars and computer science-oriented machine learning researchers studying discrimination in AI systems.", "date": "2025-01-22", "categories": []}
{"id": "2510.17873", "title": "Auditing and Mitigating Bias in Gender Classification Algorithms: A Data-Centric Approach", "abstract": "Gender classification systems often inherit and amplify demographic imbalances in their training data. We first audit five widely used gender classification datasets, revealing that all suffer from significant intersectional underrepresentation. To measure the downstream impact of these flaws, we train identical MobileNetV2 classifiers on the two most balanced of these datasets, UTKFace and FairFace. Our fairness evaluation shows that even these models exhibit significant bias, misclassifying female faces at a higher rate than male faces and amplifying existing racial skew. To counter these data-induced biases, we construct BalancedFace, a new public dataset created by blending images from FairFace and UTKFace, supplemented with images from other collections to fill missing demographic gaps. It is engineered to equalize subgroup shares across 189 intersections of age, race, and gender using only real, unedited images. When a standard classifier is trained on BalancedFace, it reduces the maximum True Positive Rate gap across racial subgroups by over 50% and brings the average Disparate Impact score 63% closer to the ideal of 1.0 compared to the next-best dataset, all with a minimal loss of overall accuracy. These results underline the profound value of data-centric interventions and provide an openly available resource for fair gender classification research.", "date": "2025-10-17", "categories": []}
{"id": "2601.13448", "title": "Fairness-informed Pareto Optimization : An Efficient Bilevel Framework", "abstract": "Despite their promise, fair machine learning methods often yield Pareto-inefficient models, in which the performance of certain groups can be improved without degrading that of others. This issue arises frequently in traditional in-processing approaches such as fairness-through-regularization. In contrast, existing Pareto-efficient approaches are biased towards a certain perspective on fairness and fail to adapt to the broad range of fairness metrics studied in the literature. In this paper, we present BADR, a simple framework to recover the optimal Pareto-efficient model for any fairness metric. Our framework recovers its models through a Bilevel Adaptive Rescalarisation procedure. The lower level is a weighted empirical risk minimization task where the weights are a convex combination of the groups, while the upper level optimizes the chosen fairness objective. We equip our framework with two novel large-scale, single-loop algorithms, BADR-GD and BADR-SGD, and establish their convergence guarantees. We release badr, an open-source Python toolbox implementing our framework for a variety of learning tasks and fairness metrics. Finally, we conduct extensive numerical experiments demonstrating the advantages of BADR over existing Pareto-efficient approaches to fairness.", "date": "2026-01-19", "categories": []}
{"id": "2601.15306", "title": "Uncovering Latent Bias in LLM-Based Emergency Department Triage Through Proxy Variables", "abstract": "Recent advances in large language models (LLMs) have enabled their integration into clinical decision-making; however, hidden biases against patients across racial, social, economic, and clinical backgrounds persist. In this study, we investigate bias in LLM-based medical AI systems applied to emergency department (ED) triage. We employ 32 patient-level proxy variables, each represented by paired positive and negative qualifiers, and evaluate their effects using both public (MIMIC-IV-ED Demo, MIMIC-IV Demo) and restricted-access credentialed (MIMIC-IV-ED and MIMIC-IV) datasets as appropriate~\\cite{mimiciv_ed_demo,mimiciv_ed,mimiciv}. Our results reveal discriminatory behavior mediated through proxy variables in ED triage scenarios, as well as a systematic tendency for LLMs to modify perceived patient severity when specific tokens appear in the input context, regardless of whether they are framed positively or negatively. These findings indicate that AI systems is still imperfectly trained on noisy, sometimes non-causal signals that do not reliably reflect true patient acuity. Consequently, more needs to be done to ensure the safe and responsible deployment of AI technologies in clinical settings.", "date": "2026-01-13", "categories": []}
{"id": "2508.17825", "title": "FAIRGAMER: Evaluating Social Biases in LLM-Based Video Game NPCs", "abstract": "Large Language Models (LLMs) have increasingly enhanced or replaced traditional Non-Player Characters (NPCs) in video games. However, these LLM-based NPCs inherit underlying social biases (e.g., race or class), posing fairness risks during in-game interactions. To address the limited exploration of this issue, we introduce FairGamer, the first benchmark to evaluate social biases across three interaction patterns: transaction, cooperation, and competition. FairGamer assesses four bias types, including class, race, age, and nationality, across 12 distinct evaluation tasks using a novel metric, FairMCV. Our evaluation of seven frontier LLMs reveals that: (1) models exhibit biased decision-making, with Grok-4-Fast demonstrating the highest bias (average FairMCV = 76.9%); and (2) larger LLMs display more severe social biases, suggesting that increased model capacity inadvertently amplifies these biases. We release FairGamer at https://github.com/Anonymous999-xxx/FairGamer to facilitate future research on NPC fairness.", "date": "2025-08-25", "categories": []}
{"id": "2601.14553", "title": "Self-Blinding and Counterfactual Self-Simulation Mitigate Biases and Sycophancy in Large Language Models", "abstract": "Fair decisions require ignoring irrelevant, potentially biasing, information. To achieve this, decision-makers need to approximate what decision they would have made had they not known certain facts, such as the gender or race of a job candidate. This counterfactual self-simulation is notoriously hard for humans, leading to biased judgments even by well-meaning actors. Here we show that large language models (LLMs) suffer from similar limitations in their ability to approximate what decisions they would make under counterfactual knowledge in offsetting gender and race biases and overcoming sycophancy. We show that prompting models to ignore or pretend not to know biasing information fails to offset these biases and occasionally backfires. However, unlike humans, LLMs can be given access to a ground-truth model of their own counterfactual cognition -- their own API. We show that this access to the responses of a blinded replica enables fairer decisions, while providing greater transparency to distinguish implicit from intentionally biased behavior.", "date": "2026-01-21", "categories": []}
{"id": "2503.06431", "title": "Fairness-aware kidney exchange and kidney paired donation", "abstract": "The kidney paired donation (KPD) program provides an innovative solution to overcome incompatibility challenges in kidney transplants by matching incompatible donor-patient pairs and facilitating kidney exchanges. To address unequal access to transplant opportunities, there are two widely used fairness criteria: group fairness and individual fairness. However, these criteria do not consider protected patient features, which refer to characteristics legally or ethically recognized as needing protection from discrimination, such as race and gender. Motivated by the calibration principle in machine learning, we introduce a new fairness criterion: the matching outcome should be conditionally independent of the protected feature, given the sensitization level. We integrate this fairness criterion as a constraint within the KPD optimization framework and propose a computationally efficient solution using linearization strategies and column-generation methods. Theoretically, we analyze the associated price of fairness using random graph models. Empirically, we compare our fairness criterion with group fairness and individual fairness through both simulations and a real-data example.", "date": "2025-03-09", "categories": []}
{"id": "2508.01412", "title": "Bias Association Discovery Framework for Open-Ended LLM Generations", "abstract": "Social biases embedded in Large Language Models (LLMs) raise critical concerns, resulting in representational harms -- unfair or distorted portrayals of demographic groups -- that may be expressed in subtle ways through generated language. Existing evaluation methods often depend on predefined identity-concept associations, limiting their ability to surface new or unexpected forms of bias. In this work, we present the Bias Association Discovery Framework (BADF), a systematic approach for extracting both known and previously unrecognized associations between demographic identities and descriptive concepts from open-ended LLM outputs. Through comprehensive experiments spanning multiple models and diverse real-world contexts, BADF enables robust mapping and analysis of the varied concepts that characterize demographic identities. Our findings advance the understanding of biases in open-ended generation and provide a scalable tool for identifying and analyzing bias associations in LLMs.", "date": "2025-08-02", "categories": []}
{"id": "2509.24340", "title": "humancompatible.detect: a Python Toolkit for Detecting Bias in AI Models", "abstract": "There is a strong recent emphasis on trustworthy AI. In particular, international regulations, such as the AI Act, demand that AI practitioners measure data quality on the input and estimate bias on the output of high-risk AI systems. However, there are many challenges involved, including scalability (MMD) and computability (Wasserstein-1) issues of traditional methods for estimating distances on measure spaces. Here, we present humancompatible.detect, a toolkit for bias detection that addresses these challenges. It incorporates two newly developed methods to detect and evaluate bias: maximum subgroup discrepancy (MSD) and subsampled $\\ell_\\infty$ distances. It has an easy-to-use API documented with multiple examples. humancompatible.detect is licensed under the Apache License, Version 2.0.", "date": "2025-09-29", "categories": []}
{"id": "2601.09394", "title": "FairGE: Fairness-Aware Graph Encoding in Incomplete Social Networks", "abstract": "Graph Transformers (GTs) are increasingly applied to social network analysis, yet their deployment is often constrained by fairness concerns. This issue is particularly critical in incomplete social networks, where sensitive attributes are frequently missing due to privacy and ethical restrictions. Existing solutions commonly generate these incomplete attributes, which may introduce additional biases and further compromise user privacy. To address this challenge, FairGE (Fair Graph Encoding) is introduced as a fairness-aware framework for GTs in incomplete social networks. Instead of generating sensitive attributes, FairGE encodes fairness directly through spectral graph theory. By leveraging the principal eigenvector to represent structural information and padding incomplete sensitive attributes with zeros to maintain independence, FairGE ensures fairness without data reconstruction. Theoretical analysis demonstrates that the method suppresses the influence of non-principal spectral components, thereby enhancing fairness. Extensive experiments on seven real-world social network datasets confirm that FairGE achieves at least a 16% improvement in both statistical parity and equality of opportunity compared with state-of-the-art baselines. The source code is shown in https://github.com/LuoRenqiang/FairGE.", "date": "2026-01-14", "categories": []}
{"id": "2601.09841", "title": "A pipeline for enabling path-specific causal fairness in observational health data", "abstract": "When training machine learning (ML) models for potential deployment in a healthcare setting, it is essential to ensure that they do not replicate or exacerbate existing healthcare biases. Although many definitions of fairness exist, we focus on path-specific causal fairness, which allows us to better consider the social and medical contexts in which biases occur (e.g., direct discrimination by a clinician or model versus bias due to differential access to the healthcare system) and to characterize how these biases may appear in learned models. In this work, we map the structural fairness model to the observational healthcare setting and create a generalizable pipeline for training causally fair models. The pipeline explicitly considers specific healthcare context and disparities to define a target \"fair\" model. Our work fills two major gaps: first, we expand on characterizations of the \"fairness-accuracy\" tradeoff by detangling direct and indirect sources of bias and jointly presenting these fairness considerations alongside considerations of accuracy in the context of broadly known biases. Second, we demonstrate how a foundation model trained without fairness constraints on observational health data can be leveraged to generate causally fair downstream predictions in tasks with known social and medical disparities. This work presents a model-agnostic pipeline for training causally fair machine learning models that address both direct and indirect forms of healthcare bias.", "date": "2026-01-14", "categories": []}
{"id": "2601.11897", "title": "Task-tailored Pre-processing: Fair Downstream Supervised Learning", "abstract": "Fairness-aware machine learning has recently attracted various communities to mitigate discrimination against certain societal groups in data-driven tasks. For fair supervised learning, particularly in pre-processing, there have been two main categories: data fairness and task-tailored fairness. The former directly finds an intermediate distribution among the groups, independent of the type of the downstream model, so a learned downstream classification/regression model returns similar predictive scores to individuals inputting the same covariates irrespective of their sensitive attributes. The latter explicitly takes the supervised learning task into account when constructing the pre-processing map. In this work, we study algorithmic fairness for supervised learning and argue that the data fairness approaches impose overly strong regularization from the perspective of the HGR correlation. This motivates us to devise a novel pre-processing approach tailored to supervised learning. We account for the trade-off between fairness and utility in obtaining the pre-processing map. Then we study the behavior of arbitrary downstream supervised models learned on the transformed data to find sufficient conditions to guarantee their fairness improvement and utility preservation. To our knowledge, no prior work in the branch of task-tailored methods has theoretically investigated downstream guarantees when using pre-processed data. We further evaluate our framework through comparison studies based on tabular and image data sets, showing the superiority of our framework which preserves consistent trade-offs among multiple downstream models compared to recent competing models. Particularly for computer vision data, we see our method alters only necessary semantic features related to the central machine learning task to achieve fairness.", "date": "2026-01-17", "categories": []}
{"id": "2601.12033", "title": "Preserving Fairness and Safety in Quantized LLMs Through Critical Weight Protection", "abstract": "Quantization is widely adopted to reduce the computational cost of large language models (LLMs); however, its implications for fairness and safety, particularly in dynamic quantization and multilingual contexts, remain underexplored. In this work, we conduct a systematic study of how static and dynamic quantization methods impact fairness and safety across benchmarks measuring intrinsic and extrinsic bias and safety alignment. For fairness, we evaluate English, French, Dutch, Spanish, and Turkish; for safety, we focus on English, Korean, and Arabic. Our findings reveal that quantization consistently degrades fairness and safety, with dynamic methods demonstrating greater stability than static ones. Moreover, fairness degradation varies across languages, while safety deterioration is especially pronounced in non-English settings. To address these risks, we introduce Critical Weight Protection, a novel technique that identifies and preserves fairness- and safety-critical weights during quantization. This approach effectively mitigates bias and safety deterioration without costly retraining or alignment, maintaining trustworthiness while retaining efficiency.", "date": "2026-01-17", "categories": []}
{"id": "2601.12091", "title": "Mitigating Cultural Bias in LLMs via Multi-Agent Cultural Debate", "abstract": "Large language models (LLMs) exhibit systematic Western-centric bias, yet whether prompting in non-Western languages (e.g., Chinese) can mitigate this remains understudied. Answering this question requires rigorous evaluation and effective mitigation, but existing approaches fall short on both fronts: evaluation methods force outputs into predefined cultural categories without a neutral option, while mitigation relies on expensive multi-cultural corpora or agent frameworks that use functional roles (e.g., Planner--Critique) lacking explicit cultural representation. To address these gaps, we introduce CEBiasBench, a Chinese--English bilingual benchmark, and Multi-Agent Vote (MAV), which enables explicit ``no bias'' judgments. Using this framework, we find that Chinese prompting merely shifts bias toward East Asian perspectives rather than eliminating it. To mitigate such persistent bias, we propose Multi-Agent Cultural Debate (MACD), a training-free framework that assigns agents distinct cultural personas and orchestrates deliberation via a \"Seeking Common Ground while Reserving Differences\" strategy. Experiments demonstrate that MACD achieves 57.6% average No Bias Rate evaluated by LLM-as-judge and 86.0% evaluated by MAV (vs. 47.6% and 69.0% baseline using GPT-4o as backbone) on CEBiasBench and generalizes to the Arabic CAMeL benchmark, confirming that explicit cultural representation in agent frameworks is essential for cross-cultural fairness.", "date": "2026-01-17", "categories": []}
{"id": "2601.11379", "title": "Evaluating LLM Behavior in Hiring: Implicit Weights, Fairness Across Groups, and Alignment with Human Preferences", "abstract": "General-purpose Large Language Models (LLMs) show significant potential in recruitment applications, where decisions require reasoning over unstructured text, balancing multiple criteria, and inferring fit and competence from indirect productivity signals. Yet, it is still uncertain how LLMs assign importance to each attribute and whether such assignments are in line with economic principles, recruiter preferences or broader societal norms. We propose a framework to evaluate an LLM's decision logic in recruitment, by drawing on established economic methodologies for analyzing human hiring behavior. We build synthetic datasets from real freelancer profiles and project descriptions from a major European online freelance marketplace and apply a full factorial design to estimate how a LLM weighs different match-relevant criteria when evaluating freelancer-project fit. We identify which attributes the LLM prioritizes and analyze how these weights vary across project contexts and demographic subgroups. Finally, we explain how a comparable experimental setup could be implemented with human recruiters to assess alignment between model and human decisions. Our findings reveal that the LLM weighs core productivity signals, such as skills and experience, but interprets certain features beyond their explicit matching value. While showing minimal average discrimination against minority groups, intersectional effects reveal that productivity signals carry different weights between demographic groups.", "date": "2026-01-16", "categories": []}
{"id": "2406.12805", "title": "AITTI: Learning Adaptive Inclusive Token for Text-to-Image Generation", "abstract": "Despite the high-quality results of text-to-image generation, stereotypical biases have been spotted in their generated contents, compromising the fairness of generative models. In this work, we propose to learn adaptive inclusive tokens to shift the attribute distribution of the final generative outputs. Unlike existing de-biasing approaches, our method requires neither explicit attribute specification nor prior knowledge of the bias distribution. Specifically, the core of our method is a lightweight adaptive mapping network, which can customize the inclusive tokens for the concepts to be de-biased, making the tokens generalizable to unseen concepts regardless of their original bias distributions. This is achieved by tuning the adaptive mapping network with a handful of balanced and inclusive samples using an anchor loss. Experimental results demonstrate that our method outperforms previous bias mitigation methods without attribute specification while preserving the alignment between generative results and text descriptions. Moreover, our method achieves comparable performance to models that require specific attributes or editing directions for generation. Extensive experiments showcase the effectiveness of our adaptive inclusive tokens in mitigating stereotypical bias in text-to-image generation. The code will be available at https://github.com/itsmag11/AITTI.", "date": "2024-06-18", "categories": []}
{"id": "2407.18454", "title": "Fairness Definitions in Language Models Explained", "abstract": "Language Models (LMs) have demonstrated exceptional performance across various Natural Language Processing (NLP) tasks. Despite these advancements, LMs can inherit and amplify societal biases related to sensitive attributes such as gender and race, limiting their adoption in real-world applications. Therefore, fairness has been extensively explored in LMs, leading to the proposal of various fairness notions. However, the lack of clear agreement on which fairness definition to apply in specific contexts and the complexity of understanding the distinctions between these definitions can create confusion and impede further progress. To this end, this paper proposes a systematic survey that clarifies the definitions of fairness as they apply to LMs. Specifically, we begin with a brief introduction to LMs and fairness in LMs, followed by a comprehensive, up-to-date overview of existing fairness notions in LMs and the introduction of a novel taxonomy that categorizes these concepts based on their transformer architecture: encoder-only, decoder-only, and encoder-decoder LMs. We further illustrate each definition through experiments, showcasing their practical implications and outcomes. Finally, we discuss current research challenges and open questions, aiming to foster innovative ideas and advance the field. The repository is publicly available online at https://github.com/vanbanTruong/Fairness-in-Large-Language-Models/tree/main/definitions.", "date": "2024-07-26", "categories": []}
{"id": "2504.03716", "title": "Evaluating Large Language Models for Fair and Reliable Organ Allocation", "abstract": "Medical institutions are considering the use of LLMs in high-stakes clinical decision-making, such as organ allocation. In such sensitive use cases, evaluating fairness is imperative. However, existing evaluation methods often fall short; benchmarks are too simplistic to capture real-world complexity, and accuracy-based metrics fail to address the absence of a clear ground truth. To realistically and fairly model organ allocation, specifically kidney allocation, we begin by testing the medical knowledge of LLMs to determine whether they understand the clinical factors required to make sound allocation decisions. Building on this foundation, we design two tasks: (1) Choose-One and (2) Rank-All. In Choose-One, LLMs select a single candidate from a list of potential candidates to receive a kidney. In this scenario, we assess fairness across demographics using traditional fairness metrics, such as proportional parity. In Rank-All, LLMs rank all candidates waiting for a kidney, reflecting real-world allocation processes more closely, where an organ is passed down a ranked list until allocated. Our evaluation on three LLMs reveals a divergence between fairness metrics: while exposure-based metrics suggest equitable outcomes, probability-based metrics uncover systematic preferential sorting, where specific groups were clustered in upper-ranking tiers. Furthermore, we observe that demographic preferences are highly task-dependent, showing inverted trends between Choose-One and Rank-All tasks, even when considering the topmost rank. Overall, our results indicate that current LLMs can introduce inequalities in real-world allocation scenarios, underscoring the urgent need for rigorous fairness evaluation and human oversight before their use in high-stakes decision-making.", "date": "2025-03-29", "categories": []}
{"id": "2601.09421", "title": "Bias Dynamics in BabyLMs: Towards a Compute-Efficient Sandbox for Democratising Pre-Training Debiasing", "abstract": "Pre-trained language models (LMs) have, over the last few years, grown substantially in both societal adoption and training costs. This rapid growth in size has constrained progress in understanding and mitigating their biases. Since re-training LMs is prohibitively expensive, most debiasing work has focused on post-hoc or masking-based strategies, which often fail to address the underlying causes of bias. In this work, we seek to democratise pre-model debiasing research by using low-cost proxy models. Specifically, we investigate BabyLMs, compact BERT-like models trained on small and mutable corpora that can approximate bias acquisition and learning dynamics of larger models. We show that BabyLMs display closely aligned patterns of intrinsic bias formation and performance development compared to standard BERT models, despite their drastically reduced size. Furthermore, correlations between BabyLMs and BERT hold across multiple intra-model and post-model debiasing methods. Leveraging these similarities, we conduct pre-model debiasing experiments with BabyLMs, replicating prior findings and presenting new insights regarding the influence of gender imbalance and toxicity on bias formation. Our results demonstrate that BabyLMs can serve as an effective sandbox for large-scale LMs, reducing pre-training costs from over 500 GPU-hours to under 30 GPU-hours. This provides a way to democratise pre-model debiasing research and enables faster, more accessible exploration of methods for building fairer LMs.", "date": "2026-01-14", "categories": []}
{"id": "2601.10460", "title": "Contextual StereoSet: Stress-Testing Bias Alignment Robustness in Large Language Models", "abstract": "A model that avoids stereotypes in a lab benchmark may not avoid them in deployment. We show that measured bias shifts dramatically when prompts mention different places, times, or audiences -- no adversarial prompting required.\n  We introduce Contextual StereoSet, a benchmark that holds stereotype content fixed while systematically varying contextual framing. Testing 13 models across two protocols, we find striking patterns: anchoring to 1990 (vs. 2030) raises stereotype selection in all models tested on this contrast (p<0.05); gossip framing raises it in 5 of 6 full-grid models; out-group observer framing shifts it by up to 13 percentage points. These effects replicate in hiring, lending, and help-seeking vignettes.\n  We propose Context Sensitivity Fingerprints (CSF): a compact profile of per-dimension dispersion and paired contrasts with bootstrap CIs and FDR correction. Two evaluation tracks support different use cases -- a 360-context diagnostic grid for deep analysis and a budgeted protocol covering 4,229 items for production screening.\n  The implication is methodological: bias scores from fixed-condition tests may not generalize.This is not a claim about ground-truth bias rates; it is a stress test of evaluation robustness. CSF forces evaluators to ask, \"Under what conditions does bias appear?\" rather than \"Is this model biased?\" We release our benchmark, code, and results.", "date": "2026-01-15", "categories": []}
{"id": "2502.16841", "title": "Fair Foundation Models for Medical Image Analysis: Challenges and Perspectives", "abstract": "Ensuring equitable Artificial Intelligence (AI) in healthcare demands systems that make unbiased decisions across all demographic groups, bridging technical innovation with ethical principles. Foundation Models (FMs), trained on vast datasets through self-supervised learning, enable efficient adaptation across medical imaging tasks while reducing dependency on labeled data. These models demonstrate potential for enhancing fairness, though significant challenges remain in achieving consistent performance across demographic groups. Our review indicates that effective bias mitigation in FMs requires systematic interventions throughout all stages of development. While previous approaches focused primarily on model-level bias mitigation, our analysis reveals that fairness in FMs requires integrated interventions throughout the development pipeline, from data documentation to deployment protocols. This comprehensive framework advances current knowledge by demonstrating how systematic bias mitigation, combined with policy engagement, can effectively address both technical and institutional barriers to equitable AI in healthcare. The development of equitable FMs represents a critical step toward democratizing advanced healthcare technologies, particularly for underserved populations and regions with limited medical infrastructure and computational resources.", "date": "2025-02-24", "categories": []}
{"id": "2512.17340", "title": "Penalized Fair Regression for Multiple Groups in Chronic Kidney Disease", "abstract": "Fair regression methods have the potential to mitigate societal bias concerns in health care, but there has been little work on penalized fair regression when multiple groups experience such bias. We propose a general regression framework that addresses this gap with unfairness penalties for multiple groups. Our approach is demonstrated for binary outcomes with true positive rate disparity penalties. It can be efficiently implemented through reduction to a cost-sensitive classification problem. We additionally introduce novel score functions for automatically selecting penalty weights. Our penalized fair regression methods are empirically studied in simulations, where they achieve a fairness-accuracy frontier beyond that of existing comparison methods. Finally, we apply these methods to a national multi-site primary care study of chronic kidney disease to develop a fair classifier for end-stage renal disease. There we find substantial improvements in fairness for multiple race and ethnicity groups who experience societal bias in the health care system without any appreciable loss in overall fit.", "date": "2025-12-19", "categories": []}
{"id": "2601.08860", "title": "Bias Detection and Rotation-Robustness Mitigation in Vision-Language Models and Generative Image Models", "abstract": "Vision-Language Models (VLMs) and generative image models have achieved remarkable performance across multimodal tasks, yet their robustness and fairness under input transformations remain insufficiently explored. This work investigates bias propagation and robustness degradation in state-of-the-art vision-language and generative models, with a particular focus on image rotation and distributional shifts. We analyze how rotation-induced perturbations affect model predictions, confidence calibration, and demographic bias patterns. To address these issues, we propose rotation-robust mitigation strategies that combine data augmentation, representation alignment, and model-level regularization. Experimental results across multiple datasets demonstrate that the proposed methods significantly improve robustness while reducing bias amplification without sacrificing overall performance. This study highlights critical limitations of current multimodal systems and provides practical mitigation techniques for building more reliable and fair AI models.", "date": "2026-01-09", "categories": []}
{"id": "2601.08953", "title": "Fairness risk and its privacy-enabled solution in AI-driven robotic applications", "abstract": "Complex decision-making by autonomous machines and algorithms could underpin the foundations of future society. Generative AI is emerging as a powerful engine for such transitions. However, we show that Generative AI-driven developments pose a critical pitfall: fairness concerns. In robotic applications, although intuitions about fairness are common, a precise and implementable definition that captures user utility and inherent data randomness is missing. Here we provide a utility-aware fairness metric for robotic decision making and analyze fairness jointly with user-data privacy, deriving conditions under which privacy budgets govern fairness metrics. This yields a unified framework that formalizes and quantifies fairness and its interplay with privacy, which is tested in a robot navigation task. In view of the fact that under legal requirements, most robotic systems will enforce user privacy, the approach shows surprisingly that such privacy budgets can be jointly used to meet fairness targets. Addressing fairness concerns in the creative combined consideration of privacy is a step towards ethical use of AI and strengthens trust in autonomous robots deployed in everyday environments.", "date": "2026-01-13", "categories": []}
{"id": "2601.08976", "title": "Continuous Fairness On Data Streams", "abstract": "We study the problem of enforcing continuous group fairness over windows in data streams. We propose a novel fairness model that ensures group fairness at a finer granularity level (referred to as block) within each sliding window. This formulation is particularly useful when the window size is large, making it desirable to enforce fairness at a finer granularity. Within this framework, we address two key challenges: efficiently monitoring whether each sliding window satisfies block-level group fairness, and reordering the current window as effectively as possible when fairness is violated. To enable real-time monitoring, we design sketch-based data structures that maintain attribute distributions with minimal overhead. We also develop optimal, efficient algorithms for the reordering task, supported by rigorous theoretical guarantees. Our evaluation on four real-world streaming scenarios demonstrates the practical effectiveness of our approach. We achieve millisecond-level processing and a throughput of approximately 30,000 queries per second on average, depending on system parameters. The stream reordering algorithm improves block-level group fairness by up to 95% in certain cases, and by 50-60% on average across datasets. A qualitative study further highlights the advantages of block-level fairness compared to window-level fairness.", "date": "2026-01-13", "categories": []}
{"id": "2601.09141", "title": "Identity-Robust Language Model Generation via Content Integrity Preservation", "abstract": "Large Language Model (LLM) outputs often vary across user sociodemographic attributes, leading to disparities in factual accuracy, utility, and safety, even for objective questions where demographic information is irrelevant. Unlike prior work on stereotypical or representational bias, this paper studies identity-dependent degradation of core response quality. We show empirically that such degradation arises from biased generation behavior, despite factual knowledge being robustly encoded across identities. Motivated by this mismatch, we propose a lightweight, training-free framework for identity-robust generation that selectively neutralizes non-critical identity information while preserving semantically essential attributes, thus maintaining output content integrity. Experiments across four benchmarks and 18 sociodemographic identities demonstrate an average 77% reduction in identity-dependent bias compared to vanilla prompting and a 45% reduction relative to prompt-based defenses. Our work addresses a critical gap in mitigating the impact of user identity cues in prompts on core generation quality.", "date": "2026-01-14", "categories": []}
{"id": "2601.08784", "title": "On the use of graph models to achieve individual and group fairness", "abstract": "Machine Learning algorithms are ubiquitous in key decision-making contexts such as justice, healthcare and finance, which has spawned a great demand for fairness in these procedures. However, the theoretical properties of such models in relation with fairness are still poorly understood, and the intuition behind the relationship between group and individual fairness is still lacking. In this paper, we provide a theoretical framework based on Sheaf Diffusion to leverage tools based on dynamical systems and homology to model fairness. Concretely, the proposed method projects input data into a bias-free space that encodes fairness constrains, resulting in fair solutions. Furthermore, we present a collection of network topologies handling different fairness metrics, leading to a unified method capable of dealing with both individual and group bias. The resulting models have a layer of interpretability in the form of closed-form expressions for their SHAP values, consolidating their place in the responsible Artificial Intelligence landscape. Finally, these intuitions are tested on a simulation study and standard fairness benchmarks, where the proposed methods achieve satisfactory results. More concretely, the paper showcases the performance of the proposed models in terms of accuracy and fairness, studying available trade-offs on the Pareto frontier, checking the effects of changing the different hyper-parameters, and delving into the interpretation of its outputs.", "date": "2026-01-13", "categories": []}
{"id": "2511.06148", "title": "Large Language Models Develop Novel Social Biases Through Adaptive Exploration", "abstract": "As large language models (LLMs) are adopted into frameworks that grant them the capacity to make real decisions, it is increasingly important to ensure that they are unbiased. In this paper, we argue that the predominant approach of simply removing existing biases from models is not enough. Using a paradigm from the psychology literature, we demonstrate that LLMs can spontaneously develop novel social biases about artificial demographic groups even when no inherent differences exist. These biases result in highly stratified task allocations, which are less fair than assignments by human participants and are exacerbated by newer and larger models. In social science, emergent biases like these have been shown to result from exploration-exploitation trade-offs, where the decision-maker explores too little, allowing early observations to strongly influence impressions about entire demographic groups. To alleviate this effect, we examine a series of interventions targeting model inputs, problem structure, and explicit steering. We find that explicitly incentivizing exploration most robustly reduces stratification, highlighting the need for better multifaceted objectives to mitigate bias. These results reveal that LLMs are not merely passive mirrors of human social biases, but can actively create new ones from experience, raising urgent questions about how these systems will shape societies over time.", "date": "2025-11-08", "categories": []}
{"id": "2601.06761", "title": "Comparative Separation: Evaluating Separation on Comparative Judgment Test Data", "abstract": "This research seeks to benefit the software engineering society by proposing comparative separation, a novel group fairness notion to evaluate the fairness of machine learning software on comparative judgment test data. Fairness issues have attracted increasing attention since machine learning software is increasingly used for high-stakes and high-risk decisions. It is the responsibility of all software developers to make their software accountable by ensuring that the machine learning software do not perform differently on different sensitive groups -- satisfying the separation criterion. However, evaluation of separation requires ground truth labels for each test data point. This motivates our work on analyzing whether separation can be evaluated on comparative judgment test data. Instead of asking humans to provide the ratings or categorical labels on each test data point, comparative judgments are made between pairs of data points such as A is better than B. According to the law of comparative judgment, providing such comparative judgments yields a lower cognitive burden for humans than providing ratings or categorical labels. This work first defines the novel fairness notion comparative separation on comparative judgment test data, and the metrics to evaluate comparative separation. Then, both theoretically and empirically, we show that in binary classification problems, comparative separation is equivalent to separation. Lastly, we analyze the number of test data points and test data pairs required to achieve the same level of statistical power in the evaluation of separation and comparative separation, respectively. This work is the first to explore fairness evaluation on comparative judgment test data. It shows the feasibility and the practical benefits of using comparative judgment test data for model evaluations.", "date": "2026-01-11", "categories": []}
{"id": "2601.05663", "title": "Tracing Stereotypes in Pre-trained Transformers: From Biased Neurons to Fairer Models", "abstract": "The advent of transformer-based language models has reshaped how AI systems process and generate text. In software engineering (SE), these models now support diverse activities, accelerating automation and decision-making. Yet, evidence shows that these models can reproduce or amplify social biases, raising fairness concerns. Recent work on neuron editing has shown that internal activations in pre-trained transformers can be traced and modified to alter model behavior. Building on the concept of knowledge neurons, neurons that encode factual information, we hypothesize the existence of biased neurons that capture stereotypical associations within pre-trained transformers. To test this hypothesis, we build a dataset of biased relations, i.e., triplets encoding stereotypes across nine bias types, and adapt neuron attribution strategies to trace and suppress biased neurons in BERT models. We then assess the impact of suppression on SE tasks. Our findings show that biased knowledge is localized within small neuron subsets, and suppressing them substantially reduces bias with minimal performance loss. This demonstrates that bias in transformers can be traced and mitigated at the neuron level, offering an interpretable approach to fairness in SE.", "date": "2026-01-09", "categories": []}
{"id": "2601.05909", "title": "Auditing Fairness under Model Updates: Fundamental Complexity and Property-Preserving Updates", "abstract": "As machine learning models become increasingly embedded in societal infrastructure, auditing them for bias is of growing importance. However, in real-world deployments, auditing is complicated by the fact that model owners may adaptively update their models in response to changing environments, such as financial markets. These updates can alter the underlying model class while preserving certain properties of interest, raising fundamental questions about what can be reliably audited under such shifts.\n  In this work, we study group fairness auditing under arbitrary updates. We consider general shifts that modify the pre-audit model class while maintaining invariance of the audited property. Our goals are two-fold: (i) to characterize the information complexity of allowable updates, by identifying which strategic changes preserve the property under audit; and (ii) to efficiently estimate auditing properties, such as group fairness, using a minimal number of labeled samples.\n  We propose a generic framework for PAC auditing based on an Empirical Property Optimization (EPO) oracle. For statistical parity, we establish distribution-free auditing bounds characterized by the SP dimension, a novel combinatorial measure that captures the complexity of admissible strategic updates. Finally, we demonstrate that our framework naturally extends to other auditing objectives, including prediction error and robust risk.", "date": "2026-01-09", "categories": []}
{"id": "2402.12937", "title": "GRAPHGINI: Fostering Individual and Group Fairness in Graph Neural Networks", "abstract": "Graph Neural Networks (GNNs) have demonstrated impressive performance across various tasks, leading to their increased adoption in high-stakes decision-making systems. However, concerns have arisen about GNNs potentially generating unfair decisions for underprivileged groups or individuals when lacking fairness constraints. This work addresses this issue by introducing GraphGini, a novel approach that incorporates the Gini coefficient to enhance both individual and group fairness within the GNN framework. We rigorously establish that the Gini coefficient offers greater robustness and promotes equal opportunity among GNN outcomes, advantages not afforded by the prevailing Lipschitz constant methodology. Additionally, we employ the Nash social welfare program to ensure our solution yields a Pareto optimal distribution of group fairness. Extensive experimentation on real-world datasets demonstrates GraphGini's efficacy in significantly improving individual fairness compared to state-of-the-art methods while maintaining utility and group fairness.", "date": "2024-02-20", "categories": []}
{"id": "2601.04730", "title": "Automatic Classifiers Underdetect Emotions Expressed by Men", "abstract": "The widespread adoption of automatic sentiment and emotion classifiers makes it important to ensure that these tools perform reliably across different populations. Yet their reliability is typically assessed using benchmarks that rely on third-party annotators rather than the individuals experiencing the emotions themselves, potentially concealing systematic biases. In this paper, we use a unique, large-scale dataset of more than one million self-annotated posts and a pre-registered research design to investigate gender biases in emotion detection across 414 combinations of models and emotion-related classes. We find that across different types of automatic classifiers and various underlying emotions, error rates are consistently higher for texts authored by men compared to those authored by women. We quantify how this bias could affect results in downstream applications and show that current machine learning tools, including large language models, should be applied with caution when the gender composition of a sample is not known or variable. Our findings demonstrate that sentiment analysis is not yet a solved problem, especially in ensuring equitable model behaviour across demographic groups.", "date": "2026-01-08", "categories": []}
{"id": "2601.05184", "title": "Observations and Remedies for Large Language Model Bias in Self-Consuming Performative Loop", "abstract": "The rapid advancement of large language models (LLMs) has led to growing interest in using synthetic data to train future models. However, this creates a self-consuming retraining loop, where models are trained on their own outputs and may cause performance drops and induce emerging biases. In real-world applications, previously deployed LLMs may influence the data they generate, leading to a dynamic system driven by user feedback. For example, if a model continues to underserve users from a group, less query data will be collected from this particular demographic of users. In this study, we introduce the concept of \\textbf{S}elf-\\textbf{C}onsuming \\textbf{P}erformative \\textbf{L}oop (\\textbf{SCPL}) and investigate the role of synthetic data in shaping bias during these dynamic iterative training processes under controlled performative feedback. This controlled setting is motivated by the inaccessibility of real-world user preference data from dynamic production systems, and enables us to isolate and analyze feedback-driven bias evolution in a principled manner. We focus on two types of loops, including the typical retraining setting and the incremental fine-tuning setting, which is largely underexplored. Through experiments on three real-world tasks, we find that the performative loop increases preference bias and decreases disparate bias. We design a reward-based rejection sampling strategy to mitigate the bias, moving towards more trustworthy self-improving systems.", "date": "2026-01-08", "categories": []}
{"id": "2310.19917", "title": "Unmasking Bias in AI: A Systematic Review of Bias Detection and Mitigation Strategies in Electronic Health Record-based Models", "abstract": "Objectives: Leveraging artificial intelligence (AI) in conjunction with electronic health records (EHRs) holds transformative potential to improve healthcare. Yet, addressing bias in AI, which risks worsening healthcare disparities, cannot be overlooked. This study reviews methods to detect and mitigate diverse forms of bias in AI models developed using EHR data. Methods: We conducted a systematic review following the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) guidelines, analyzing articles from PubMed, Web of Science, and IEEE published between January 1, 2010, and Dec 17, 2023. The review identified key biases, outlined strategies for detecting and mitigating bias throughout the AI model development process, and analyzed metrics for bias assessment. Results: Of the 450 articles retrieved, 20 met our criteria, revealing six major bias types: algorithmic, confounding, implicit, measurement, selection, and temporal. The AI models were primarily developed for predictive tasks in healthcare settings. Four studies concentrated on the detection of implicit and algorithmic biases employing fairness metrics like statistical parity, equal opportunity, and predictive equity. Sixty proposed various strategies for mitigating biases, especially targeting implicit and selection biases. These strategies, evaluated through both performance (e.g., accuracy, AUROC) and fairness metrics, predominantly involved data collection and preprocessing techniques like resampling, reweighting, and transformation. Discussion: This review highlights the varied and evolving nature of strategies to address bias in EHR-based AI models, emphasizing the urgent needs for the establishment of standardized, generalizable, and interpretable methodologies to foster the creation of ethical AI systems that promote fairness and equity in healthcare.", "date": "2023-10-30", "categories": []}
{"id": "2502.03826", "title": "FairT2I: Mitigating Social Bias in Text-to-Image Generation via Large Language Model-Assisted Detection and Attribute Rebalancing", "abstract": "Text-to-image (T2I) models have advanced creative content generation, yet their reliance on large uncurated datasets often reproduces societal biases. We present FairT2I, a training-free and interactive framework grounded in a mathematically principled latent variable guidance formulation. This formulation decomposes the generative score function into attribute-conditioned components and reweights them according to a defined distribution, providing a unified and flexible mechanism for bias-aware generation that also subsumes many existing ad hoc debiasing approaches as special cases. Building upon this foundation, FairT2I incorporates (1) latent variable guidance as the core mechanism, (2) LLM-based bias detection to automatically infer bias-prone categories and attributes from text prompts as part of the latent structure, and (3) attribute resampling, which allows users to adjust or redefine the attribute distribution based on uniform, real-world, or user-specified statistics. The accompanying user interface supports this pipeline by enabling users to inspect detected biases, modify attributes or weights, and generate debiased images in real time. Experimental results show that LLMs outperform average human annotators in the number and granularity of detected bias categories and attributes. Moreover, FairT2I achieves superior performance to baseline models in both societal bias mitigation and image diversity, while preserving image quality and prompt fidelity.", "date": "2025-02-06", "categories": []}
{"id": "2601.03453", "title": "Measures of classification bias derived from sample size analysis", "abstract": "We propose the use of a simple intuitive principle for measuring algorithmic classification bias: the significance of the differences in a classifier's error rates across the various demographics is inversely commensurate with the sample size required to statistically detect them. That is, if large sample sizes are required to statistically establish biased behavior, the algorithm is less biased, and vice versa. In a simple setting, we assume two distinct demographics, and non-parametric estimates of the error rates on them, e1 and e2, respectively. We use a well-known approximate formula for the sample size of the chi-squared test, and verify some basic desirable properties of the proposed measure. Next, we compare the proposed measure with two other commonly used statistics, the difference e2-e1 and the ratio e2/e1 of the error rates. We establish that the proposed measure is essentially different in that it can rank algorithms for bias differently, and we discuss some of its advantages over the other two measures. Finally, we briefly discuss how some of the desirable properties of the proposed measure emanate from fundamental characteristics of the method, rather than the approximate sample size formula we used, and thus, are expected to hold in more complex settings with more than two demographics.", "date": "2026-01-06", "categories": []}
{"id": "2601.03032", "title": "Causal Manifold Fairness: Enforcing Geometric Invariance in Representation Learning", "abstract": "Fairness in machine learning is increasingly critical, yet standard approaches often treat data as static points in a high-dimensional space, ignoring the underlying generative structure. We posit that sensitive attributes (e.g., race, gender) do not merely shift data distributions but causally warp the geometry of the data manifold itself. To address this, we introduce Causal Manifold Fairness (CMF), a novel framework that bridges causal inference and geometric deep learning. CMF learns a latent representation where the local Riemannian geometry, defined by the metric tensor and curvature, remains invariant under counterfactual interventions on sensitive attributes. By enforcing constraints on the Jacobian and Hessian of the decoder, CMF ensures that the rules of the latent space (distances and shapes) are preserved across demographic groups. We validate CMF on synthetic Structural Causal Models (SCMs), demonstrating that it effectively disentangles sensitive geometric warping while preserving task utility, offering a rigorous quantification of the fairness-utility trade-off via geometric metrics.", "date": "2026-01-06", "categories": []}
{"id": "2601.03087", "title": "Audit Me If You Can: Query-Efficient Active Fairness Auditing of Black-Box LLMs", "abstract": "Large Language Models (LLMs) exhibit systematic biases across demographic groups. Auditing is proposed as an accountability tool for black-box LLM applications, but suffers from resource-intensive query access. We conceptualise auditing as uncertainty estimation over a target fairness metric and introduce BAFA, the Bounded Active Fairness Auditor for query-efficient auditing of black-box LLMs. BAFA maintains a version space of surrogate models consistent with queried scores and computes uncertainty intervals for fairness metrics (e.g., $\\Delta$ AUC) via constrained empirical risk minimisation. Active query selection narrows these intervals to reduce estimation error. We evaluate BAFA on two standard fairness dataset case studies: \\textsc{CivilComments} and \\textsc{Bias-in-Bios}, comparing against stratified sampling, power sampling, and ablations. BAFA achieves target error thresholds with up to 40$\\times$ fewer queries than stratified sampling (e.g., 144 vs 5,956 queries at $\\varepsilon=0.02$ for \\textsc{CivilComments}) for tight thresholds, demonstrates substantially better performance over time, and shows lower variance across runs. These results suggest that active sampling can reduce resources needed for independent fairness auditing with LLMs, supporting continuous model evaluations.", "date": "2026-01-06", "categories": []}
{"id": "2601.03090", "title": "LesionTABE: Equitable AI for Skin Lesion Detection", "abstract": "Bias remains a major barrier to the clinical adoption of AI in dermatology, as diagnostic models underperform on darker skin tones. We present LesionTABE, a fairness-centric framework that couples adversarial debiasing with dermatology-specific foundation model embeddings. Evaluated across multiple datasets covering both malignant and inflammatory conditions, LesionTABE achieves over a 25\\% improvement in fairness metrics compared to a ResNet-152 baseline, outperforming existing debiasing methods while simultaneously enhancing overall diagnostic accuracy. These results highlight the potential of foundation model debiasing as a step towards equitable clinical AI adoption.", "date": "2026-01-06", "categories": []}
{"id": "2601.03203", "title": "Counterfactual Fairness with Graph Uncertainty", "abstract": "Evaluating machine learning (ML) model bias is key to building trustworthy and robust ML systems. Counterfactual Fairness (CF) audits allow the measurement of bias of ML models with a causal framework, yet their conclusions rely on a single causal graph that is rarely known with certainty in real-world scenarios. We propose CF with Graph Uncertainty (CF-GU), a bias evaluation procedure that incorporates the uncertainty of specifying a causal graph into CF. CF-GU (i) bootstraps a Causal Discovery algorithm under domain knowledge constraints to produce a bag of plausible Directed Acyclic Graphs (DAGs), (ii) quantifies graph uncertainty with the normalized Shannon entropy, and (iii) provides confidence bounds on CF metrics. Experiments on synthetic data show how contrasting domain knowledge assumptions support or refute audits of CF, while experiments on real-world data (COMPAS and Adult datasets) pinpoint well-known biases with high confidence, even when supplied with minimal domain knowledge constraints.", "date": "2026-01-06", "categories": []}
{"id": "2505.09503", "title": "Towards Fair In-Context Learning with Tabular Foundation Models", "abstract": "Transformer-based tabular foundation models have recently demonstrated promising in-context learning (ICL) performance on structured data, emerging as competitive alternatives to gradient-boosted trees. However, the fairness implications of this new paradigm remain largely unexplored. We present the first investigation of fairness in tabular ICL, evaluating three recently proposed foundation models--TabPFNv2, TabICL, and TabDPT--on multiple benchmark datasets. To mitigate biases, we explore three pre-processing fairness-enhancing methods: correlation removal (decorrelating input features from the sensitive attribute), group-balanced sample selection (ensuring equal representation of protected groups in context examples), and uncertainty-based sample selection (prioritizing context examples with high sensitive-attribute prediction uncertainty). Our experiments show that the uncertainty-based strategy consistently improves group fairness metrics (e.g., demographic parity, equalized odds, and equal opportunity) with minimal impact on predictive accuracy. We release our code to facilitate reproducibility https://github.com/patrikken/Fair-TabICL.", "date": "2025-05-14", "categories": []}
{"id": "2601.00996", "title": "VEAT Quantifies Implicit Associations in Text-to-Video Generator Sora and Reveals Challenges in Bias Mitigation", "abstract": "Text-to-Video (T2V) generators such as Sora raise concerns about whether generated content reflects societal bias. We extend embedding-association tests from words and images to video by introducing the Video Embedding Association Test (VEAT) and Single-Category VEAT (SC-VEAT). We validate these methods by reproducing the direction and magnitude of associations from widely used baselines, including Implicit Association Test (IAT) scenarios and OASIS image categories. We then quantify race (African American vs. European American) and gender (women vs. men) associations with valence (pleasant vs. unpleasant) across 17 occupations and 7 awards. Sora videos associate European Americans and women more with pleasantness (both d>0.8). Effect sizes correlate with real-world demographic distributions: percent men and White in occupations (r=0.93, r=0.83) and percent male and non-Black among award recipients (r=0.88, r=0.99). Applying explicit debiasing prompts generally reduces effect-size magnitudes, but can backfire: two Black-associated occupations (janitor, postal service) become more Black-associated after debiasing. Together, these results reveal that easily accessible T2V generators can actually amplify representational harms if not rigorously evaluated and responsibly deployed.", "date": "2026-01-02", "categories": []}
{"id": "2507.05823", "title": "Fair Domain Generalization: An Information-Theoretic View", "abstract": "Domain generalization (DG) and algorithmic fairness are two critical challenges in machine learning. However, most DG methods focus only on minimizing expected risk in the unseen target domain without considering algorithmic fairness. Conversely, fairness methods typically do not account for domain shifts, so the fairness achieved during training may not generalize to unseen test domains. In this work, we bridge these gaps by studying the problem of Fair Domain Generalization (FairDG), which aims to minimize both expected risk and fairness violations in unseen target domains. We derive novel mutual information-based upper bounds for expected risk and fairness violations in multi-class classification tasks with multi-group sensitive attributes. These bounds provide key insights for algorithm design from an information-theoretic perspective. Guided by these insights, we introduce PAFDG (Pareto-Optimal Fairness for Domain Generalization), a practical framework that solves the FairDG problem and models the utility-fairness trade-off through Pareto optimization. Experiments on real-world vision and language datasets show that PAFDG achieves superior utility-fairness trade-offs compared to existing methods.", "date": "2025-07-08", "categories": []}
{"id": "2512.24747", "title": "Fairness-Aware Insurance Pricing: A Multi-Objective Optimization Approach", "abstract": "Machine learning improves predictive accuracy in insurance pricing but exacerbates trade-offs between competing fairness criteria across different discrimination measures, challenging regulators and insurers to reconcile profitability with equitable outcomes. While existing fairness-aware models offer partial solutions under GLM and XGBoost estimation methods, they remain constrained by single-objective optimization, failing to holistically navigate a conflicting landscape of accuracy, group fairness, individual fairness, and counterfactual fairness. To address this, we propose a novel multi-objective optimization framework that jointly optimizes all four criteria via the Non-dominated Sorting Genetic Algorithm II (NSGA-II), generating a diverse Pareto front of trade-off solutions. We use a specific selection mechanism to extract a premium on this front. Our results show that XGBoost outperforms GLM in accuracy but amplifies fairness disparities; the Orthogonal model excels in group fairness, while Synthetic Control leads in individual and counterfactual fairness. Our method consistently achieves a balanced compromise, outperforming single-model approaches.", "date": "2025-12-31", "categories": []}
{"id": "2410.01324", "title": "Fair Class-Incremental Learning using Sample Weighting", "abstract": "Model fairness is becoming important in class-incremental learning for Trustworthy AI. While accuracy has been a central focus in class-incremental learning, fairness has been relatively understudied. However, naively using all the samples of the current task for training results in unfair catastrophic forgetting for certain sensitive groups including classes. We theoretically analyze that forgetting occurs if the average gradient vector of the current task data is in an \"opposite direction\" compared to the average gradient vector of a sensitive group, which means their inner products are negative. We then propose a fair class-incremental learning framework that adjusts the training weights of current task samples to change the direction of the average gradient vector and thus reduce the forgetting of underperforming groups and achieve fairness. For various group fairness measures, we formulate optimization problems to minimize the overall losses of sensitive groups while minimizing the disparities among them. We also show the problems can be solved with linear programming and propose an efficient Fairness-aware Sample Weighting (FSW) algorithm. Experiments show that FSW achieves better accuracy-fairness tradeoff results than state-of-the-art approaches on real datasets.", "date": "2024-10-02", "categories": []}
{"id": "2512.22210", "title": "Toward Equitable Recovery: A Fairness-Aware AI Framework for Prioritizing Post-Flood Aid in Bangladesh", "abstract": "Post-disaster aid allocation in developing nations often suffers from systematic biases that disadvantage vulnerable regions, perpetuating historical inequities. This paper presents a fairness-aware artificial intelligence framework for prioritizing post-flood aid distribution in Bangladesh, a country highly susceptible to recurring flood disasters. Using real data from the 2022 Bangladesh floods that affected 7.2 million people and caused 405.5 million US dollars in damages, we develop an adversarial debiasing model that predicts flood vulnerability while actively removing biases against marginalized districts and rural areas. Our approach adapts fairness-aware representation learning techniques from healthcare AI to disaster management, employing a gradient reversal layer that forces the model to learn bias-invariant representations. Experimental results on 87 upazilas across 11 districts demonstrate that our framework reduces statistical parity difference by 41.6 percent, decreases regional fairness gaps by 43.2 percent, and maintains strong predictive accuracy (R-squared=0.784 vs baseline 0.811). The model generates actionable priority rankings ensuring aid reaches the most vulnerable populations based on genuine need rather than historical allocation patterns. This work demonstrates how algorithmic fairness techniques can be effectively applied to humanitarian contexts, providing decision-makers with tools to implement more equitable disaster recovery strategies.", "date": "2025-12-22", "categories": []}
{"id": "2512.22242", "title": "Fairness Evaluation of Risk Estimation Models for Lung Cancer Screening", "abstract": "Lung cancer is the leading cause of cancer-related mortality in adults worldwide. Screening high-risk individuals with annual low-dose CT (LDCT) can support earlier detection and reduce deaths, but widespread implementation may strain the already limited radiology workforce. AI models have shown potential in estimating lung cancer risk from LDCT scans. However, high-risk populations for lung cancer are diverse, and these models' performance across demographic groups remains an open question. In this study, we drew on the considerations on confounding factors and ethically significant biases outlined in the JustEFAB framework to evaluate potential performance disparities and fairness in two deep learning risk estimation models for lung cancer screening: the Sybil lung cancer risk model and the Venkadesh21 nodule risk estimator. We also examined disparities in the PanCan2b logistic regression model recommended in the British Thoracic Society nodule management guideline. Both deep learning models were trained on data from the US-based National Lung Screening Trial (NLST), and assessed on a held-out NLST validation set. We evaluated AUROC, sensitivity, and specificity across demographic subgroups, and explored potential confounding from clinical risk factors. We observed a statistically significant AUROC difference in Sybil's performance between women (0.88, 95% CI: 0.86, 0.90) and men (0.81, 95% CI: 0.78, 0.84, p < .001). At 90% specificity, Venkadesh21 showed lower sensitivity for Black (0.39, 95% CI: 0.23, 0.59) than White participants (0.69, 95% CI: 0.65, 0.73). These differences were not explained by available clinical confounders and thus may be classified as unfair biases according to JustEFAB. Our findings highlight the importance of improving and monitoring model performance across underrepresented subgroups, and further research on algorithmic fairness, in lung cancer screening.", "date": "2025-12-23", "categories": []}
{"id": "2512.22313", "title": "Optimistic Feasible Search for Closed-Loop Fair Threshold Decision-Making", "abstract": "Closed-loop decision-making systems (e.g., lending, screening, or recidivism risk assessment) often operate under fairness and service constraints while inducing feedback effects: decisions change who appears in the future, yielding non-stationary data and potentially amplifying disparities. We study online learning of a one-dimensional threshold policy from bandit feedback under demographic parity (DP) and, optionally, service-rate constraints. The learner observes only a scalar score each round and selects a threshold; reward and constraint residuals are revealed only for the chosen threshold.\n  We propose Optimistic Feasible Search (OFS), a simple grid-based method that maintains confidence bounds for reward and constraint residuals for each candidate threshold. At each round, OFS selects a threshold that appears feasible under confidence bounds and, among those, maximizes optimistic reward; if no threshold appears feasible, OFS selects the threshold minimizing optimistic constraint violation. This design directly targets feasible high-utility thresholds and is particularly effective for low-dimensional, interpretable policy classes where discretization is natural.\n  We evaluate OFS on (i) a synthetic closed-loop benchmark with stable contraction dynamics and (ii) two semi-synthetic closed-loop benchmarks grounded in German Credit and COMPAS, constructed by training a score model and feeding group-dependent acceptance decisions back into population composition. Across all environments, OFS achieves higher reward with smaller cumulative constraint violation than unconstrained and primal-dual bandit baselines, and is near-oracle relative to the best feasible fixed threshold under the same sweep procedure. Experiments are reproducible and organized with double-blind-friendly relative outputs.", "date": "2025-12-26", "categories": []}
{"id": "2511.13540", "title": "Fairness-Aware Graph Representation Learning with Limited Demographic Information", "abstract": "Ensuring fairness in Graph Neural Networks is fundamental to promoting trustworthy and socially responsible machine learning systems. In response, numerous fair graph learning methods have been proposed in recent years. However, most of them assume full access to demographic information, a requirement rarely met in practice due to privacy, legal, or regulatory restrictions. To this end, this paper introduces a novel fair graph learning framework that mitigates bias in graph learning under limited demographic information. Specifically, we propose a mechanism guided by partial demographic data to generate proxies for demographic information and design a strategy that enforces consistent node embeddings across demographic groups. In addition, we develop an adaptive confidence strategy that dynamically adjusts each node's contribution to fairness and utility based on prediction confidence. We further provide theoretical analysis demonstrating that our framework, FairGLite, achieves provable upper bounds on group fairness metrics, offering formal guarantees for bias mitigation. Through extensive experiments on multiple datasets and fair graph learning frameworks, we demonstrate the framework's effectiveness in both mitigating bias and maintaining model utility.", "date": "2025-11-17", "categories": []}
{"id": "2512.21348", "title": "Fairness Is Not Just Ethical: Performance Trade-Off via Data Correlation Tuning to Mitigate Bias in ML Software", "abstract": "Traditional software fairness research typically emphasizes ethical and social imperatives, neglecting that fairness fundamentally represents a core software quality issue arising directly from performance disparities across sensitive user groups. Recognizing fairness explicitly as a software quality dimension yields practical benefits beyond ethical considerations, notably improved predictive performance for unprivileged groups, enhanced out-of-distribution generalization, and increased geographic transferability in real-world deployments. Nevertheless, existing bias mitigation methods face a critical dilemma: while pre-processing methods offer broad applicability across model types, they generally fall short in effectiveness compared to post-processing techniques. To overcome this challenge, we propose Correlation Tuning (CoT), a novel pre-processing approach designed to mitigate bias by adjusting data correlations. Specifically, CoT introduces the Phi-coefficient, an intuitive correlation measure, to systematically quantify correlation between sensitive attributes and labels, and employs multi-objective optimization to address the proxy biases. Extensive evaluations demonstrate that CoT increases the true positive rate of unprivileged groups by an average of 17.5% and reduces three key bias metrics, including statistical parity difference (SPD), average odds difference (AOD), and equal opportunity difference (EOD), by more than 50% on average. CoT outperforms state-of-the-art methods by three and ten percentage points in single attribute and multiple attributes scenarios, respectively. We will publicly release our experimental results and source code to facilitate future research.", "date": "2025-12-19", "categories": []}
{"id": "2511.11767", "title": "Learning Fair Representations with Kolmogorov-Arnold Networks", "abstract": "Despite recent advances in fairness-aware machine learning, predictive models often exhibit discriminatory behavior towards marginalized groups. Such unfairness might arise from biased training data, model design, or representational disparities across groups, posing significant challenges in high-stakes decision-making domains such as college admissions. While existing fair learning models aim to mitigate bias, achieving an optimal trade-off between fairness and accuracy remains a challenge. Moreover, the reliance on black-box models hinders interpretability, limiting their applicability in socially sensitive domains. To circumvent these issues, we propose integrating Kolmogorov-Arnold Networks (KANs) within a fair adversarial learning framework. Leveraging the adversarial robustness and interpretability of KANs, our approach facilitates stable adversarial learning. We derive theoretical insights into the spline-based KAN architecture that ensure stability during adversarial optimization. Additionally, an adaptive fairness penalty update mechanism is proposed to strike a balance between fairness and accuracy. We back these findings with empirical evidence on two real-world admissions datasets, demonstrating the proposed framework's efficiency in achieving fairness across sensitive attributes while preserving predictive performance.", "date": "2025-11-14", "categories": []}
{"id": "2512.20233", "title": "How I Met Your Bias: Investigating Bias Amplification in Diffusion Models", "abstract": "Diffusion-based generative models demonstrate state-of-the-art performance across various image synthesis tasks, yet their tendency to replicate and amplify dataset biases remains poorly understood. Although previous research has viewed bias amplification as an inherent characteristic of diffusion models, this work provides the first analysis of how sampling algorithms and their hyperparameters influence bias amplification. We empirically demonstrate that samplers for diffusion models -- commonly optimized for sample quality and speed -- have a significant and measurable effect on bias amplification. Through controlled studies with models trained on Biased MNIST, Multi-Color MNIST and BFFHQ, and with Stable Diffusion, we show that sampling hyperparameters can induce both bias reduction and amplification, even when the trained model is fixed. Source code is available at https://github.com/How-I-met-your-bias/how_i_met_your_bias.", "date": "2025-12-23", "categories": []}
{"id": "2311.02757", "title": "Certified Defense on the Fairness of Graph Neural Networks", "abstract": "Graph Neural Networks (GNNs) have emerged as a prominent graph learning model in various graph-based tasks over the years. Nevertheless, due to the vulnerabilities of GNNs, it has been empirically shown that malicious attackers could easily corrupt the fairness level of their predictions by adding perturbations to the input graph data. In this paper, we take crucial steps to study a novel problem of certifiable defense on the fairness level of GNNs. Specifically, we propose a principled framework named ELEGANT and present a detailed theoretical certification analysis for the fairness of GNNs. ELEGANT takes {\\em any} GNN as its backbone, and the fairness level of such a backbone is theoretically impossible to be corrupted under certain perturbation budgets for attackers. Notably, ELEGANT does not make any assumptions over the GNN structure or parameters, and does not require re-training the GNNs to realize certification. Hence it can serve as a plug-and-play framework for any optimized GNNs ready to be deployed. We verify the satisfactory effectiveness of ELEGANT in practice through extensive experiments on real-world datasets across different backbones of GNNs and parameter settings.", "date": "2023-11-05", "categories": []}
{"id": "2407.16292", "title": "Visual Stereotypes of Autism Spectrum in Janus-Pro-7B, DALL-E, Stable Diffusion, SDXL, FLUX, and Midjourney", "abstract": "Avoiding systemic discrimination of neurodiverse individuals is an ongoing challenge in training AI models, which often propagate negative stereotypes. This study examined whether six text-to-image models (Janus-Pro-7B VL2 vs. VL3, DALL-E 3 v. April 2024 vs. August 2025, Stable Diffusion v. 1.6 vs. 3.5, SDXL v. April 2024 vs. FLUX.1 Pro, and Midjourney v. 5.1 vs. 7) perpetuate non-rational beliefs regarding autism by comparing images generated in 2024-2025 with controls. 53 prompts aimed at neutrally visualizing concrete objects and abstract concepts related to autism were used against 53 controls (baseline total N=302, follow-up experimental 280 images plus 265 controls). Expert assessment measuring the presence of common autism-related stereotypes employed a framework of 10 deductive codes followed by statistical analysis. Autistic individuals were depicted with striking homogeneity in skin color (white), gender (male), and age (young), often engaged in solitary activities, interacting with objects rather than people, and exhibiting stereotypical emotional expressions such as sadness, anger, or emotional flatness. In contrast, the images of neurotypical individuals were more diverse and lacked such traits. We found significant differences between the models; however, with a moderate effect size, and no differences between baseline and follow-up summary values, with the ratio of stereotypical themes to the number of images similar across all models. The control prompts showed a significantly lower degree of stereotyping with large size effects, confirming the hidden biases of the models. In summary, despite improvements in the technical aspects of image generation, the level of reproduction of potentially harmful autism-related stereotypes remained largely unaffected.", "date": "2024-07-23", "categories": []}
{"id": "2512.18180", "title": "FairExpand: Individual Fairness on Graphs with Partial Similarity Information", "abstract": "Individual fairness, which requires that similar individuals should be treated similarly by algorithmic systems, has become a central principle in fair machine learning. Individual fairness has garnered traction in graph representation learning due to its practical importance in high-stakes Web areas such as user modeling, recommender systems, and search. However, existing methods assume the existence of predefined similarity information over all node pairs, an often unrealistic requirement that prevents their operationalization in practice. In this paper, we assume the similarity information is only available for a limited subset of node pairs and introduce FairExpand, a flexible framework that promotes individual fairness in this more realistic partial information scenario. FairExpand follows a two-step pipeline that alternates between refining node representations using a backbone model (e.g., a graph neural network) and gradually propagating similarity information, which allows fairness enforcement to effectively expand to the entire graph. Extensive experiments show that FairExpand consistently enhances individual fairness while preserving performance, making it a practical solution for enabling graph-based individual fairness in real-world applications with partial similarity information.", "date": "2025-12-20", "categories": []}
{"id": "2512.19238", "title": "Identifying Features Associated with Bias Against 93 Stigmatized Groups in Language Models and Guardrail Model Safety Mitigation", "abstract": "Large language models (LLMs) have been shown to exhibit social bias, however, bias towards non-protected stigmatized identities remain understudied. Furthermore, what social features of stigmas are associated with bias in LLM outputs is unknown. From psychology literature, it has been shown that stigmas contain six shared social features: aesthetics, concealability, course, disruptiveness, origin, and peril. In this study, we investigate if human and LLM ratings of the features of stigmas, along with prompt style and type of stigma, have effect on bias towards stigmatized groups in LLM outputs. We measure bias against 93 stigmatized groups across three widely used LLMs (Granite 3.0-8B, Llama-3.1-8B, Mistral-7B) using SocialStigmaQA, a benchmark that includes 37 social scenarios about stigmatized identities; for example deciding wether to recommend them for an internship. We find that stigmas rated by humans to be highly perilous (e.g., being a gang member or having HIV) have the most biased outputs from SocialStigmaQA prompts (60% of outputs from all models) while sociodemographic stigmas (e.g. Asian-American or old age) have the least amount of biased outputs (11%). We test if the amount of biased outputs could be decreased by using guardrail models, models meant to identify harmful input, using each LLM's respective guardrail model (Granite Guardian 3.0, Llama Guard 3.0, Mistral Moderation API). We find that bias decreases significantly by 10.4%, 1.4%, and 7.8%, respectively. However, we show that features with significant effect on bias remain unchanged post-mitigation and that guardrail models often fail to recognize the intent of bias in prompts. This work has implications for using LLMs in scenarios involving stigmatized groups and we suggest future work towards improving guardrail models for bias mitigation.", "date": "2025-12-22", "categories": []}
{"id": "2412.00720", "title": "Fairness via Independence: A (Conditional) Distance Covariance Framework", "abstract": "We explore fairness from a statistical perspective by selectively utilizing either conditional distance covariance or distance covariance statistics as measures to assess the independence between predictions and sensitive attributes. We boost fairness with independence by adding a distance covariance-based penalty to the model's training. Additionally, we present the matrix form of empirical (conditional) distance covariance for parallel calculations to enhance computational efficiency. Theoretically, we provide a proof for the convergence between empirical and population (conditional) distance covariance, establishing necessary guarantees for batch computations. Through experiments conducted on a range of real-world datasets, we have demonstrated that our method effectively bridges the fairness gap in machine learning. Our code is available at \\url{https://github.com/liuhaixias1/Fair_dc/}.", "date": "2024-12-01", "categories": []}
{"id": "2511.04505", "title": "Alternative Fairness and Accuracy Optimization in Criminal Justice", "abstract": "Algorithmic fairness has grown rapidly as a research area, yet key concepts remain unsettled, especially in criminal justice. We review group, individual, and process fairness and map the conditions under which they conflict. We then develop a simple modification to standard group fairness. Rather than exact parity across protected groups, we minimize a weighted error loss while keeping differences in false negative rates within a small tolerance. This makes solutions easier to find, can raise predictive accuracy, and surfaces the ethical choice of error costs. We situate this proposal within three classes of critique: biased and incomplete data, latent affirmative action, and the explosion of subgroup constraints. Finally, we offer a practical framework for deployment in public decision systems built on three pillars: need-based decisions, Transparency and accountability, and narrowly tailored definitions and solutions. Together, these elements link technical design to legitimacy and provide actionable guidance for agencies that use risk assessment and related tools.", "date": "2025-11-06", "categories": []}
{"id": "2512.16386", "title": "Quantitative Verification of Fairness in Tree Ensembles", "abstract": "This work focuses on quantitative verification of fairness in tree ensembles. Unlike traditional verification approaches that merely return a single counterexample when the fairness is violated, quantitative verification estimates the ratio of all counterexamples and characterizes the regions where they occur, which is important information for diagnosing and mitigating bias. To date, quantitative verification has been explored almost exclusively for deep neural networks (DNNs). Representative methods, such as DeepGemini and FairQuant, all build on the core idea of Counterexample-Guided Abstraction Refinement, a generic framework that could be adapted to other model classes. We extended the framework into a model-agnostic form, but discovered two limitations: (i) it can provide only lower bounds, and (ii) its performance scales poorly. Exploiting the discrete structure of tree ensembles, our work proposes an efficient quantification technique that delivers any-time upper and lower bounds. Experiments on five widely used datasets demonstrate its effectiveness and efficiency. When applied to fairness testing, our quantification method significantly outperforms state-of-the-art testing techniques.", "date": "2025-12-18", "categories": []}
{"id": "2512.16433", "title": "Emergent Bias and Fairness in Multi-Agent Decision Systems", "abstract": "Multi-agent systems have demonstrated the ability to improve performance on a variety of predictive tasks by leveraging collaborative decision making. However, the lack of effective evaluation methodologies has made it difficult to estimate the risk of bias, making deployment of such systems unsafe in high stakes domains such as consumer finance, where biased decisions can translate directly into regulatory breaches and financial loss. To address this challenge, we need to develop fairness evaluation methodologies for multi-agent predictive systems and measure the fairness characteristics of these systems in the financial tabular domain. Examining fairness metrics using large-scale simulations across diverse multi-agent configurations, with varying communication and collaboration mechanisms, we reveal patterns of emergent bias in financial decision-making that cannot be traced to individual agent components, indicating that multi-agent systems may exhibit genuinely collective behaviors. Our findings highlight that fairness risks in financial multi-agent systems represent a significant component of model risk, with tangible impacts on tasks such as credit scoring and income estimation. We advocate that multi-agent decision systems must be evaluated as holistic entities rather than through reductionist analyses of their constituent components.", "date": "2025-12-18", "categories": []}
{"id": "2512.16532", "title": "From Personalization to Prejudice: Bias and Discrimination in Memory-Enhanced AI Agents for Recruitment", "abstract": "Large Language Models (LLMs) have empowered AI agents with advanced capabilities for understanding, reasoning, and interacting across diverse tasks. The addition of memory further enhances them by enabling continuity across interactions, learning from past experiences, and improving the relevance of actions and responses over time; termed as memory-enhanced personalization. Although such personalization through memory offers clear benefits, it also introduces risks of bias. While several previous studies have highlighted bias in ML and LLMs, bias due to memory-enhanced personalized agents is largely unexplored. Using recruitment as an example use case, we simulate the behavior of a memory-enhanced personalized agent, and study whether and how bias is introduced and amplified in and across various stages of operation. Our experiments on agents using safety-trained LLMs reveal that bias is systematically introduced and reinforced through personalization, emphasizing the need for additional protective measures or agent guardrails in memory-enhanced LLM-based AI agents.", "date": "2025-12-18", "categories": []}
{"id": "2512.16901", "title": "Impacts of Racial Bias in Historical Training Data for News AI", "abstract": "AI technologies have rapidly moved into business and research applications that involve large text corpora, including computational journalism research and newsroom settings. These models, trained on extant data from various sources, can be conceptualized as historical artifacts that encode decades-old attitudes and stereotypes. This paper investigates one such example trained on the broadly-used New York Times Annotated Corpus to create a multi-label classifier. Our use in research settings surfaced the concerning \"blacks\" thematic topic label. Through quantitative and qualitative means we investigate this label's use in the training corpus, what concepts it might be encoding in the trained classifier, and how those concepts impact our model use. Via the application of explainable AI methods, we find that the \"blacks\" label operates partially as a general \"racism detector\" across some minoritized groups. However, it performs poorly against expectations on modern examples such as COVID-19 era anti-Asian hate stories, and reporting on the Black Lives Matter movement. This case study of interrogating embedded biases in a model reveals how similar applications in newsroom settings can lead to unexpected outputs that could impact a wide variety of potential uses of any large language model-story discovery, audience targeting, summarization, etc. The fundamental tension this exposes for newsrooms is how to adopt AI-enabled workflow tools while reducing the risk of reproducing historical biases in news coverage.", "date": "2025-12-18", "categories": []}
{"id": "2512.14711", "title": "Promoting Fairness in Information Access within Social Networks", "abstract": "The advent of online social networks has facilitated fast and wide spread of information. However, some users, especially members of minority groups, may be less likely to receive information spreading on the network, due to their disadvantaged network position. We study the optimization problem of adding new connections to a network to enhance fairness in information access among different demographic groups.\n  We provide a concrete formulation of this problem where information access is measured in terms of resistance distance, {offering a new perspective that emphasizes global network structure and multi-path connectivity.} The problem is shown to be NP-hard. We propose a simple greedy algorithm which turns out to output accurate solutions, but its run time is cubic, which makes it undesirable for large networks. As our main technical contribution, we reduce its time complexity to linear, leveraging several novel approximation techniques. In addition to our theoretical findings, we also conduct an extensive set of experiments using both real-world and synthetic datasets. We demonstrate that our linear-time algorithm can produce accurate solutions for networks with millions of nodes.", "date": "2025-12-08", "categories": []}
{"id": "2506.01594", "title": "Selecting for Less Discriminatory Algorithms: A Relational Search Framework for Navigating Fairness-Accuracy Trade-offs in Practice", "abstract": "As machine learning models are increasingly embedded into society through high-stakes decision-making, selecting the right algorithm for a given task, audience, and sector presents a critical challenge, particularly in the context of fairness. Traditional assessments of model fairness have often framed fairness as an objective mathematical property, treating model selection as an optimization problem under idealized informational conditions. This overlooks model multiplicity as a consideration--that multiple models can deliver similar performance while exhibiting different fairness characteristics. Legal scholars have engaged this challenge through the concept of Less Discriminatory Algorithms (LDAs), which frames model selection as a civil rights obligation. In real-world deployment, this normative challenge is bounded by constraints on fairness experimentation, e.g., regulatory standards, institutional priorities, and resource capacity.\n  Against these considerations, the paper revisits Lee and Floridi (2021)'s relational fairness approach using updated 2021 Home Mortgage Disclosure Act (HMDA) data, and proposes an expansion of the scope of the LDA search process. We argue that extending the LDA search horizontally, considering fairness across model families themselves, provides a lightweight complement, or alternative, to within-model hyperparameter optimization, when operationalizing fairness in non-experimental, resource constrained settings. Fairness metrics alone offer useful, but insufficient signals to accurately evaluate candidate LDAs. Rather, by using a horizontal LDA search approach with the relational trade-off framework, we demonstrate a responsible minimum viable LDA search on real-world lending outcomes. Organizations can modify this approach to systematically compare, evaluate, and select LDAs that optimize fairness and accuracy in a sector-based contextualized manner.", "date": "2025-06-02", "categories": []}
{"id": "2512.13427", "title": "MineTheGap: Automatic Mining of Biases in Text-to-Image Models", "abstract": "Text-to-Image (TTI) models generate images based on text prompts, which often leave certain aspects of the desired image ambiguous. When faced with these ambiguities, TTI models have been shown to exhibit biases in their interpretations. These biases can have societal impacts, e.g., when showing only a certain race for a stated occupation. They can also affect user experience when creating redundancy within a set of generated images instead of spanning diverse possibilities. Here, we introduce MineTheGap - a method for automatically mining prompts that cause a TTI model to generate biased outputs. Our method goes beyond merely detecting bias for a given prompt. Rather, it leverages a genetic algorithm to iteratively refine a pool of prompts, seeking for those that expose biases. This optimization process is driven by a novel bias score, which ranks biases according to their severity, as we validate on a dataset with known biases. For a given prompt, this score is obtained by comparing the distribution of generated images to the distribution of LLM-generated texts that constitute variations on the prompt. Code and examples are available on the project's webpage.", "date": "2025-12-15", "categories": []}
{"id": "2403.15780", "title": "A Fairness-Oriented Reinforcement Learning Approach for the Operation and Control of Shared Micromobility Services", "abstract": "As Machine Learning grows in popularity across various fields, equity has become a key focus for the AI community. However, fairness-oriented approaches are still underexplored in smart mobility. Addressing this gap, our study investigates the balance between performance optimization and algorithmic fairness in shared micromobility services providing a novel framework based on Reinforcement Learning. Exploiting Q-learning, the proposed methodology achieves equitable outcomes in terms of the Gini index across different areas characterized by their distance from central hubs. Through vehicle rebalancing, the provided scheme maximizes operator performance while ensuring fairness principles for users, reducing iniquity by up to 85% while only increasing costs by 30% (w.r.t. applying no equity adjustment). A case study with synthetic data validates our insights and highlights the importance of fairness in urban micromobility (source code: https://github.com/mcederle99/FairMSS.git).", "date": "2024-03-23", "categories": []}
{"id": "2407.21202", "title": "Rolling in the deep of cognitive and AI biases", "abstract": "Nowadays, we delegate many of our decisions to Artificial Intelligence (AI) that acts either in solo or as a human companion in decisions made to support several sensitive domains, like healthcare, financial services and law enforcement. AI systems, even carefully designed to be fair, are heavily criticized for delivering misjudged and discriminated outcomes against individuals and groups. Numerous work on AI algorithmic fairness is devoted on Machine Learning pipelines which address biases and quantify fairness under a pure computational view. However, the continuous unfair and unjust AI outcomes, indicate that there is urgent need to understand AI as a sociotechnical system, inseparable from the conditions in which it is designed, developed and deployed. Although, the synergy of humans and machines seems imperative to make AI work, the significant impact of human and societal factors on AI bias is currently overlooked. We address this critical issue by following a radical new methodology under which human cognitive biases become core entities in our AI fairness overview. Inspired by the cognitive science definition and taxonomy of human heuristics, we identify how harmful human actions influence the overall AI lifecycle, and reveal human to AI biases hidden pathways. We introduce a new mapping, which justifies the human heuristics to AI biases reflections and we detect relevant fairness intensities and inter-dependencies. We envision that this approach will contribute in revisiting AI fairness under deeper human-centric case studies, revealing hidden biases cause and effects.", "date": "2024-07-30", "categories": []}
{"id": "2512.08055", "title": "Fairness-aware PageRank via Edge Reweighting", "abstract": "Link-analysis algorithms, such as PageRank, are instrumental in understanding the structural dynamics of networks by evaluating the importance of individual vertices based on their connectivity. Recently, with the rising importance of responsible AI, the question of fairness in link-analysis algorithms has gained traction. In this paper, we present a new approach for incorporating group fairness into the PageRank algorithm by reweighting the transition probabilities in the underlying transition matrix. We formulate the problem of achieving fair PageRank by seeking to minimize the fairness loss, which is the difference between the original group-wise PageRank distribution and a target PageRank distribution. We further define a group-adapted fairness notion, which accounts for group homophily by considering random walks with group-biased restart for each group. Since the fairness loss is non-convex, we propose an efficient projected gradient-descent method for computing locally-optimal edge weights. Unlike earlier approaches, we do not recommend adding new edges to the network, nor do we adjust the restart vector. Instead, we keep the topology of the underlying network unchanged and only modify the relative importance of existing edges. We empirically compare our approach with state-of-the-art baselines and demonstrate the efficacy of our method, where very small changes in the transition matrix lead to significant improvement in the fairness of the PageRank algorithm.", "date": "2025-12-08", "categories": []}
{"id": "2512.11015", "title": "Leveraging Text Guidance for Enhancing Demographic Fairness in Gender Classification", "abstract": "In the quest for fairness in artificial intelligence, novel approaches to enhance it in facial image based gender classification algorithms using text guided methodologies are presented. The core methodology involves leveraging semantic information from image captions during model training to improve generalization capabilities. Two key strategies are presented: Image Text Matching (ITM) guidance and Image Text fusion. ITM guidance trains the model to discern fine grained alignments between images and texts to obtain enhanced multimodal representations. Image text fusion combines both modalities into comprehensive representations for improved fairness. Exensive experiments conducted on benchmark datasets demonstrate these approaches effectively mitigate bias and improve accuracy across gender racial groups compared to existing methods. Additionally, the unique integration of textual guidance underscores an interpretable and intuitive training paradigm for computer vision systems. By scrutinizing the extent to which semantic information reduces disparities, this research offers valuable insights into cultivating more equitable facial analysis algorithms. The proposed methodologies contribute to addressing the pivotal challenge of demographic bias in gender classification from facial images. Furthermore, this technique operates in the absence of demographic labels and is application agnostic.", "date": "2025-12-11", "categories": []}
{"id": "2512.11131", "title": "Fairness-Regularized Online Optimization with Switching Costs", "abstract": "Fairness and action smoothness are two crucial considerations in many online optimization problems, but they have yet to be addressed simultaneously. In this paper, we study a new and challenging setting of fairness-regularized smoothed online convex optimization with switching costs. First, to highlight the fundamental challenges introduced by the long-term fairness regularizer evaluated based on the entire sequence of actions, we prove that even without switching costs, no online algorithms can possibly achieve a sublinear regret or finite competitive ratio compared to the offline optimal algorithm as the problem episode length $T$ increases. Then, we propose FairOBD (Fairness-regularized Online Balanced Descent), which reconciles the tension between minimizing the hitting cost, switching cost, and fairness cost. Concretely, FairOBD decomposes the long-term fairness cost into a sequence of online costs by introducing an auxiliary variable and then leverages the auxiliary variable to regularize the online actions for fair outcomes. Based on a new approach to account for switching costs, we prove that FairOBD offers a worst-case asymptotic competitive ratio against a novel benchmark -- the optimal offline algorithm with parameterized constraints -- by considering $T\\to\\infty$. Finally, we run trace-driven experiments of dynamic computing resource provisioning for socially responsible AI inference to empirically evaluate FairOBD, showing that FairOBD can effectively reduce the total fairness-regularized cost and better promote fair outcomes compared to existing baseline solutions.", "date": "2025-12-11", "categories": []}
{"id": "2512.08981", "title": "Mitigating Bias with Words: Inducing Demographic Ambiguity in Face Recognition Templates by Text Encoding", "abstract": "Face recognition (FR) systems are often prone to demographic biases, partially due to the entanglement of demographic-specific information with identity-relevant features in facial embeddings. This bias is extremely critical in large multicultural cities, especially where biometrics play a major role in smart city infrastructure. The entanglement can cause demographic attributes to overshadow identity cues in the embedding space, resulting in disparities in verification performance across different demographic groups. To address this issue, we propose a novel strategy, Unified Text-Image Embedding (UTIE), which aims to induce demographic ambiguity in face embeddings by enriching them with information related to other demographic groups. This encourages face embeddings to emphasize identity-relevant features and thus promotes fairer verification performance across groups. UTIE leverages the zero-shot capabilities and cross-modal semantic alignment of Vision-Language Models (VLMs). Given that VLMs are naturally trained to align visual and textual representations, we enrich the facial embeddings of each demographic group with text-derived demographic features extracted from other demographic groups. This encourages a more neutral representation in terms of demographic attributes. We evaluate UTIE using three VLMs, CLIP, OpenCLIP, and SigLIP, on two widely used benchmarks, RFW and BFW, designed to assess bias in FR. Experimental results show that UTIE consistently reduces bias metrics while maintaining, or even improving in several cases, the face verification accuracy.", "date": "2025-12-05", "categories": []}
{"id": "2512.09467", "title": "Cauchy-Schwarz Fairness Regularizer", "abstract": "Group fairness in machine learning is often enforced by adding a regularizer that reduces the dependence between model predictions and sensitive attributes. However, existing regularizers are built on heterogeneous distance measures and design choices, which makes their behavior hard to reason about and their performance inconsistent across tasks. This raises a basic question: what properties make a good fairness regularizer? We address this question by first organizing existing in-process methods into three families: (i) matching prediction statistics across sensitive groups, (ii) aligning latent representations, and (iii) directly minimizing dependence between predictions and sensitive attributes. Through this lens, we identify desirable properties of the underlying distance measure, including tight generalization bounds, robustness to scale differences, and the ability to handle arbitrary prediction distributions. Motivated by these properties, we propose a Cauchy-Schwarz (CS) fairness regularizer that penalizes the empirical CS divergence between prediction distributions conditioned on sensitive groups. Under a Gaussian comparison, we show that CS divergence yields a tighter bound than Kullback-Leibler divergence, Maximum Mean Discrepancy, and the mean disparity used in Demographic Parity, and we discuss how these advantages translate to a distribution-free, kernel-based estimator that naturally extends to multiple sensitive attributes. Extensive experiments on four tabular benchmarks and one image dataset demonstrate that the proposed CS regularizer consistently improves Demographic Parity and Equal Opportunity metrics while maintaining competitive accuracy, and achieves a more stable utility-fairness trade-off across hyperparameter settings compared to prior regularizers.", "date": "2025-12-10", "categories": []}
{"id": "2512.07990", "title": "A Gray Literature Study on Fairness Requirements in AI-enabled Software Engineering", "abstract": "Today, with the growing obsession with applying Artificial Intelligence (AI), particularly Machine Learning (ML), to software across various contexts, much of the focus has been on the effectiveness of AI models, often measured through common metrics such as F1- score, while fairness receives relatively little attention. This paper presents a review of existing gray literature, examining fairness requirements in AI context, with a focus on how they are defined across various application domains, managed throughout the Software Development Life Cycle (SDLC), and the causes, as well as the corresponding consequences of their violation by AI models. Our gray literature investigation shows various definitions of fairness requirements in AI systems, commonly emphasizing non-discrimination and equal treatment across different demographic and social attributes. Fairness requirement management practices vary across the SDLC, particularly in model training and bias mitigation, fairness monitoring and evaluation, and data handling practices. Fairness requirement violations are frequently linked, but not limited, to data representation bias, algorithmic and model design bias, human judgment, and evaluation and transparency gaps. The corresponding consequences include harm in a broad sense, encompassing specific professional and societal impacts as key examples, stereotype reinforcement, data and privacy risks, and loss of trust and legitimacy in AI-supported decisions. These findings emphasize the need for consistent frameworks and practices to integrate fairness into AI software, paying as much attention to fairness as to effectiveness.", "date": "2025-12-08", "categories": []}
{"id": "2512.08733", "title": "Mitigating Individual Skin Tone Bias in Skin Lesion Classification through Distribution-Aware Reweighting", "abstract": "Skin color has historically been a focal point of discrimination, yet fairness research in machine learning for medical imaging often relies on coarse subgroup categories, overlooking individual-level variations. Such group-based approaches risk obscuring biases faced by outliers within subgroups. This study introduces a distribution-based framework for evaluating and mitigating individual fairness in skin lesion classification. We treat skin tone as a continuous attribute rather than a categorical label, and employ kernel density estimation (KDE) to model its distribution. We further compare twelve statistical distance metrics to quantify disparities between skin tone distributions and propose a distance-based reweighting (DRW) loss function to correct underrepresentation in minority tones. Experiments across CNN and Transformer models demonstrate: (i) the limitations of categorical reweighting in capturing individual-level disparities, and (ii) the superior performance of distribution-based reweighting, particularly with Fidelity Similarity (FS), Wasserstein Distance (WD), Hellinger Metric (HM), and Harmonic Mean Similarity (HS). These findings establish a robust methodology for advancing fairness at individual level in dermatological AI systems, and highlight broader implications for sensitive continuous attributes in medical image analysis.", "date": "2025-12-09", "categories": []}
{"id": "2410.02799", "title": "A Data Envelopment Analysis Approach for Assessing Fairness in Resource Allocation: Application to Kidney Exchange Programs", "abstract": "Kidney exchange programs have substantially increased transplantation rates but also raise critical concerns about fairness in organ allocation. We propose a novel framework leveraging Data Envelopment Analysis (DEA) to evaluate multiple dimensions of fairness-Priority, Access, and Outcome-within a unified model. This approach captures complexities often missed in single-metric analyses. Using data from the United Network for Organ Sharing, we separately quantify fairness across these dimensions: Priority fairness through waitlist durations, Access fairness via the Living Kidney Donor Profile Index (LKDPI) scores, and Outcome fairness based on graft lifespan. We then apply our conditional DEA model with covariate adjustment to demonstrate significant disparities in kidney allocation efficiency across ethnic groups. To quantify uncertainty, we employ conformal prediction within a novel Reference Frontier Mapping (RFM) framework, yielding group-conditional prediction intervals with finite-sample coverage guarantees. Our findings show notable differences in efficiency distributions between ethnic groups. Our study provides a rigorous framework for evaluating fairness in complex resource allocation systems with resource scarcity and mutual compatibility constraints.", "date": "2024-09-18", "categories": []}
{"id": "2512.06944", "title": "A Unifying Human-Centered AI Fairness Framework", "abstract": "The increasing use of Artificial Intelligence (AI) in critical societal domains has amplified concerns about fairness, particularly regarding unequal treatment across sensitive attributes such as race, gender, and socioeconomic status. While there has been substantial work on ensuring AI fairness, navigating trade-offs between competing notions of fairness as well as predictive accuracy remains challenging, creating barriers to the practical deployment of fair AI systems. To address this, we introduce a unifying human-centered fairness framework that systematically covers eight distinct fairness metrics, formed by combining individual and group fairness, infra-marginal and intersectional assumptions, and outcome-based and equality-of-opportunity (EOO) perspectives. This structure allows stakeholders to align fairness interventions with their values and contextual considerations. The framework uses a consistent and easy-to-understand formulation for all metrics to reduce the learning curve for non-experts. Rather than privileging a single fairness notion, the framework enables stakeholders to assign weights across multiple fairness objectives, reflecting their priorities and facilitating multi-stakeholder compromises. We apply this approach to four real-world datasets: the UCI Adult census dataset for income prediction, the COMPAS dataset for criminal recidivism, the German Credit dataset for credit risk assessment, and the MEPS dataset for healthcare utilization. We show that adjusting weights reveals nuanced trade-offs between different fairness metrics. Finally, through case studies in judicial decision-making and healthcare, we demonstrate how the framework can inform practical and value-sensitive deployment of fair AI systems.", "date": "2025-12-07", "categories": []}
{"id": "2512.07249", "title": "IFFair: Influence Function-driven Sample Reweighting for Fair Classification", "abstract": "Because machine learning has significantly improved efficiency and convenience in the society, it's increasingly used to assist or replace human decision-making. However, the data-based pattern makes related algorithms learn and even exacerbate potential bias in samples, resulting in discriminatory decisions against certain unprivileged groups, depriving them of the rights to equal treatment, thus damaging the social well-being and hindering the development of related applications. Therefore, we propose a pre-processing method IFFair based on the influence function. Compared with other fairness optimization approaches, IFFair only uses the influence disparity of training samples on different groups as a guidance to dynamically adjust the sample weights during training without modifying the network structure, data features and decision boundaries. To evaluate the validity of IFFair, we conduct experiments on multiple real-world datasets and metrics. The experimental results show that our approach mitigates bias of multiple accepted metrics in the classification setting, including demographic parity, equalized odds, equality of opportunity and error rate parity without conflicts. It also demonstrates that IFFair achieves better trade-off between multiple utility and fairness metrics compared with previous pre-processing methods.", "date": "2025-12-08", "categories": []}
{"id": "2512.07433", "title": "Mitigating Bias in Graph Hyperdimensional Computing", "abstract": "Graph hyperdimensional computing (HDC) has emerged as a promising paradigm for cognitive tasks, emulating brain-like computation with high-dimensional vectors known as hypervectors. While HDC offers robustness and efficiency on graph-structured data, its fairness implications remain largely unexplored. In this paper, we study fairness in graph HDC, where biases in data representation and decision rules can lead to unequal treatment of different groups. We show how hypervector encoding and similarity-based classification can propagate or even amplify such biases, and we propose a fairness-aware training framework, FairGHDC, to mitigate them. FairGHDC introduces a bias correction term, derived from a gap-based demographic-parity regularizer, and converts it into a scalar fairness factor that scales the update of the class hypervector for the ground-truth label. This enables debiasing directly in the hypervector space without modifying the graph encoder or requiring backpropagation. Experimental results on six benchmark datasets demonstrate that FairGHDC substantially reduces demographic-parity and equal-opportunity gaps while maintaining accuracy comparable to standard GNNs and fairness-aware GNNs. At the same time, FairGHDC preserves the computational advantages of HDC, achieving up to about one order of magnitude ($\\approx 10\\times$) speedup in training time on GPU compared to GNN and fairness-aware GNN baselines.", "date": "2025-12-08", "categories": []}
{"id": "2512.07608", "title": "Metric-Fair Prompting: Treating Similar Samples Similarly", "abstract": "We introduce \\emph{Metric-Fair Prompting}, a fairness-aware prompting framework that guides large language models (LLMs) to make decisions under metric-fairness constraints. In the application of multiple-choice medical question answering, each {(question, option)} pair is treated as a binary instance with label $+1$ (correct) or $-1$ (incorrect). To promote {individual fairness}~--~treating similar instances similarly~--~we compute question similarity using NLP embeddings and solve items in \\emph{joint pairs of similar questions} rather than in isolation. The prompt enforces a global decision protocol: extract decisive clinical features, map each \\((\\text{question}, \\text{option})\\) to a score $f(x)$ that acts as confidence, and impose a Lipschitz-style constraint so that similar inputs receive similar scores and, hence, consistent outputs. Evaluated on the {MedQA (US)} benchmark, Metric-Fair Prompting is shown to improve performance over standard single-item prompting, demonstrating that fairness-guided, confidence-oriented reasoning can enhance LLM accuracy on high-stakes clinical multiple-choice questions.", "date": "2025-12-08", "categories": []}
{"id": "2503.07691", "title": "Fair Text Classification via Transferable Representations", "abstract": "Group fairness is a central research topic in text classification, where reaching fair treatment between sensitive groups (e.g., women and men) remains an open challenge. We propose an approach that extends the use of the Wasserstein Dependency Measure for learning unbiased neural text classifiers. Given the challenge of distinguishing fair from unfair information in a text encoder, we draw inspiration from adversarial training by inducing independence between representations learned for the target label and those for a sensitive attribute. We further show that Domain Adaptation can be efficiently leveraged to remove the need for access to the sensitive attributes in the dataset we cure. We provide both theoretical and empirical evidence that our approach is well-founded.", "date": "2025-03-10", "categories": []}
{"id": "2312.15478", "title": "A Group Fairness Lens for Large Language Models", "abstract": "The need to assess LLMs for bias and fairness is critical, with current evaluations often being narrow, missing a broad categorical view. In this paper, we propose evaluating the bias and fairness of LLMs from a group fairness lens using a novel hierarchical schema characterizing diverse social groups. Specifically, we construct a dataset, GFAIR, encapsulating target-attribute combinations across multiple dimensions. Moreover, we introduce statement organization, a new open-ended text generation task, to uncover complex biases in LLMs. Extensive evaluations of popular LLMs reveal inherent safety concerns. To mitigate the biases of LLMs from a group fairness perspective, we pioneer a novel chainof-thought method GF-THINK to mitigate biases of LLMs from a group fairness perspective. Experimental results demonstrate its efficacy in mitigating bias and achieving fairness in LLMs. Our dataset and codes are available at https://github.com/surika/Group-Fairness-LLMs.", "date": "2023-12-24", "categories": []}
{"id": "2407.14766", "title": "Fairness Interventions: A Study in AI Explainability", "abstract": "This paper presents a philosophical and experimental study of fairness interventions in AI classification, centered on the explainability of corrective methods. We argue that ensuring fairness requires not only satisfying a target criterion, but also explaining which variables constrain its realization. When corrections are used to mitigate advantage transparently, they must remain sensitive to the distribution of true labels. To illustrate this approach, we built FairDream, a fairness package whose mechanism is made transparent for lay users, increasing the model's weights of errors on disadvantaged groups. While a user may intend to achieve Demographic Parity by the correction method, experiments show that FairDream tends towards Equalized Odds, revealing a conservative bias inherent to the data environment. We clarify the relationship between these fairness criteria, analyze FairDream's reweighting process, and compare its trade-offs with closely related GridSearch models. Finally, we justify the normative preference for Equalized Odds via an epistemological interpretation of the results, using their proximity with Simpson's paradox. The paper thus unites normative, epistemological, and empirical explanations of fairness interventions, to ensure transparency for the users.", "date": "2024-07-20", "categories": []}
{"id": "2512.03074", "title": "Model-Agnostic Fairness Regularization for GNNs with Incomplete Sensitive Information", "abstract": "Graph Neural Networks (GNNs) have demonstrated exceptional efficacy in relational learning tasks, including node classification and link prediction. However, their application raises significant fairness concerns, as GNNs can perpetuate and even amplify societal biases against protected groups defined by sensitive attributes such as race or gender. These biases are often inherent in the node features, structural topology, and message-passing mechanisms of the graph itself. A critical limitation of existing fairness-aware GNN methods is their reliance on the strong assumption that sensitive attributes are fully available for all nodes during training--a condition that poses a practical impediment due to privacy concerns and data collection constraints. To address this gap, we propose a novel, model-agnostic fairness regularization framework designed for the realistic scenario where sensitive attributes are only partially available. Our approach formalizes a fairness-aware objective function that integrates both equal opportunity and statistical parity as differentiable regularization terms. Through a comprehensive empirical evaluation across five real-world benchmark datasets, we demonstrate that the proposed method significantly mitigates bias across key fairness metrics while maintaining competitive node classification performance. Results show that our framework consistently outperforms baseline models in achieving a favorable fairness-accuracy trade-off, with minimal degradation in predictive accuracy. The datasets and source code will be publicly released at https://github.com/mtavassoli/GNN-FC.", "date": "2025-11-27", "categories": []}
{"id": "2508.03079", "title": "Bias Beyond Demographics: Probing Decision Boundaries in Black-Box LVLMs via Counterfactual VQA", "abstract": "Recent advances in large vision-language models (LVLMs) have amplified concerns about fairness, yet existing evaluations remain confined to demographic attributes and often conflate fairness with refusal behavior. This paper broadens the scope of fairness by introducing a counterfactual VQA benchmark that probes the decision boundaries of closed-source LVLMs under controlled context shifts. Each image pair differs in a single visual attribute that has been validated as irrelevant to the question, enabling ground-truth-free and refusal-aware analysis of reasoning stability. Comprehensive experiments reveal that non-demographic attributes, such as environmental context or social behavior, distort LVLM decision-making more strongly than demographic ones. Moreover, instruction-based debiasing shows limited effectiveness and can even amplify these asymmetries, whereas exposure to a small number of human norm validated examples from our benchmark encourages more consistent and balanced responses, highlighting its potential not only as an evaluative framework but also as a means for understanding and improving model behavior. Together, these results provide a practial basis for auditing contextual biases even in black-box LVLMs and contribute to more transparent and equitable multimodal reasoning.", "date": "2025-08-05", "categories": []}
{"id": "2512.02265", "title": "The Effect of Enforcing Fairness on Reshaping Explanations in Machine Learning Models", "abstract": "Trustworthy machine learning in healthcare requires strong predictive performance, fairness, and explanations. While it is known that improving fairness can affect predictive performance, little is known about how fairness improvements influence explainability, an essential ingredient for clinical trust. Clinicians may hesitate to rely on a model whose explanations shift after fairness constraints are applied. In this study, we examine how enhancing fairness through bias mitigation techniques reshapes Shapley-based feature rankings. We quantify changes in feature importance rankings after applying fairness constraints across three datasets: pediatric urinary tract infection risk, direct anticoagulant bleeding risk, and recidivism risk. We also evaluate multiple model classes on the stability of Shapley-based rankings. We find that increasing model fairness across racial subgroups can significantly alter feature importance rankings, sometimes in different ways across groups. These results highlight the need to jointly consider accuracy, fairness, and explainability in model assessment rather than in isolation.", "date": "2025-12-01", "categories": []}
{"id": "2512.02689", "title": "An Empirical Survey of Model Merging Algorithms for Social Bias Mitigation", "abstract": "Large language models (LLMs) are known to inherit and even amplify societal biases present in their pre-training corpora, threatening fairness and social trust. To address this issue, recent work has explored ``editing'' LLM parameters to mitigate social bias with model merging approaches; however, there is no empirical comparison. In this work, we empirically survey seven algorithms: Linear, Karcher Mean, SLERP, NuSLERP, TIES, DELLA, and Nearswap, applying 13 open weight models in the GPT, LLaMA, and Qwen families. We perform a comprehensive evaluation using three bias datasets (BBQ, BOLD, and HONEST) and measure the impact of these techniques on LLM performance in downstream tasks of the SuperGLUE benchmark. We find a trade-off between bias reduction and downstream performance: methods achieving greater bias mitigation degrade accuracy, particularly on tasks requiring reading comprehension and commonsense and causal reasoning. Among the merging algorithms, Linear, SLERP, and Nearswap consistently reduce bias while maintaining overall performance, with SLERP at moderate interpolation weights emerging as the most balanced choice. These results highlight the potential of model merging algorithms for bias mitigation, while indicating that excessive debiasing or inappropriate merging methods may lead to the degradation of important linguistic abilities.", "date": "2025-12-02", "categories": []}
{"id": "2511.08711", "title": "Harnessing Diffusion-Generated Synthetic Images for Fair Image Classification", "abstract": "Image classification systems often inherit biases from uneven group representation in training data. For example, in face datasets for hair color classification, blond hair may be disproportionately associated with females, reinforcing stereotypes. A recent approach leverages the Stable Diffusion model to generate balanced training data, but these models often struggle to preserve the original data distribution. In this work, we explore multiple diffusion-finetuning techniques, e.g., LoRA and DreamBooth, to generate images that more accurately represent each training group by learning directly from their samples. Additionally, in order to prevent a single DreamBooth model from being overwhelmed by excessive intra-group variations, we explore a technique of clustering images within each group and train a DreamBooth model per cluster. These models are then used to generate group-balanced data for pretraining, followed by fine-tuning on real data. Experiments on multiple benchmarks demonstrate that the studied finetuning approaches outperform vanilla Stable Diffusion on average and achieve results comparable to SOTA debiasing techniques like Group-DRO, while surpassing them as the dataset bias severity increases.", "date": "2025-11-11", "categories": []}
{"id": "2512.00469", "title": "FairMT: Fairness for Heterogeneous Multi-Task Learning", "abstract": "Fairness in machine learning has been extensively studied in single-task settings, while fair multi-task learning (MTL), especially with heterogeneous tasks (classification, detection, regression) and partially missing labels, remains largely unexplored. Existing fairness methods are predominantly classification-oriented and fail to extend to continuous outputs, making a unified fairness objective difficult to formulate. Further, existing MTL optimization is structurally misaligned with fairness: constraining only the shared representation, allowing task heads to absorb bias and leading to uncontrolled task-specific disparities. Finally, most work treats fairness as a zero-sum trade-off with utility, enforcing symmetric constraints that achieve parity by degrading well-served groups. We introduce FairMT, a unified fairness-aware MTL framework that accommodates all three task types under incomplete supervision. At its core is an Asymmetric Heterogeneous Fairness Constraint Aggregation mechanism, which consolidates task-dependent asymmetric violations into a unified fairness constraint. Utility and fairness are jointly optimized via a primal--dual formulation, while a head-aware multi-objective optimization proxy provides a tractable descent geometry that explicitly accounts for head-induced anisotropy. Across three homogeneous and heterogeneous MTL benchmarks encompassing diverse modalities and supervision regimes, FairMT consistently achieves substantial fairness gains while maintaining superior task utility. Code will be released upon paper acceptance.", "date": "2025-11-29", "categories": []}
{"id": "2512.00545", "title": "DQ4FairIM: Fairness-aware Influence Maximization using Deep Reinforcement Learning", "abstract": "The Influence Maximization (IM) problem aims to select a set of seed nodes within a given budget to maximize the spread of influence in a social network. However, real-world social networks have several structural inequalities, such as dominant majority groups and underrepresented minority groups. If these inequalities are not considered while designing IM algorithms, the outcomes might be biased, disproportionately benefiting majority groups while marginalizing minorities. In this work, we address this gap by designing a fairness-aware IM method using Reinforcement Learning (RL) that ensures equitable influence outreach across all communities, regardless of protected attributes. Fairness is incorporated using a maximin fairness objective, which prioritizes improving the outreach of the least-influenced group, pushing the solution toward an equitable influence distribution. We propose a novel fairness-aware deep RL method, called DQ4FairIM, that maximizes the expected number of influenced nodes by learning an RL policy. The learnt policy ensures that minority groups formulate the IM problem as a Markov Decision Process (MDP) and use deep Q-learning, combined with the Structure2Vec network embedding, earning together with Structure2Vec network embedding to solve the MDP. We perform extensive experiments on synthetic benchmarks and real-world networks to compare our method with fairness-agnostic and fairness-aware baselines. The results show that our method achieves a higher level of fairness while maintaining a better fairness-performance trade-off than baselines. Additionally, our approach learns effective seeding policies that generalize across problem instances without retraining, such as varying the network size or the number of seed nodes.", "date": "2025-11-29", "categories": []}
{"id": "2512.00598", "title": "Developing Fairness-Aware Task Decomposition to Improve Equity in Post-Spinal Fusion Complication Prediction", "abstract": "Fairness in clinical prediction models remains a persistent challenge, particularly in high-stakes applications such as spinal fusion surgery for scoliosis, where patient outcomes exhibit substantial heterogeneity. Many existing fairness approaches rely on coarse demographic adjustments or post-hoc corrections, which fail to capture the latent structure of clinical populations and may unintentionally reinforce bias. We propose FAIR-MTL, a fairness-aware multitask learning framework designed to provide equitable and fine-grained prediction of postoperative complication severity.\n  Instead of relying on explicit sensitive attributes during model training, FAIR-MTL employs a data-driven subgroup inference mechanism. We extract a compact demographic embedding, and apply k-means clustering to uncover latent patient subgroups that may be differentially affected by traditional models. These inferred subgroup labels determine task routing within a shared multitask architecture. During training, subgroup imbalance is mitigated through inverse-frequency weighting, and regularization prevents overfitting to smaller groups.\n  Applied to postoperative complication prediction with four severity levels, FAIR-MTL achieves an AUC of 0.86 and an accuracy of 75%, outperforming single-task baselines while substantially reducing bias. For gender, the demographic parity difference decreases to 0.055 and equalized odds to 0.094; for age, these values reduce to 0.056 and 0.148, respectively. Model interpretability is ensured through SHAP and Gini importance analyses, which consistently highlight clinically meaningful predictors such as hemoglobin, hematocrit, and patient weight. Our findings show that incorporating unsupervised subgroup discovery into a multitask framework enables more equitable, interpretable, and clinically actionable predictions for surgical risk stratification.", "date": "2025-11-29", "categories": []}
{"id": "2512.00807", "title": "BioPro: On Difference-Aware Gender Fairness for Vision-Language Models", "abstract": "Vision-Language Models (VLMs) inherit significant social biases from their training data, notably in gender representation. Current fairness interventions often adopt a difference-unaware perspective that enforces uniform treatment across demographic groups. These approaches, however, fail to distinguish between contexts where neutrality is required and those where group-specific attributes are legitimate and must be preserved. Building upon recent advances in difference-aware fairness for text-only models, we extend this concept to the multimodal domain and formalize the problem of difference-aware gender fairness for image captioning and text-to-image generation. We advocate for selective debiasing, which aims to mitigate unwanted bias in neutral contexts while preserving valid distinctions in explicit ones. To achieve this, we propose BioPro (Bias Orthogonal Projection), an entirely training-free framework. BioPro identifies a low-dimensional gender-variation subspace through counterfactual embeddings and applies projection to selectively neutralize gender-related information. Experiments show that BioPro effectively reduces gender bias in neutral cases while maintaining gender faithfulness in explicit ones, thus providing a promising direction toward achieving selective fairness in VLMs. Beyond gender bias, we further demonstrate that BioPro can effectively generalize to continuous bias variables, such as scene brightness, highlighting its broader applicability.", "date": "2025-11-30", "categories": []}
{"id": "2403.10624", "title": "Configurable Fairness: Direct Optimization of Parity Metrics via Vision-Language Models", "abstract": "Performance disparities of image recognition across demographic groups are known to exist in deep learning-based models, due to imbalanced group representations or spurious correlation between group and target labels. Previous work has addressed such challenges without relying on expensive group labels, typically by upweighting high-loss samples or balancing discovered clusters. However, these heuristic strategies lack direct connection to specific fairness metrics and cannot guarantee optimization of parity-based criteria like equal opportunity, which ensures equal chance to receive positive outcomes across groups. In this work, we propose a novel paradigm that directly optimizes parity-based fairness metrics through specifically designed training objectives, without requiring group labels. We leverage vision-language models to analyze sensitive attribute relevancy for individual samples, then formulate loss functions that mathematically connect to each target fairness metric. This enables flexible optimization of different fairness criteria based on application needs. Experiments on multiple image classification datasets show that our metric-specific approach significantly improves parity-based fairness criteria and outperforms existing methods.", "date": "2024-03-15", "categories": []}
{"id": "2511.21711", "title": "Addressing Stereotypes in Large Language Models: A Critical Examination and Mitigation", "abstract": "Large Language models (LLMs), such as ChatGPT, have gained popularity in recent years with the advancement of Natural Language Processing (NLP), with use cases spanning many disciplines and daily lives as well. LLMs inherit explicit and implicit biases from the datasets they were trained on; these biases can include social, ethical, cultural, religious, and other prejudices and stereotypes. It is important to comprehensively examine such shortcomings by identifying the existence and extent of such biases, recognizing the origin, and attempting to mitigate such biased outputs to ensure fair outputs to reduce harmful stereotypes and misinformation. This study inspects and highlights the need to address biases in LLMs amid growing generative Artificial Intelligence (AI). We utilize bias-specific benchmarks such StereoSet and CrowSPairs to evaluate the existence of various biases in many different generative models such as BERT, GPT 3.5, and ADA. To detect both explicit and implicit biases, we adopt a three-pronged approach for thorough and inclusive analysis. Results indicate fine-tuned models struggle with gender biases but excel at identifying and avoiding racial biases. Our findings also illustrated that despite some cases of success, LLMs often over-rely on keywords in prompts and its outputs. This demonstrates the incapability of LLMs to attempt to truly understand the accuracy and authenticity of its outputs. Finally, in an attempt to bolster model performance, we applied an enhancement learning strategy involving fine-tuning, models using different prompting techniques, and data augmentation of the bias benchmarks. We found fine-tuned models to exhibit promising adaptability during cross-dataset testing and significantly enhanced performance on implicit bias benchmarks, with performance gains of up to 20%.", "date": "2025-11-18", "categories": []}
{"id": "2511.22016", "title": "AfriStereo: A Culturally Grounded Dataset for Evaluating Stereotypical Bias in Large Language Models", "abstract": "Existing AI bias evaluation benchmarks largely reflect Western perspectives, leaving African contexts underrepresented and enabling harmful stereotypes in applications across various domains. To address this gap, we introduce AfriStereo, the first open-source African stereotype dataset and evaluation framework grounded in local socio-cultural contexts. Through community engaged efforts across Senegal, Kenya, and Nigeria, we collected 1,163 stereotypes spanning gender, ethnicity, religion, age, and profession. Using few-shot prompting with human-in-the-loop validation, we augmented the dataset to over 5,000 stereotype-antistereotype pairs. Entries were validated through semantic clustering and manual annotation by culturally informed reviewers. Preliminary evaluation of language models reveals that nine of eleven models exhibit statistically significant bias, with Bias Preference Ratios (BPR) ranging from 0.63 to 0.78 (p <= 0.05), indicating systematic preferences for stereotypes over antistereotypes, particularly across age, profession, and gender dimensions. Domain-specific models appeared to show weaker bias in our setup, suggesting task-specific training may mitigate some associations. Looking ahead, AfriStereo opens pathways for future research on culturally grounded bias evaluation and mitigation, offering key methodologies for the AI community on building more equitable, context-aware, and globally inclusive NLP technologies.", "date": "2025-11-27", "categories": []}
{"id": "2511.23097", "title": "Fairness in the Multi-Secretary Problem", "abstract": "This paper bridges two perspectives: it studies the multi-secretary problem through the fairness lens of social choice, and examines multi-winner elections from the viewpoint of online decision making. After identifying the limitations of the prominent proportionality notion of Extended Justified Representation (EJR) in the online domain, the work proposes a set of mechanisms that merge techniques from online algorithms with rules from social choice -- such as the Method of Equal Shares and the Nash Rule -- and supports them through both theoretical analysis and extensive experimental evaluation.", "date": "2025-11-28", "categories": []}
{"id": "2503.03401", "title": "Evolutionary Prediction Games", "abstract": "When a prediction algorithm serves a collection of users, disparities in prediction quality are likely to emerge. If users respond to accurate predictions by increasing engagement, inviting friends, or adopting trends, repeated learning creates a feedback loop that shapes both the model and the population of its users. In this work, we introduce evolutionary prediction games, a framework grounded in evolutionary game theory which models such feedback loops as natural-selection processes among groups of users. Our theoretical analysis reveals a gap between idealized and real-world learning settings: In idealized settings with unlimited data and computational power, repeated learning creates competition and promotes competitive exclusion across a broad class of behavioral dynamics. However, under realistic constraints such as finite data, limited compute, or risk of overfitting, we show that stable coexistence and mutualistic symbiosis between groups becomes possible. We analyze these possibilities in terms of their stability and feasibility, present mechanisms that can sustain their existence, and empirically demonstrate our findings.", "date": "2025-03-05", "categories": []}
{"id": "2511.21109", "title": "Interpretable Fair Clustering", "abstract": "Fair clustering has gained increasing attention in recent years, especially in applications involving socially sensitive attributes. However, existing fair clustering methods often lack interpretability, limiting their applicability in high-stakes scenarios where understanding the rationale behind clustering decisions is essential. In this work, we address this limitation by proposing an interpretable and fair clustering framework, which integrates fairness constraints into the structure of decision trees. Our approach constructs interpretable decision trees that partition the data while ensuring fair treatment across protected groups. To further enhance the practicality of our framework, we also introduce a variant that requires no fairness hyperparameter tuning, achieved through post-pruning a tree constructed without fairness constraints. Extensive experiments on both real-world and synthetic datasets demonstrate that our method not only delivers competitive clustering performance and improved fairness, but also offers additional advantages such as interpretability and the ability to handle multiple sensitive attributes. These strengths enable our method to perform robustly under complex fairness constraints, opening new possibilities for equitable and transparent clustering.", "date": "2025-11-26", "categories": []}
{"id": "2503.15815", "title": "Attention Pruning: Automated Fairness Repair of Language Models via Surrogate Simulated Annealing", "abstract": "This paper explores pruning attention heads as a post-processing bias mitigation method for large language models (LLMs). Modern AI systems such as LLMs are expanding into sensitive social contexts where fairness concerns become especially crucial. Since LLMs develop decision-making patterns by training on massive datasets of human-generated content, they naturally encode and perpetuate societal biases. While modifying training datasets and algorithms is expensive and requires significant resources; post-processing techniques-such as selectively deactivating neurons and attention heads in pre-trained LLMs-can provide feasible and effective approaches to improve fairness. However, identifying the optimal subset of parameters to prune presents a combinatorial challenge within LLMs' immense parameter space, requiring solutions that efficiently balance competing objectives across the frontiers of model fairness and utility.\n  To address the computational challenges, we explore a search-based program repair approach via randomized simulated annealing. Given the prohibitive evaluation costs in billion-parameter LLMs, we develop surrogate deep neural networks that efficiently model the relationship between attention head states (active/inactive) and their corresponding fairness/utility metrics. This allows us to perform optimization over the surrogate models and efficiently identify optimal subsets of attention heads for selective pruning rather than directly searching through the LLM parameter space. This paper introduces Attention Pruning, a fairness-aware surrogate simulated annealing approach to prune attention heads in LLMs that disproportionately contribute to bias while minimally impacting overall model utility. Our experiments show that Attention Pruning achieves up to $40\\%$ reduction in gender bias and outperforms the state-of-the-art bias mitigation strategies.", "date": "2025-03-20", "categories": []}
{"id": "2511.19722", "title": "Individual and group fairness in geographical partitioning", "abstract": "Socioeconomic segregation often arises in school districting and other contexts, causing some groups to be over- or under-represented within a particular district. This phenomenon is closely linked with disparities in opportunities and outcomes. We formulate a new class of geographical partitioning problems in which the population is heterogeneous, and it is necessary to ensure fair representation for each group at each facility. We prove that the optimal solution is a novel generalization of the additively weighted Voronoi diagram, and we propose a simple and efficient algorithm to compute it, thus resolving an open question dating back to Dvoretzky et al. (1951). The efficacy and potential for practical insight of the approach are demonstrated in a realistic case study involving seven demographic groups and $78$ district offices.", "date": "2025-11-24", "categories": []}
{"id": "2511.19956", "title": "Prompt Fairness: Sub-group Disparities in LLMs", "abstract": "Large Language Models (LLMs), though shown to be effective in many applications, can vary significantly in their response quality. In this paper, we investigate this problem of prompt fairness: specifically, the phrasing of a prompt by different users/styles, despite the same question being asked in principle, may elicit different responses from an LLM. To quantify this disparity, we propose to use information-theoretic metrics that can capture two dimensions of bias: subgroup sensitivity, the variability of responses within a subgroup and cross group consistency, the variability of responses across subgroups. Our analysis reveals that certain subgroups exhibit both higher internal variability and greater divergence from others. Our empirical analysis reveals that certain demographic sub groups experience both higher internal variability and greater divergence from others, indicating structural inequities in model behavior. To mitigate these disparities, we propose practical interventions, including majority voting across multiple generations and prompt neutralization, which together improve response stability and enhance fairness across user populations. In the experiments, we observe clear prompt sensitivity disparities across demographic subgroups: before mitigation, cross-group divergence values reach 0.28 and typically fall in the from 0.14 to 0.22 range. After applying our neutralization and multi generation strategy, these divergences consistently decrease, with the largest gap reduced to 0.22 and many distances falling to 0.17 or below, indicating more stable and consistent outputs across subgroups.", "date": "2025-11-25", "categories": []}
{"id": "2502.00313", "title": "Distributive Fairness in Large Language Models: Evaluating Alignment with Human Values", "abstract": "The growing interest in employing large language models (LLMs) for decision-making in social and economic contexts has raised questions about their potential to function as agents in these domains. A significant number of societal problems involve the distribution of resources, where fairness, along with economic efficiency, play a critical role in the desirability of outcomes. In this paper, we examine whether LLM responses adhere to fundamental fairness concepts such as equitability, envy-freeness, and Rawlsian maximin, and investigate their alignment with human preferences. We evaluate the performance of several LLMs, providing a comparative benchmark of their ability to reflect these measures. Our results demonstrate a lack of alignment between current LLM responses and human distributional preferences. Moreover, LLMs are unable to utilize money as a transferable resource to mitigate inequality. Nonetheless, we demonstrate a stark contrast when (some) LLMs are tasked with selecting from a predefined menu of options rather than generating one. In addition, we analyze the robustness of LLM responses to variations in semantic factors (e.g., intentions or personas) or non-semantic prompting changes (e.g., templates or orderings). Finally, we highlight potential strategies aimed at enhancing the alignment of LLM behavior with well-established fairness concepts.", "date": "2025-02-01", "categories": []}
{"id": "2509.19659", "title": "Bias in the Picture: Benchmarking VLMs with Social-Cue News Images and LLM-as-Judge Assessment", "abstract": "Large vision-language models (VLMs) can jointly interpret images and text, but they are also prone to absorbing and reproducing harmful social stereotypes when visual cues such as age, gender, race, clothing, or occupation are present. To investigate these risks, we introduce a news-image benchmark consisting of 1,343 image-question pairs drawn from diverse outlets, which we annotated with ground-truth answers and demographic attributes (age, gender, race, occupation, and sports). We evaluate a range of state-of-the-art VLMs and employ a large language model (LLM) as judge, with human verification. Our findings show that: (i) visual context systematically shifts model outputs in open-ended settings; (ii) bias prevalence varies across attributes and models, with particularly high risk for gender and occupation; and (iii) higher faithfulness does not necessarily correspond to lower bias. We release the benchmark prompts, evaluation rubric, and code to support reproducible and fairness-aware multimodal assessment.", "date": "2025-09-24", "categories": []}
{"id": "2509.25662", "title": "On Explaining Proxy Discrimination and Unfairness in Individual Decisions Made by AI Systems", "abstract": "Artificial intelligence (AI) systems in high-stakes domains raise concerns about proxy discrimination, unfairness, and explainability. Existing audits often fail to reveal why unfairness arises, particularly when rooted in structural bias. We propose a novel framework using formal abductive explanations to explain proxy discrimination in individual AI decisions. Leveraging background knowledge, our method identifies which features act as unjustified proxies for protected attributes, revealing hidden structural biases. Central to our approach is the concept of aptitude, a task-relevant property independent of group membership, with a mapping function aligning individuals of equivalent aptitude across groups to assess fairness substantively. As a proof of concept, we showcase the framework with examples taken from the German credit dataset, demonstrating its applicability in real-world cases.", "date": "2025-09-30", "categories": []}
{"id": "2511.18876", "title": "Fairness Meets Privacy: Integrating Differential Privacy and Demographic Parity in Multi-class Classification", "abstract": "The increasing use of machine learning in sensitive applications demands algorithms that simultaneously preserve data privacy and ensure fairness across potentially sensitive sub-populations. While privacy and fairness have each been extensively studied, their joint treatment remains poorly understood. Existing research often frames them as conflicting objectives, with multiple studies suggesting that strong privacy notions such as differential privacy inevitably compromise fairness. In this work, we challenge that perspective by showing that differential privacy can be integrated into a fairness-enhancing pipeline with minimal impact on fairness guarantees. We design a postprocessing algorithm, called DP2DP, that enforces both demographic parity and differential privacy. Our analysis reveals that our algorithm converges towards its demographic parity objective at essentially the same rate (up logarithmic factor) as the best non-private methods from the literature. Experiments on both synthetic and real datasets confirm our theoretical results, showing that the proposed algorithm achieves state-of-the-art accuracy/fairness/privacy trade-offs.", "date": "2025-11-24", "categories": []}
{"id": "2505.12114", "title": "Behind the Screens: Uncovering Bias in AI-Driven Video Interview Assessments Using Counterfactuals", "abstract": "AI-enhanced personality assessments are increasingly shaping hiring decisions, using affective computing to predict traits from the Big Five (OCEAN) model. However, integrating AI into these assessments raises ethical concerns, especially around bias amplification rooted in training data. These biases can lead to discriminatory outcomes based on protected attributes like gender, ethnicity, and age. To address this, we introduce a counterfactual-based framework to systematically evaluate and quantify bias in AI-driven personality assessments. Our approach employs generative adversarial networks (GANs) to generate counterfactual representations of job applicants by altering protected attributes, enabling fairness analysis without access to the underlying model. Unlike traditional bias assessments that focus on unimodal or static data, our method supports multimodal evaluation-spanning visual, audio, and textual features. This comprehensive approach is particularly important in high-stakes applications like hiring, where third-party vendors often provide AI systems as black boxes. Applied to a state-of-the-art personality prediction model, our method reveals significant disparities across demographic groups. We also validate our framework using a protected attribute classifier to confirm the effectiveness of our counterfactual generation. This work provides a scalable tool for fairness auditing of commercial AI hiring platforms, especially in black-box settings where training data and model internals are inaccessible. Our results highlight the importance of counterfactual approaches in improving ethical transparency in affective computing.", "date": "2025-05-17", "categories": []}
{"id": "2507.04224", "title": "Fairness Evaluation of Large Language Models in Academic Library Reference Services", "abstract": "As libraries explore large language models (LLMs) for use in virtual reference services, a key question arises: Can LLMs serve all users equitably, regardless of demographics or social status? While they offer great potential for scalable support, LLMs may also reproduce societal biases embedded in their training data, risking the integrity of libraries' commitment to equitable service. To address this concern, we evaluate whether LLMs differentiate responses across user identities by prompting six state-of-the-art LLMs to assist patrons differing in sex, race/ethnicity, and institutional role. We find no evidence of differentiation by race or ethnicity, and only minor evidence of stereotypical bias against women in one model. LLMs demonstrate nuanced accommodation of institutional roles through the use of linguistic choices related to formality, politeness, and domain-specific vocabularies, reflecting professional norms rather than discriminatory treatment. These findings suggest that current LLMs show a promising degree of readiness to support equitable and contextually appropriate communication in academic library reference services.", "date": "2025-07-06", "categories": []}
{"id": "2511.16248", "title": "Revisiting Fairness-aware Interactive Recommendation: Item Lifecycle as a Control Knob", "abstract": "This paper revisits fairness-aware interactive recommendation (e.g., TikTok, KuaiShou) by introducing a novel control knob, i.e., the lifecycle of items. We make threefold contributions. First, we conduct a comprehensive empirical analysis and uncover that item lifecycles in short-video platforms follow a compressed three-phase pattern, i.e., rapid growth, transient stability, and sharp decay, which significantly deviates from the classical four-stage model (introduction, growth, maturity, decline). Second, we introduce LHRL, a lifecycle-aware hierarchical reinforcement learning framework that dynamically harmonizes fairness and accuracy by leveraging phase-specific exposure dynamics. LHRL consists of two key components: (1) PhaseFormer, a lightweight encoder combining STL decomposition and attention mechanisms for robust phase detection; (2) a two-level HRL agent, where the high-level policy imposes phase-aware fairness constraints, and the low-level policy optimizes immediate user engagement. This decoupled optimization allows for effective reconciliation between long-term equity and short-term utility. Third, experiments on multiple real-world interactive recommendation datasets demonstrate that LHRL significantly improves both fairness and user engagement. Furthermore, the integration of lifecycle-aware rewards into existing RL-based models consistently yields performance gains, highlighting the generalizability and practical value of our approach.", "date": "2025-11-20", "categories": []}
{"id": "2511.16549", "title": "FairLRF: Achieving Fairness through Sparse Low Rank Factorization", "abstract": "As deep learning (DL) techniques become integral to various applications, ensuring model fairness while maintaining high performance has become increasingly critical, particularly in sensitive fields such as medical diagnosis. Although a variety of bias-mitigation methods have been proposed, many rely on computationally expensive debiasing strategies or suffer substantial drops in model accuracy, which limits their practicality in real-world, resource-constrained settings. To address this issue, we propose a fairness-oriented low rank factorization (LRF) framework that leverages singular value decomposition (SVD) to improve DL model fairness. Unlike traditional SVD, which is mainly used for model compression by decomposing and reducing weight matrices, our work shows that SVD can also serve as an effective tool for fairness enhancement. Specifically, we observed that elements in the unitary matrices obtained from SVD contribute unequally to model bias across groups defined by sensitive attributes. Motivated by this observation, we propose a method, named FairLRF, that selectively removes bias-inducing elements from unitary matrices to reduce group disparities, thus enhancing model fairness. Extensive experiments show that our method outperforms conventional LRF methods as well as state-of-the-art fairness-enhancing techniques. Additionally, an ablation study examines how major hyper-parameters may influence the performance of processed models. To the best of our knowledge, this is the first work utilizing SVD not primarily for compression but for fairness enhancement.", "date": "2025-11-20", "categories": []}
{"id": "2406.17385", "title": "Native Design Bias: Studying the Impact of English Nativeness on Language Model Performance", "abstract": "Large Language Models (LLMs) excel at providing information acquired during pretraining on large-scale corpora and following instructions through user prompts. This study investigates whether the quality of LLM responses varies depending on the demographic profile of users. Considering English as the global lingua franca, along with the diversity of its dialects among speakers of different native languages, we explore whether non-native English speakers receive lower-quality or even factually incorrect responses from LLMs more frequently. Our results show that performance discrepancies occur when LLMs are prompted by native versus non-native English speakers and persist when comparing native speakers from Western countries with others. Additionally, we find a strong anchoring effect when the model recognizes or is made aware of the user's nativeness, which further degrades the response quality when interacting with non-native speakers. Our analysis is based on a newly collected dataset with over 12,000 unique annotations from 124 annotators, including information on their native language and English proficiency.", "date": "2024-06-25", "categories": []}
{"id": "2412.06632", "title": "MAVias: Mitigate any Visual Bias", "abstract": "Mitigating biases in computer vision models is an essential step towards the trustworthiness of artificial intelligence models. Existing bias mitigation methods focus on a small set of predefined biases, limiting their applicability in visual datasets where multiple, possibly unknown biases exist. To address this limitation, we introduce MAVias, an open-set bias mitigation approach leveraging foundation models to discover spurious associations between visual attributes and target classes. MAVias first captures a wide variety of visual features in natural language via a foundation image tagging model, and then leverages a large language model to select those visual features defining the target class, resulting in a set of language-coded potential visual biases. We then translate this set of potential biases into vision-language embeddings and introduce an in-processing bias mitigation approach to prevent the model from encoding information related to them. Our experiments on diverse datasets, including CelebA, Waterbirds, ImageNet, and UrbanCars, show that MAVias effectively detects and mitigates a wide range of biases in visual recognition tasks outperforming current state-of-the-art.", "date": "2024-12-09", "categories": []}
{"id": "2508.18708", "title": "Skill-Aligned Fairness in Multi-Agent Learning for Collaboration in Healthcare", "abstract": "Fairness in multi-agent reinforcement learning (MARL) is often framed as a workload balance problem, overlooking agent expertise and the structured coordination required in real-world domains. In healthcare, equitable task allocation requires workload balance or expertise alignment to prevent burnout and overuse of highly skilled agents. Workload balance refers to distributing an approximately equal number of subtasks or equalised effort across healthcare workers, regardless of their expertise. We make two contributions to address this problem. First, we propose FairSkillMARL, a framework that defines fairness as the dual objective of workload balance and skill-task alignment. Second, we introduce MARLHospital, a customizable healthcare-inspired environment for modeling team compositions and energy-constrained scheduling impacts on fairness, as no existing simulators are well-suited for this problem. We conducted experiments to compare FairSkillMARL in conjunction with four standard MARL methods, and against two state-of-the-art fairness metrics. Our results suggest that fairness based solely on equal workload might lead to task-skill mismatches and highlight the need for more robust metrics that capture skill-task misalignment. Our work provides tools and a foundation for studying fairness in heterogeneous multi-agent systems where aligning effort with expertise is critical.", "date": "2025-08-26", "categories": []}
{"id": "2509.04128", "title": "Revisiting (Un)Fairness in Recourse by Minimizing Worst-Case Social Burden", "abstract": "Machine learning based predictions are increasingly used in sensitive decision-making applications that directly affect our lives. This has led to extensive research into ensuring the fairness of classifiers. Beyond just fair classification, emerging legislation now mandates that when a classifier delivers a negative decision, it must also offer actionable steps an individual can take to reverse that outcome. This concept is known as algorithmic recourse. Nevertheless, many researchers have expressed concerns about the fairness guarantees within the recourse process itself. In this work, we provide a holistic theoretical characterization of unfairness in algorithmic recourse, formally linking fairness guarantees in recourse and classification, and highlighting limitations of the standard equal cost paradigm. We then introduce a novel fairness framework based on social burden, along with a practical algorithm (MISOB), broadly applicable under real-world conditions. Empirical results on real-world datasets show that MISOB reduces the social burden across all groups without compromising overall classifier accuracy.", "date": "2025-09-04", "categories": []}
{"id": "2511.13793", "title": "Modeling Fairness in Recruitment AI via Information Flow", "abstract": "Avoiding bias and understanding the real-world consequences of AI-supported decision-making are critical to address fairness and assign accountability. Existing approaches often focus either on technical aspects, such as datasets and models, or on high-level socio-ethical considerations - rarely capturing how these elements interact in practice. In this paper, we apply an information flow-based modeling framework to a real-world recruitment process that integrates automated candidate matching with human decision-making. Through semi-structured stakeholder interviews and iterative modeling, we construct a multi-level representation of the recruitment pipeline, capturing how information is transformed, filtered, and interpreted across both algorithmic and human components. We identify where biases may emerge, how they can propagate through the system, and what downstream impacts they may have on candidates. This case study illustrates how information flow modeling can support structured analysis of fairness risks, providing transparency across complex socio-technical systems.", "date": "2025-11-16", "categories": []}
{"id": "2511.14153", "title": "A Comprehensive Study of Implicit and Explicit Biases in Large Language Models", "abstract": "Large Language Models (LLMs) inherit explicit and implicit biases from their training datasets. Identifying and mitigating biases in LLMs is crucial to ensure fair outputs, as they can perpetuate harmful stereotypes and misinformation. This study highlights the need to address biases in LLMs amid growing generative AI. We studied bias-specific benchmarks such as StereoSet and CrowSPairs to evaluate the existence of various biases in multiple generative models such as BERT and GPT 3.5. We proposed an automated Bias-Identification Framework to recognize various social biases in LLMs such as gender, race, profession, and religion. We adopted a two-pronged approach to detect explicit and implicit biases in text data. Results indicated fine-tuned models struggle with gender biases but excelled at identifying and avoiding racial biases. Our findings illustrated that despite having some success, LLMs often over-relied on keywords. To illuminate the capability of the analyzed LLMs in detecting implicit biases, we employed Bag-of-Words analysis and unveiled indications of implicit stereotyping within the vocabulary. To bolster the model performance, we applied an enhancement strategy involving fine-tuning models using prompting techniques and data augmentation of the bias benchmarks. The fine-tuned models exhibited promising adaptability during cross-dataset testing and significantly enhanced performance on implicit bias benchmarks, with performance gains of up to 20%.", "date": "2025-11-18", "categories": []}
{"id": "2411.02671", "title": "Fair In-Context Learning via Latent Concept Variables", "abstract": "The emerging in-context learning (ICL) ability of large language models (LLMs) has prompted their use for predictive tasks in various domains with different data types, including tabular data, facilitated by serialization methods. However, with increasing applications in high-stakes domains, it has been shown that LLMs can inherit social bias and discrimination from their pre-training data. In this work, we investigate inherent bias in LLMs during in-context learning with tabular data. We focus on an optimal demonstration selection approach that utilizes latent concept variables for resource-efficient task adaptation. We design data augmentation strategies that reduce the correlation between predictive outcomes and sensitive variables, helping promote fairness during latent concept learning. We utilize the learned concept to select demonstrations and obtain fair predictions. The latent concept variables are learned using a smaller internal LLM and generalized to larger external LLMs. We empirically verify that the fair latent variable approach improves fairness results on tabular datasets compared to multiple heuristic demonstration selection methods.", "date": "2024-11-04", "categories": []}
{"id": "2505.00631", "title": "Bayes-Optimal Fair Classification with Multiple Sensitive Features", "abstract": "Existing theoretical work on Bayes-optimal fair classifiers usually considers a single (binary) sensitive feature. In practice, individuals are often defined by multiple sensitive features. In this paper, we characterize the Bayes-optimal fair classifier for multiple sensitive features under general approximate fairness measures, including mean difference and mean ratio. We show that these approximate measures for existing group fairness notions, including Demographic Parity, Equal Opportunity, Predictive Equality, and Accuracy Parity, are linear transformations of selection rates for specific groups defined by both labels and sensitive features. We then characterize that Bayes-optimal fair classifiers for multiple sensitive features become instance-dependent thresholding rules that rely on a weighted sum of these group membership probabilities. Our framework applies to both attribute-aware and attribute-blind settings and can accommodate composite fairness notions like Equalized Odds. Building on this, we propose two practical algorithms for Bayes-optimal fair classification via in-processing and post-processing. We show empirically that our methods compare favorably to existing methods.", "date": "2025-05-01", "categories": []}
{"id": "2508.04511", "title": "Argumentative Debates for Transparent Bias Detection [Technical Report]", "abstract": "As the use of AI in society grows, addressing emerging biases is essential to prevent systematic discrimination. Several bias detection methods have been proposed, but, with few exceptions, these tend to ignore transparency. Instead, interpretability and explainability are core requirements for algorithmic fairness, even more so than for other algorithmic solutions, given the human-oriented nature of fairness. We present ABIDE (Argumentative BIas detection by DEbate), a novel framework that structures bias detection transparently as debate, guided by an underlying argument graph as understood in (formal and computational) argumentation. The arguments are about the success chances of groups in local neighbourhoods and the significance of these neighbourhoods. We evaluate ABIDE experimentally and demonstrate its strengths in performance against an argumentative baseline.", "date": "2025-08-06", "categories": []}
{"id": "2511.06568", "title": "Breaking the Dyadic Barrier: Rethinking Fairness in Link Prediction Beyond Demographic Parity", "abstract": "Link prediction is a fundamental task in graph machine learning with applications, ranging from social recommendation to knowledge graph completion. Fairness in this setting is critical, as biased predictions can exacerbate societal inequalities. Prior work adopts a dyadic definition of fairness, enforcing fairness through demographic parity between intra-group and inter-group link predictions. However, we show that this dyadic framing can obscure underlying disparities across subgroups, allowing systemic biases to go undetected. Moreover, we argue that demographic parity does not meet desired properties for fairness assessment in ranking-based tasks such as link prediction. We formalize the limitations of existing fairness evaluations and propose a framework that enables a more expressive assessment. Additionally, we propose a lightweight post-processing method combined with decoupled link predictors that effectively mitigates bias and achieves state-of-the-art fairness-utility trade-offs.", "date": "2025-11-09", "categories": []}
{"id": "2511.10089", "title": "T2IBias: Uncovering Societal Bias Encoded in the Latent Space of Text-to-Image Generative Models", "abstract": "Text-to-image (T2I) generative models are largely used in AI-powered real-world applications and value creation. However, their strategic deployment raises critical concerns for responsible AI management, particularly regarding the reproduction and amplification of race- and gender-related stereotypes that can undermine organizational ethics. In this work, we investigate whether such societal biases are systematically encoded within the pretrained latent spaces of state-of-the-art T2I models. We conduct an empirical study across the five most popular open-source models, using ten neutral, profession-related prompts to generate 100 images per profession, resulting in a dataset of 5,000 images evaluated by diverse human assessors representing different races and genders. We demonstrate that all five models encode and amplify pronounced societal skew: caregiving and nursing roles are consistently feminized, while high-status professions such as corporate CEO, politician, doctor, and lawyer are overwhelmingly represented by males and mostly White individuals. We further identify model-specific patterns, such as QWEN-Image's near-exclusive focus on East Asian outputs, Kandinsky's dominance of White individuals, and SDXL's comparatively broader but still biased distributions. These results provide critical insights for AI project managers and practitioners, enabling them to select equitable AI models and customized prompts that generate images in alignment with the principles of responsible AI. We conclude by discussing the risks of these biases and proposing actionable strategies for bias mitigation in building responsible GenAI systems. The code and Data Repository: https://github.com/Sufianlab/T2IBias", "date": "2025-11-13", "categories": []}
{"id": "2511.11575", "title": "Detecting Statistically Significant Fairness Violations in Recidivism Forecasting Algorithms", "abstract": "Machine learning algorithms are increasingly deployed in critical domains such as finance, healthcare, and criminal justice [1]. The increasing popularity of algorithmic decision-making has stimulated interest in algorithmic fairness within the academic community. Researchers have introduced various fairness definitions that quantify disparities between privileged and protected groups, use causal inference to determine the impact of race on model predictions, and that test calibration of probability predictions from the model. Existing literature does not provide a way in which to assess whether observed disparities between groups are statistically significant or merely due to chance. This paper introduces a rigorous framework for testing the statistical significance of fairness violations by leveraging k-fold cross-validation [2] to generate sampling distributions of fairness metrics. This paper introduces statistical tests that can be used to identify statistically significant violations of fairness metrics based on disparities between predicted and actual outcomes, model calibration, and causal inference techniques [1]. We demonstrate this approach by testing recidivism forecasting algorithms trained on data from the National Institute of Justice. Our findings reveal that machine learning algorithms used for recidivism forecasting exhibit statistically significant bias against Black individuals under several fairness definitions, while also exhibiting no bias or bias against White individuals under other definitions. The results from this paper underscore the importance of rigorous and robust statistical testing while evaluating algorithmic decision-making systems.", "date": "2025-09-18", "categories": []}
{"id": "2511.11636", "title": "An Explainable and Fair AI Tool for PCOS Risk Assessment: Calibration, Subgroup Equity, and Interactive Clinical Deployment", "abstract": "This paper presents a fairness-audited and interpretable machine learning framework for predicting polycystic ovary syndrome (PCOS), designed to evaluate model performance and identify diagnostic disparities across patient subgroups. The framework integrated SHAP-based feature attributions with demographic audits to connect predictive explanations with observed disparities for actionable insights. Probabilistic calibration metrics (Brier Score and Expected Calibration Error) are incorporated to ensure reliable risk predictions across subgroups. Random Forest, SVM, and XGBoost models were trained with isotonic and Platt scaling for calibration and fairness comparison. A calibrated Random Forest achieved a high predictive accuracy of 90.8%. SHAP analysis identified follicle count, weight gain, and menstrual irregularity as the most influential features, which are consistent with the Rotterdam diagnostic criteria. Although the SVM with isotonic calibration achieved the lowest calibration error (ECE = 0.0541), the Random Forest model provided a better balance between calibration and interpretability (Brier = 0.0678, ECE = 0.0666). Therefore, it was selected for detailed fairness and SHAP analyses. Subgroup analysis revealed that the model performed best among women aged 25-35 (accuracy 90.9%) but underperformed in those under 25 (69.2%), highlighting age-related disparities. The model achieved perfect precision in obese women and maintained high recall in lean PCOS cases, demonstrating robustness across phenotypes. Finally, a Streamlit-based web interface enables real-time PCOS risk assessment, Rotterdam criteria evaluation, and interactive 'what-if' analysis, bridging the gap between AI research and clinical usability.", "date": "2025-11-08", "categories": []}
{"id": "2511.12132", "title": "FairGSE: Fairness-Aware Graph Neural Network without High False Positive Rates", "abstract": "Graph neural networks (GNNs) have emerged as the mainstream paradigm for graph representation learning due to their effective message aggregation. However, this advantage also amplifies biases inherent in graph topology, raising fairness concerns. Existing fairness-aware GNNs provide satisfactory performance on fairness metrics such as Statistical Parity and Equal Opportunity while maintaining acceptable accuracy trade-offs. Unfortunately, we observe that this pursuit of fairness metrics neglects the GNN's ability to predict negative labels, which renders their predictions with extremely high False Positive Rates (FPR), resulting in negative effects in high-risk scenarios. To this end, we advocate that classification performance should be carefully calibrated while improving fairness, rather than simply constraining accuracy loss. Furthermore, we propose Fair GNN via Structural Entropy (\\textbf{FairGSE}), a novel framework that maximizes two-dimensional structural entropy (2D-SE) to improve fairness without neglecting false positives. Experiments on several real-world datasets show FairGSE reduces FPR by 39\\% vs. state-of-the-art fairness-aware GNNs, with comparable fairness improvement.", "date": "2025-11-15", "categories": []}
{"id": "2511.13525", "title": "AI Fairness Beyond Complete Demographics: Current Achievements and Future Directions", "abstract": "Fairness in artificial intelligence (AI) has become a growing concern due to discriminatory outcomes in AI-based decision-making systems. While various methods have been proposed to mitigate bias, most rely on complete demographic information, an assumption often impractical due to legal constraints and the risk of reinforcing discrimination. This survey examines fairness in AI when demographics are incomplete, addressing the gap between traditional approaches and real-world challenges. We introduce a novel taxonomy of fairness notions in this setting, clarifying their relationships and distinctions. Additionally, we summarize existing techniques that promote fairness beyond complete demographics and highlight open research questions to encourage further progress in the field.", "date": "2025-11-17", "categories": []}
{"id": "2508.15374", "title": "Fairness for the People, by the People: Minority Collective Action", "abstract": "Machine learning models often preserve biases present in training data, leading to unfair treatment of certain minority groups. Despite an array of existing firm-side bias mitigation techniques, they typically incur utility costs and require organizational buy-in. Recognizing that many models rely on user-contributed data, end-users can induce fairness through the framework of Algorithmic Collective Action, where a coordinated minority group strategically relabels its own data to enhance fairness, without altering the firm's training process. We propose three practical, model-agnostic methods to approximate ideal relabeling and validate them on real-world datasets. Our findings show that a subgroup of the minority can substantially reduce unfairness with a small impact on the overall prediction error.", "date": "2025-08-21", "categories": []}
{"id": "2511.10689", "title": "Equilibrium Dynamics and Mitigation of Gender Bias in Synthetically Generated Data", "abstract": "Recursive prompting with large language models enables scalable synthetic dataset generation but introduces the risk of bias amplification. We investigate gender bias dynamics across three generations of recursive text generation using three complementary evaluation frameworks: rule-based pattern matching, embedding-based semantic similarity, and downstream task performance. Experiments with three initial bias levels (0.1, 0.3, 0.6) and four mitigation strategies reveal equilibrium dynamics rather than monotonic amplification. The low initial bias amplifies toward the model's inherent bias level (+36%), whereas the high initial bias decays toward it (-26%). Among mitigation methods, contrastive augmentation, which introduces gender-swapped variants, achieves significant downstream bias reduction (98.8% for low initial bias and 91% on average) despite producing higher embedding-based bias scores. This paradox demonstrates that semantic similarity metrics may diverge from behavioral fairness outcomes, highlighting the need for multidimensional evaluation in responsible synthetic data generation.", "date": "2025-11-12", "categories": []}
{"id": "2511.11144", "title": "Specification, Application, and Operationalization of a Metamodel of Fairness", "abstract": "This paper presents the AR fairness metamodel, aimed at formally representing, analyzing, and comparing fairness scenarios. The metamodel provides an abstract representation of fairness, enabling the formal definition of fairness notions. We instantiate the metamodel through several examples, with a particular focus on comparing the notions of equity and equality.\n  We use the Tiles framework, which offers modular components that can be interconnected to represent various definitions of fairness. Its primary objective is to support the operationalization of AR-based fairness definitions in a range of scenarios, providing a robust method for defining, comparing, and evaluating fairness.\n  Tiles has an open-source implementation for fairness modeling and evaluation.", "date": "2025-11-14", "categories": []}
{"id": "2511.11294", "title": "Decomposing Direct and Indirect Biases in Linear Models under Demographic Parity Constraint", "abstract": "Linear models are widely used in high-stakes decision-making due to their simplicity and interpretability. Yet when fairness constraints such as demographic parity are introduced, their effects on model coefficients, and thus on how predictive bias is distributed across features, remain opaque. Existing approaches on linear models often rely on strong and unrealistic assumptions, or overlook the explicit role of the sensitive attribute, limiting their practical utility for fairness assessment. We extend the work of (Chzhen and Schreuder, 2022) and (Fukuchi and Sakuma, 2023) by proposing a post-processing framework that can be applied on top of any linear model to decompose the resulting bias into direct (sensitive-attribute) and indirect (correlated-features) components. Our method analytically characterizes how demographic parity reshapes each model coefficient, including those of both sensitive and non-sensitive features. This enables a transparent, feature-level interpretation of fairness interventions and reveals how bias may persist or shift through correlated variables. Our framework requires no retraining and provides actionable insights for model auditing and mitigation. Experiments on both synthetic and real-world datasets demonstrate that our method captures fairness dynamics missed by prior work, offering a practical and interpretable tool for responsible deployment of linear models.", "date": "2025-11-14", "categories": []}
{"id": "2511.11413", "title": "Multicalibration yields better matchings", "abstract": "Consider the problem of finding the best matching in a weighted graph where we only have access to predictions of the actual stochastic weights, based on an underlying context. If the predictor is the Bayes optimal one, then computing the best matching based on the predicted weights is optimal. However, in practice, this perfect information scenario is not realistic. Given an imperfect predictor, a suboptimal decision rule may compensate for the induced error and thus outperform the standard optimal rule.\n  In this paper, we propose multicalibration as a way to address this problem. This fairness notion requires a predictor to be unbiased on each element of a family of protected sets of contexts. Given a class of matching algorithms $\\mathcal C$ and any predictor $\\gamma$ of the edge-weights, we show how to construct a specific multicalibrated predictor $\\hat \\gamma$, with the following property. Picking the best matching based on the output of $\\hat \\gamma$ is competitive with the best decision rule in $\\mathcal C$ applied onto the original predictor $\\gamma$. We complement this result by providing sample complexity bounds.", "date": "2025-11-14", "categories": []}
{"id": "2511.11459", "title": "FairReweighing: Density Estimation-Based Reweighing Framework for Improving Separation in Fair Regression", "abstract": "There has been a prevalence of applying AI software in both high-stakes public-sector and industrial contexts. However, the lack of transparency has raised concerns about whether these data-informed AI software decisions secure fairness against people of all racial, gender, or age groups. Despite extensive research on emerging fairness-aware AI software, up to now most efforts to solve this issue have been dedicated to binary classification tasks. Fairness in regression is relatively underexplored. In this work, we adopted a mutual information-based metric to assess separation violations. The metric is also extended so that it can be directly applied to both classification and regression problems with both binary and continuous sensitive attributes. Inspired by the Reweighing algorithm in fair classification, we proposed a FairReweighing pre-processing algorithm based on density estimation to ensure that the learned model satisfies the separation criterion. Theoretically, we show that the proposed FairReweighing algorithm can guarantee separation in the training data under a data independence assumption. Empirically, on both synthetic and real-world data, we show that FairReweighing outperforms existing state-of-the-art regression fairness solutions in terms of improving separation while maintaining high accuracy.", "date": "2025-11-14", "categories": []}
{"id": "2511.11539", "title": "Generalizing Fair Clustering to Multiple Groups: Algorithms and Applications", "abstract": "Clustering is a fundamental task in machine learning and data analysis, but it frequently fails to provide fair representation for various marginalized communities defined by multiple protected attributes -- a shortcoming often caused by biases in the training data. As a result, there is a growing need to enhance the fairness of clustering outcomes, ideally by making minimal modifications, possibly as a post-processing step after conventional clustering. Recently, Chakraborty et al. [COLT'25] initiated the study of \\emph{closest fair clustering}, though in a restricted scenario where data points belong to only two groups. In practice, however, data points are typically characterized by many groups, reflecting diverse protected attributes such as age, ethnicity, gender, etc.\n  In this work, we generalize the study of the \\emph{closest fair clustering} problem to settings with an arbitrary number (more than two) of groups. We begin by showing that the problem is NP-hard even when all groups are of equal size -- a stark contrast with the two-group case, for which an exact algorithm exists. Next, we propose near-linear time approximation algorithms that efficiently handle arbitrary-sized multiple groups, thereby answering an open question posed by Chakraborty et al. [COLT'25].\n  Leveraging our closest fair clustering algorithms, we further achieve improved approximation guarantees for the \\emph{fair correlation clustering} problem, advancing the state-of-the-art results established by Ahmadian et al. [AISTATS'20] and Ahmadi et al. [2020]. Additionally, we are the first to provide approximation algorithms for the \\emph{fair consensus clustering} problem involving multiple (more than two) groups, thus addressing another open direction highlighted by Chakraborty et al. [COLT'25].", "date": "2025-11-14", "categories": []}
{"id": "2511.09768", "title": "ProbLog4Fairness: A Neurosymbolic Approach to Modeling and Mitigating Bias", "abstract": "Operationalizing definitions of fairness is difficult in practice, as multiple definitions can be incompatible while each being arguably desirable. Instead, it may be easier to directly describe algorithmic bias through ad-hoc assumptions specific to a particular real-world task, e.g., based on background information on systemic biases in its context. Such assumptions can, in turn, be used to mitigate this bias during training. Yet, a framework for incorporating such assumptions that is simultaneously principled, flexible, and interpretable is currently lacking.\n  Our approach is to formalize bias assumptions as programs in ProbLog, a probabilistic logic programming language that allows for the description of probabilistic causal relationships through logic. Neurosymbolic extensions of ProbLog then allow for easy integration of these assumptions in a neural network's training process. We propose a set of templates to express different types of bias and show the versatility of our approach on synthetic tabular datasets with known biases. Using estimates of the bias distortions present, we also succeed in mitigating algorithmic bias in real-world tabular and image data. We conclude that ProbLog4Fairness outperforms baselines due to its ability to flexibly model the relevant bias assumptions, where other methods typically uphold a fixed bias type or notion of fairness.", "date": "2025-11-12", "categories": []}
{"id": "2511.09889", "title": "A General Anchor-Based Framework for Scalable Fair Clustering", "abstract": "Fair clustering is crucial for mitigating bias in unsupervised learning, yet existing algorithms often suffer from quadratic or super-quadratic computational complexity, rendering them impractical for large-scale datasets. To bridge this gap, we introduce the Anchor-based Fair Clustering Framework (AFCF), a novel, general, and plug-and-play framework that empowers arbitrary fair clustering algorithms with linear-time scalability. Our approach first selects a small but representative set of anchors using a novel fair sampling strategy. Then, any off-the-shelf fair clustering algorithm can be applied to this small anchor set. The core of our framework lies in a novel anchor graph construction module, where we formulate an optimization problem to propagate labels while preserving fairness. This is achieved through a carefully designed group-label joint constraint, which we prove theoretically ensures that the fairness of the final clustering on the entire dataset matches that of the anchor clustering. We solve this optimization efficiently using an ADMM-based algorithm. Extensive experiments on multiple large-scale benchmarks demonstrate that AFCF drastically accelerates state-of-the-art methods, which reduces computational time by orders of magnitude while maintaining strong clustering performance and fairness guarantees.", "date": "2025-11-13", "categories": []}
{"id": "2503.22454", "title": "A Causal Framework to Measure and Mitigate Non-binary Treatment Discrimination", "abstract": "Fairness studies of algorithmic decision-making systems often simplify complex decision processes, such as bail or loan approvals, into binary classification tasks. However, these approaches overlook that such decisions are not inherently binary (e.g., approve or not approve bail or loan); they also involve non-binary treatment decisions (e.g., bail conditions or loan terms) that can influence the downstream outcomes (e.g., loan repayment or reoffending). In this paper, we argue that non-binary treatment decisions are integral to the decision process and controlled by decision-makers and, therefore, should be central to fairness analyses in algorithmic decision-making. We propose a causal framework that extends fairness analyses and explicitly distinguishes between decision-subjects' covariates and the treatment decisions. This specification allows decision-makers to use our framework to (i) measure treatment disparity and its downstream effects in historical data and, using counterfactual reasoning, (ii) mitigate the impact of past unfair treatment decisions when automating decision-making. We use our framework to empirically analyze four widely used loan approval datasets to reveal potential disparity in non-binary treatment decisions and their discriminatory impact on outcomes, highlighting the need to incorporate treatment decisions in fairness assessments. Moreover, by intervening in treatment decisions, we show that our framework effectively mitigates treatment discrimination from historical data to ensure fair risk score estimation and (non-binary) decision-making processes that benefit all stakeholders.", "date": "2025-03-28", "categories": []}
{"id": "2509.04498", "title": "Where Should I Study? Biased Language Models Decide! Evaluating Fairness in LMs for Academic Recommendations", "abstract": "Large Language Models (LLMs) are increasingly used as daily recommendation systems for tasks like education planning, yet their recommendations risk perpetuating societal biases. This paper empirically examines geographic, demographic, and economic biases in university and program suggestions from three open-source LLMs: LLaMA-3.1-8B, Gemma-7B, and Mistral-7B. Using 360 simulated user profiles varying by gender, nationality, and economic status, we analyze over 25,000 recommendations. Results show strong biases: institutions in the Global North are disproportionately favored, recommendations often reinforce gender stereotypes, and institutional repetition is prevalent. While LLaMA-3.1 achieves the highest diversity, recommending 481 unique universities across 58 countries, systemic disparities persist. To quantify these issues, we propose a novel, multi-dimensional evaluation framework that goes beyond accuracy by measuring demographic and geographic representation. Our findings highlight the urgent need for bias consideration in educational LMs to ensure equitable global access to higher education.", "date": "2025-09-01", "categories": []}
{"id": "2511.08702", "title": "FAIRPLAI: A Human-in-the-Loop Approach to Fair and Private Machine Learning", "abstract": "As machine learning systems move from theory to practice, they are increasingly tasked with decisions that affect healthcare access, financial opportunities, hiring, and public services. In these contexts, accuracy is only one piece of the puzzle - models must also be fair to different groups, protect individual privacy, and remain accountable to stakeholders. Achieving all three is difficult: differential privacy can unintentionally worsen disparities, fairness interventions often rely on sensitive data that privacy restricts, and automated pipelines ignore that fairness is ultimately a human and contextual judgment. We introduce FAIRPLAI (Fair and Private Learning with Active Human Influence), a practical framework that integrates human oversight into the design and deployment of machine learning systems. FAIRPLAI works in three ways: (1) it constructs privacy-fairness frontiers that make trade-offs between accuracy, privacy guarantees, and group outcomes transparent; (2) it enables interactive stakeholder input, allowing decision-makers to select fairness criteria and operating points that reflect their domain needs; and (3) it embeds a differentially private auditing loop, giving humans the ability to review explanations and edge cases without compromising individual data security. Applied to benchmark datasets, FAIRPLAI consistently preserves strong privacy protections while reducing fairness disparities relative to automated baselines. More importantly, it provides a straightforward, interpretable process for practitioners to manage competing demands of accuracy, privacy, and fairness in socially impactful applications. By embedding human judgment where it matters most, FAIRPLAI offers a pathway to machine learning systems that are effective, responsible, and trustworthy in practice. GitHub: https://github.com/Li1Davey/Fairplai", "date": "2025-11-11", "categories": []}
{"id": "2511.09039", "title": "Fairness-Aware Few-Shot Learning for Audio-Visual Stress Detection", "abstract": "Fairness in AI-driven stress detection is critical for equitable mental healthcare, yet existing models frequently exhibit gender bias, particularly in data-scarce scenarios. To address this, we propose FairM2S, a fairness-aware meta-learning framework for stress detection leveraging audio-visual data. FairM2S integrates Equalized Odds constraints during both meta-training and adaptation phases, employing adversarial gradient masking and fairness-constrained meta-updates to effectively mitigate bias. Evaluated against five state-of-the-art baselines, FairM2S achieves 78.1% accuracy while reducing the Equal Opportunity to 0.06, demonstrating substantial fairness gains. We also release SAVSD, a smartphone-captured dataset with gender annotations, designed to support fairness research in low-resource, real-world contexts. Together, these contributions position FairM2S as a state-of-the-art approach for equitable and scalable few-shot stress detection in mental health AI. We release our dataset and FairM2S publicly with this paper.", "date": "2025-11-12", "categories": []}
{"id": "2511.07700", "title": "On the Role of Calibration in Benchmarking Algorithmic Fairness for Skin Cancer Detection", "abstract": "Artificial Intelligence (AI) models have demonstrated expert-level performance in melanoma detection, yet their clinical adoption is hindered by performance disparities across demographic subgroups such as gender, race, and age. Previous efforts to benchmark the performance of AI models have primarily focused on assessing model performance using group fairness metrics that rely on the Area Under the Receiver Operating Characteristic curve (AUROC), which does not provide insights into a model's ability to provide accurate estimates. In line with clinical assessments, this paper addresses this gap by incorporating calibration as a complementary benchmarking metric to AUROC-based fairness metrics. Calibration evaluates the alignment between predicted probabilities and observed event rates, offering deeper insights into subgroup biases. We assess the performance of the leading skin cancer detection algorithm of the ISIC 2020 Challenge on the ISIC 2020 Challenge dataset and the PROVE-AI dataset, and compare it with the second and third place models, focusing on subgroups defined by sex, race (Fitzpatrick Skin Tone), and age. Our findings reveal that while existing models enhance discriminative accuracy, they often over-diagnose risk and exhibit calibration issues when applied to new datasets. This study underscores the necessity for comprehensive model auditing strategies and extensive metadata collection to achieve equitable AI-driven healthcare solutions. All code is publicly available at https://github.com/bdominique/testing_strong_calibration.", "date": "2025-11-10", "categories": []}
{"id": "2511.08331", "title": "Adversarial Bias: Data Poisoning Attacks on Fairness", "abstract": "With the growing adoption of AI and machine learning systems in real-world applications, ensuring their fairness has become increasingly critical. The majority of the work in algorithmic fairness focus on assessing and improving the fairness of machine learning systems. There is relatively little research on fairness vulnerability, i.e., how an AI system's fairness can be intentionally compromised. In this work, we first provide a theoretical analysis demonstrating that a simple adversarial poisoning strategy is sufficient to induce maximally unfair behavior in naive Bayes classifiers. Our key idea is to strategically inject a small fraction of carefully crafted adversarial data points into the training set, biasing the model's decision boundary to disproportionately affect a protected group while preserving generalizable performance. To illustrate the practical effectiveness of our method, we conduct experiments across several benchmark datasets and models. We find that our attack significantly outperforms existing methods in degrading fairness metrics across multiple models and datasets, often achieving substantially higher levels of unfairness with a comparable or only slightly worse impact on accuracy. Notably, our method proves effective on a wide range of models, in contrast to prior work, demonstrating a robust and potent approach to compromising the fairness of machine learning systems.", "date": "2025-11-11", "categories": []}
{"id": "2511.06293", "title": "Achieving Fairness Without Harm via Selective Demographic Experts", "abstract": "As machine learning systems become increasingly integrated into human-centered domains such as healthcare, ensuring fairness while maintaining high predictive performance is critical. Existing bias mitigation techniques often impose a trade-off between fairness and accuracy, inadvertently degrading performance for certain demographic groups. In high-stakes domains like clinical diagnosis, such trade-offs are ethically and practically unacceptable. In this study, we propose a fairness-without-harm approach by learning distinct representations for different demographic groups and selectively applying demographic experts consisting of group-specific representations and personalized classifiers through a no-harm constrained selection. We evaluate our approach on three real-world medical datasets -- covering eye disease, skin cancer, and X-ray diagnosis -- as well as two face datasets. Extensive empirical results demonstrate the effectiveness of our approach in achieving fairness without harm.", "date": "2025-11-09", "categories": []}
{"id": "2511.06592", "title": "MedVoiceBias: A Controlled Study of Audio LLM Behavior in Clinical Decision-Making", "abstract": "As large language models transition from text-based interfaces to audio interactions in clinical settings, they might introduce new vulnerabilities through paralinguistic cues in audio. We evaluated these models on 170 clinical cases, each synthesized into speech from 36 distinct voice profiles spanning variations in age, gender, and emotion. Our findings reveal a severe modality bias: surgical recommendations for audio inputs varied by as much as 35% compared to identical text-based inputs, with one model providing 80% fewer recommendations. Further analysis uncovered age disparities of up to 12% between young and elderly voices, which persisted in most models despite chain-of-thought prompting. While explicit reasoning successfully eliminated gender bias, the impact of emotion was not detected due to poor recognition performance. These results demonstrate that audio LLMs are susceptible to making clinical decisions based on a patient's voice characteristics rather than medical evidence, a flaw that risks perpetuating healthcare disparities. We conclude that bias-aware architectures are essential and urgently needed before the clinical deployment of these models.", "date": "2025-11-10", "categories": []}
{"id": "2511.07032", "title": "Fair Bayesian Data Selection via Generalized Discrepancy Measures", "abstract": "Fairness concerns are increasingly critical as machine learning models are deployed in high-stakes applications. While existing fairness-aware methods typically intervene at the model level, they often suffer from high computational costs, limited scalability, and poor generalization. To address these challenges, we propose a Bayesian data selection framework that ensures fairness by aligning group-specific posterior distributions of model parameters and sample weights with a shared central distribution. Our framework supports flexible alignment via various distributional discrepancy measures, including Wasserstein distance, maximum mean discrepancy, and $f$-divergence, allowing geometry-aware control without imposing explicit fairness constraints. This data-centric approach mitigates group-specific biases in training data and improves fairness in downstream tasks, with theoretical guarantees. Experiments on benchmark datasets show that our method consistently outperforms existing data selection and model-based fairness methods in both fairness and accuracy.", "date": "2025-11-10", "categories": []}
{"id": "2406.17433", "title": "Mind the Graph When Balancing Data for Fairness or Robustness", "abstract": "Failures of fairness or robustness in machine learning predictive settings can be due to undesired dependencies between covariates, outcomes and auxiliary factors of variation. A common strategy to mitigate these failures is data balancing, which attempts to remove those undesired dependencies. In this work, we define conditions on the training distribution for data balancing to lead to fair or robust models. Our results display that, in many cases, the balanced distribution does not correspond to selectively removing the undesired dependencies in a causal graph of the task, leading to multiple failure modes and even interference with other mitigation techniques such as regularization. Overall, our results highlight the importance of taking the causal graph into account before performing data balancing.", "date": "2024-06-25", "categories": []}
{"id": "2506.03777", "title": "FedFACT: A Provable Framework for Controllable Group-Fairness Calibration in Federated Learning", "abstract": "With the emerging application of Federated Learning (FL) in decision-making scenarios, it is imperative to regulate model fairness to prevent disparities across sensitive groups (e.g., female, male). Current research predominantly focuses on two concepts of group fairness within FL: Global Fairness (overall model disparity across all clients) and Local Fairness (the disparity within each client). However, the non-decomposable, non-differentiable nature of fairness criteria poses two fundamental, unresolved challenges for fair FL: (i) Harmonizing global and local fairness, especially in multi-class setting; (ii) Enabling a controllable, optimal accuracy-fairness trade-off. To tackle these challenges, we propose a novel controllable federated group-fairness calibration framework, named FedFACT. FedFACT identifies the Bayes-optimal classifiers under both global and local fairness constraints, yielding models with minimal performance decline while guaranteeing fairness. Building on the characterization of the optimal fair classifiers, we reformulate fair federated learning as a personalized cost-sensitive learning problem for in-processing and a bi-level optimization for post-processing. Theoretically, we provide convergence and generalization guarantees for FedFACT to approach the near-optimal accuracy under given fairness levels. Extensive experiments on multiple datasets across various data heterogeneity demonstrate that FedFACT consistently outperforms baselines in balancing accuracy and global-local fairness.", "date": "2025-06-04", "categories": []}
{"id": "2511.04907", "title": "Efficient Swap Multicalibration of Elicitable Properties", "abstract": "Multicalibration [HJKRR18] is an algorithmic fairness perspective that demands that the predictions of a predictor are correct conditional on themselves and membership in a collection of potentially overlapping subgroups of a population. The work of [NR23] established a surprising connection between multicalibration for an arbitrary property $\\Gamma$ (e.g., mean or median) and property elicitation: a property $\\Gamma$ can be multicalibrated if and only if it is elicitable, where elicitability is the notion that the true property value of a distribution can be obtained by solving a regression problem over the distribution. In the online setting, [NR23] proposed an inefficient algorithm that achieves $\\sqrt T$ $\\ell_2$-multicalibration error for a hypothesis class of group membership functions and an elicitable property $\\Gamma$, after $T$ rounds of interaction between a forecaster and adversary.\n  In this paper, we generalize multicalibration for an elicitable property $\\Gamma$ from group membership functions to arbitrary bounded hypothesis classes and introduce a stronger notion -- swap multicalibration, following [GKR23]. Subsequently, we propose an oracle-efficient algorithm which, when given access to an online agnostic learner, achieves $T^{1/(r+1)}$ $\\ell_r$-swap multicalibration error with high probability (for $r\\ge2$) for a hypothesis class with bounded sequential Rademacher complexity and an elicitable property $\\Gamma$. For the special case of $r=2$, this implies an oracle-efficient algorithm that achieves $T^{1/3}$ $\\ell_2$-swap multicalibration error, which significantly improves on the previously established bounds for the problem [NR23, GMS25, LSS25a], and completely resolves an open question raised in [GJRR24] on the possibility of an oracle-efficient algorithm that achieves $\\sqrt{T}$ $\\ell_2$-mean multicalibration error by answering it in a strongly affirmative sense.", "date": "2025-11-07", "categories": []}
{"id": "2406.17737", "title": "LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users", "abstract": "While state-of-the-art large language models (LLMs) have shown impressive performance on many tasks, there has been extensive research on undesirable model behavior such as hallucinations and bias. In this work, we investigate how the quality of LLM responses changes in terms of information accuracy, truthfulness, and refusals depending on three user traits: English proficiency, education level, and country of origin. We present extensive experimentation on three state-of-the-art LLMs and two different datasets targeting truthfulness and factuality. Our findings suggest that undesirable behaviors in state-of-the-art LLMs occur disproportionately more for users with lower English proficiency, of lower education status, and originating from outside the US, rendering these models unreliable sources of information towards their most vulnerable users.", "date": "2024-06-25", "categories": []}
{"id": "2501.08441", "title": "Religious Bias Landscape in Language and Text-to-Image Models: Analysis, Detection, and Debiasing Strategies", "abstract": "Note: This paper includes examples of potentially offensive content related to religious bias, presented solely for academic purposes. The widespread adoption of language models highlights the need for critical examinations of their inherent biases, particularly concerning religion. This study systematically investigates religious bias in both language models and text-to-image generation models, analyzing both open-source and closed-source systems. We construct approximately 400 unique, naturally occurring prompts to probe language models for religious bias across diverse tasks, including mask filling, prompt completion, and image generation. Our experiments reveal concerning instances of underlying stereotypes and biases associated disproportionately with certain religions. Additionally, we explore cross-domain biases, examining how religious bias intersects with demographic factors such as gender, age, and nationality. This study further evaluates the effectiveness of targeted debiasing techniques by employing corrective prompts designed to mitigate the identified biases. Our findings demonstrate that language models continue to exhibit significant biases in both text and image generation tasks, emphasizing the urgent need to develop fairer language models to achieve global acceptability.", "date": "2025-01-14", "categories": []}
{"id": "2506.10564", "title": "Balancing Tails when Comparing Distributions: Comprehensive Equity Index (CEI) with Application to Bias Evaluation in Operational Face Biometrics", "abstract": "Demographic bias in high-performance face recognition (FR) systems often eludes detection by existing metrics, especially with respect to subtle disparities in the tails of the score distribution. We introduce the Comprehensive Equity Index (CEI), a novel metric designed to address this limitation. CEI uniquely analyzes genuine and impostor score distributions separately, enabling a configurable focus on tail probabilities while also considering overall distribution shapes. Our extensive experiments (evaluating state-of-the-art FR systems, intentionally biased models, and diverse datasets) confirm CEI's superior ability to detect nuanced biases where previous methods fall short. Furthermore, we present CEI^A, an automated version of the metric that enhances objectivity and simplifies practical application. CEI provides a robust and sensitive tool for operational FR fairness assessment. The proposed methods have been developed particularly for bias evaluation in face biometrics but, in general, they are applicable for comparing statistical distributions in any problem where one is interested in analyzing the distribution tails.", "date": "2025-06-12", "categories": []}
{"id": "2407.18745", "title": "FairAIED: Navigating Fairness, Bias, and Ethics in Educational AI Applications", "abstract": "The integration of AI in education holds immense potential for personalizing learning experiences and transforming instructional practices. However, AI systems can inadvertently encode and amplify biases present in educational data, leading to unfair or discriminatory outcomes. As researchers have sought to understand and mitigate these biases, a growing body of work has emerged examining fairness in educational AI. These studies, though expanding rapidly, remain fragmented due to differing assumptions, methodologies, and application contexts. Moreover, existing surveys either focus on algorithmic fairness without an educational setting or emphasize educational methods while overlooking fairness. To this end, this survey provides a comprehensive systematic review of algorithmic fairness within educational AI, explicitly bridging the gap between technical fairness research and educational applications. We integrate multiple dimensions, including bias sources, fairness definitions, mitigation strategies, evaluation resources, and ethical considerations, into a harmonized, education-centered framework. In addition, we explicitly examine practical challenges such as censored or partially observed learning outcomes and the persistent difficulty in quantifying and managing the trade-off between fairness and predictive utility, enhancing the applicability of fairness frameworks to real-world educational AI systems. Finally, we outline an emerging pathway toward fair AI-driven education and by situating these technologies and practical insights within broader educational and ethical contexts, this review establishes a comprehensive foundation for advancing fairness, accountability, and inclusivity in the field of AI education.", "date": "2024-07-26", "categories": []}
{"id": "2502.11559", "title": "Auto-Search and Refinement: An Automated Framework for Gender Bias Mitigation in Large Language Models", "abstract": "Pre-training large language models (LLMs) on vast text corpora enhances natural language processing capabilities but risks encoding social biases, particularly gender bias. While parameter-modification methods like fine-tuning mitigate bias, they are resource-intensive, unsuitable for closed-source models, and lack adaptability to evolving societal norms. Instruction-based approaches offer flexibility but often compromise task performance. To address these limitations, we propose $\\textbf{FaIRMaker}$, an automated and model-independent framework that employs an $\\textbf{auto-search and refinement}$ paradigm to adaptively generate Fairwords, which act as instructions integrated into input queries to reduce gender bias and enhance response quality. Extensive experiments demonstrate that FaIRMaker automatically searches for and dynamically refines Fairwords, effectively mitigating gender bias while preserving task integrity and ensuring compatibility with both API-based and open-source LLMs.", "date": "2025-02-17", "categories": []}
{"id": "2506.23467", "title": "AdFair-CLIP: Adversarial Fair Contrastive Language-Image Pre-training for Chest X-rays", "abstract": "Contrastive Language-Image Pre-training (CLIP) models have demonstrated superior performance across various visual tasks including medical image classification. However, fairness concerns, including demographic biases, have received limited attention for CLIP models. This oversight leads to critical issues, particularly those related to race and gender, resulting in disparities in diagnostic outcomes and reduced reliability for underrepresented groups. To address these challenges, we introduce AdFair-CLIP, a novel framework employing adversarial feature intervention to suppress sensitive attributes, thereby mitigating spurious correlations and improving prediction fairness. We conduct comprehensive experiments on chest X-ray (CXR) datasets, and show that AdFair-CLIP significantly enhances both fairness and diagnostic accuracy, while maintaining robust generalization in zero-shot and few-shot scenarios. These results establish new benchmarks for fairness-aware learning in CLIP-based medical diagnostic models, particularly for CXR analysis.", "date": "2025-06-30", "categories": []}
{"id": "2508.15499", "title": "Let's Grow an Unbiased Community: Guiding the Fairness of Graphs via New Links", "abstract": "Graph Neural Networks (GNNs) have achieved remarkable success across diverse applications. However, due to the biases in the graph structures, graph neural networks face significant challenges in fairness. Although the original user graph structure is generally biased, it is promising to guide these existing structures toward unbiased ones by introducing new links. The fairness guidance via new links could foster unbiased communities, thereby enhancing fairness in downstream applications. To address this issue, we propose a novel framework named FairGuide. Specifically, to ensure fairness in downstream tasks trained on fairness-guided graphs, we introduce a differentiable community detection task as a pseudo downstream task. Our theoretical analysis further demonstrates that optimizing fairness within this pseudo task effectively enhances structural fairness, promoting fairness generalization across diverse downstream applications. Moreover, FairGuide employs an effective strategy which leverages meta-gradients derived from the fairness-guidance objective to identify new links that significantly enhance structural fairness. Extensive experimental results demonstrate the effectiveness and generalizability of our proposed method across a variety of graph-based fairness tasks.", "date": "2025-08-21", "categories": []}
{"id": "2511.00079", "title": "flowengineR: A Modular and Extensible Framework for Fair and Reproducible Workflow Design in R", "abstract": "flowengineR is an R package designed to provide a modular and extensible framework for building reproducible algorithmic workflows for general-purpose machine learning pipelines. It is motivated by the rapidly evolving field of algorithmic fairness where new metrics, mitigation strategies, and machine learning methods continuously emerge. A central challenge in fairness, but also far beyond, is that existing toolkits either focus narrowly on single interventions or treat reproducibility and extensibility as secondary considerations rather than core design principles. flowengineR addresses this by introducing a unified architecture of standardized engines for data splitting, execution, preprocessing, training, inprocessing, postprocessing, evaluation, and reporting. Each engine encapsulates one methodological task yet communicates via a lightweight interface, ensuring workflows remain transparent, auditable, and easily extensible. Although implemented in R, flowengineR builds on ideas from workflow languages (CWL, YAWL), graph-oriented visual programming languages (KNIME), and R frameworks (BatchJobs, batchtools). Its emphasis, however, is less on orchestrating engines for resilient parallel execution but rather on the straightforward setup and management of distinct engines as data structures. This orthogonalization enables distributed responsibilities, independent development, and streamlined integration. In fairness context, by structuring fairness methods as interchangeable engines, flowengineR lets researchers integrate, compare, and evaluate interventions across the modeling pipeline. At the same time, the architecture generalizes to explainability, robustness, and compliance metrics without core modifications. While motivated by fairness, it ultimately provides a general infrastructure for any workflow context where reproducibility, transparency, and extensibility are essential.", "date": "2025-10-29", "categories": []}
{"id": "2511.00359", "title": "Toward Unifying Group Fairness Evaluation from a Sparsity Perspective", "abstract": "Ensuring algorithmic fairness remains a significant challenge in machine learning, particularly as models are increasingly applied across diverse domains. While numerous fairness criteria exist, they often lack generalizability across different machine learning problems. This paper examines the connections and differences among various sparsity measures in promoting fairness and proposes a unified sparsity-based framework for evaluating algorithmic fairness. The framework aligns with existing fairness criteria and demonstrates broad applicability to a wide range of machine learning tasks. We demonstrate the effectiveness of the proposed framework as an evaluation metric through extensive experiments on a variety of datasets and bias mitigation methods. This work provides a novel perspective to algorithmic fairness by framing it through the lens of sparsity and social equity, offering potential for broader impact on fairness research and applications.", "date": "2025-11-01", "categories": []}
{"id": "2511.00854", "title": "TriCon-Fair: Triplet Contrastive Learning for Mitigating Social Bias in Pre-trained Language Models", "abstract": "The increasing utilization of large language models raises significant concerns about the propagation of social biases, which may result in harmful and unfair outcomes. However, existing debiasing methods treat the biased and unbiased samples independently, thus ignoring their mutual relationship. This oversight enables a hidden negative-positive coupling, where improvements for one group inadvertently compromise the other, allowing residual social bias to persist. In this paper, we introduce TriCon-Fair, a contrastive learning framework that employs a decoupled loss that combines triplet and language modeling terms to eliminate positive-negative coupling. Our TriCon-Fair assigns each anchor an explicitly biased negative and an unbiased positive, decoupling the push-pull dynamics and avoiding positive-negative coupling, and jointly optimizes a language modeling (LM) objective to preserve general capability. Experimental results demonstrate that TriCon-Fair reduces discriminatory output beyond existing debiasing baselines while maintaining strong downstream performance. This suggests that our proposed TriCon-Fair offers a practical and ethical solution for sensitive NLP applications.", "date": "2025-11-02", "categories": []}
{"id": "2511.00875", "title": "Controlling Gender Bias in Retrieval via a Backpack Architecture", "abstract": "The presence of social biases in large language models (LLMs) has become a significant concern in AI research. These biases, often embedded in training data, can perpetuate harmful stereotypes and distort decision-making processes. When LLMs are integrated into ranking systems, they can propagate these biases, leading to unfair outcomes in critical applications such as search engines and recommendation systems. Backpack Language Models, unlike traditional transformer-based models that treat text sequences as monolithic structures, generate outputs as weighted combinations of non-contextual, learned word aspects, also known as senses. Leveraging this architecture, we propose a framework for debiasing ranking tasks. Our experimental results show that this framework effectively mitigates gender bias in text retrieval and ranking with minimal degradation in performance.", "date": "2025-11-02", "categories": []}
{"id": "2511.01054", "title": "MedEqualizer: A Framework Investigating Bias in Synthetic Medical Data and Mitigation via Augmentation", "abstract": "Synthetic healthcare data generation presents a viable approach to enhance data accessibility and support research by overcoming limitations associated with real-world medical datasets. However, ensuring fairness across protected attributes in synthetic data is critical to avoid biased or misleading results in clinical research and decision-making. In this study, we assess the fairness of synthetic data generated by multiple generative adversarial network (GAN)-based models using the MIMIC-III dataset, with a focus on representativeness across protected demographic attributes. We measure subgroup representation using the logarithmic disparity metric and observe significant imbalances, with many subgroups either underrepresented or overrepresented in the synthetic data, compared to the real data. To mitigate these disparities, we introduce MedEqualizer, a model-agnostic augmentation framework that enriches the underrepresented subgroups prior to synthetic data generation. Our results show that MedEqualizer significantly improves demographic balance in the resulting synthetic datasets, offering a viable path towards more equitable and representative healthcare data synthesis.", "date": "2025-11-02", "categories": []}
{"id": "2511.01069", "title": "Happiness as a Measure of Fairness", "abstract": "In this paper, we propose a novel fairness framework grounded in the concept of happiness, a measure of the utility each group gains fromdecisionoutcomes. Bycapturingfairness through this intuitive lens, we not only offer a more human-centered approach, but also one that is mathematically rigorous: In order to compute the optimal, fair post-processing strategy, only a linear program needs to be solved. This makes our method both efficient and scalable with existing optimization tools. Furthermore, it unifies and extends several well-known fairness definitions, and our empirical results highlight its practical strengths across diverse scenarios.", "date": "2025-11-02", "categories": []}
{"id": "2412.11060", "title": "DPA: A one-stop metric to measure bias amplification in classification datasets", "abstract": "Most ML datasets today contain biases. When we train models on these datasets, they often not only learn these biases but can worsen them -- a phenomenon known as bias amplification. Several co-occurrence-based metrics have been proposed to measure bias amplification in classification datasets. They measure bias amplification between a protected attribute (e.g., gender) and a task (e.g., cooking). These metrics also support fine-grained bias analysis by identifying the direction in which a model amplifies biases. However, co-occurrence-based metrics have limitations -- some fail to measure bias amplification in balanced datasets, while others fail to measure negative bias amplification. To solve these issues, recent work proposed a predictability-based metric called leakage amplification (LA). However, LA cannot identify the direction in which a model amplifies biases. We propose Directional Predictability Amplification (DPA), a predictability-based metric that is (1) directional, (2) works with balanced and unbalanced datasets, and (3) correctly identifies positive and negative bias amplification. DPA eliminates the need to evaluate models on multiple metrics to verify these three aspects. DPA also improves over prior predictability-based metrics like LA: it is less sensitive to the choice of attacker function (a hyperparameter in predictability-based metrics), reports scores within a bounded range, and accounts for dataset bias by measuring relative changes in predictability. Our experiments on well-known datasets like COMPAS (a tabular dataset), COCO, and ImSitu (image datasets) show that DPA is the most reliable metric to measure bias amplification in classification problems. To compare DPA with existing bias amplification metrics, we released a one-stop library of major bias amplification metrics at https://github.com/kerner-lab/Bias-Amplification.", "date": "2024-12-15", "categories": []}
{"id": "2501.10484", "title": "Bias in Decision-Making for AI's Ethical Dilemmas: A Comparative Study of ChatGPT and Claude", "abstract": "Recent advances in Large Language Models (LLMs) have enabled human-like responses across various tasks, raising questions about their ethical decision-making capabilities and potential biases. This study systematically evaluates how nine popular LLMs (both open-source and closed-source) respond to ethical dilemmas involving protected attributes. Across 50,400 trials spanning single and intersectional attribute combinations in four dilemma scenarios (protective vs. harmful), we assess models' ethical preferences, sensitivity, stability, and clustering patterns. Results reveal significant biases in protected attributes in all models, with differing preferences depending on model type and dilemma context. Notably, open-source LLMs show stronger preferences for marginalized groups and greater sensitivity in harmful scenarios, while closed-source models are more selective in protective situations and tend to favor mainstream groups. We also find that ethical behavior varies across dilemma types: LLMs maintain consistent patterns in protective scenarios but respond with more diverse and cognitively demanding decisions in harmful ones. Furthermore, models display more pronounced ethical tendencies under intersectional conditions than in single-attribute settings, suggesting that complex inputs reveal deeper biases. These findings highlight the need for multi-dimensional, context-aware evaluation of LLMs' ethical behavior and offer a systematic evaluation and approach to understanding and addressing fairness in LLM decision-making.", "date": "2025-01-17", "categories": []}
{"id": "2503.00333", "title": "More of the Same: Persistent Representational Harms Under Increased Representation", "abstract": "To recognize and mitigate the harms of generative AI systems, it is crucial to consider whether and how different societal groups are represented by these systems. A critical gap emerges when naively measuring or improving who is represented, as this does not consider how people are represented. In this work, we develop GAS(P), an evaluation methodology for surfacing distribution-level group representational biases in generated text, tackling the setting where groups are unprompted (i.e., groups are not specified in the input to generative systems). We apply this novel methodology to investigate gendered representations in occupations across state-of-the-art large language models. We show that, even though the gender distribution when models are prompted to generate biographies leads to a large representation of women, even representational biases persist in how different genders are represented. Our evaluation methodology reveals that there are statistically significant distribution-level differences in the word choice used to describe biographies and personas of different genders across occupations, and we show that many of these differences are associated with representational harms and stereotypes. Our empirical findings caution that naively increasing (unprompted) representation may inadvertently proliferate representational biases, and our proposed evaluation methodology enables systematic and rigorous measurement of the problem.", "date": "2025-03-01", "categories": []}
{"id": "2505.07985", "title": "Fair Play for Individuals, Foul Play for Groups? Auditing Anonymization's Impact on ML Fairness", "abstract": "Machine learning (ML) algorithms are heavily based on the availability of training data, which, depending on the domain, often includes sensitive information about data providers. This raises critical privacy concerns. Anonymization techniques have emerged as a practical solution to address these issues by generalizing features or suppressing data to make it more difficult to accurately identify individuals. Although recent studies have shown that privacy-enhancing technologies can influence ML predictions across different subgroups, thus affecting fair decision-making, the specific effects of anonymization techniques, such as $k$-anonymity, $\\ell$-diversity, and $t$-closeness, on ML fairness remain largely unexplored. In this work, we systematically audit the impact of anonymization techniques on ML fairness, evaluating both individual and group fairness. Our quantitative study reveals that anonymization can degrade group fairness metrics by up to fourfold. Conversely, similarity-based individual fairness metrics tend to improve under stronger anonymization, largely as a result of increased input homogeneity. By analyzing varying levels of anonymization across diverse privacy settings and data distributions, this study provides critical insights into the trade-offs between privacy, fairness, and utility, offering actionable guidelines for responsible AI development. Our code is publicly available at: https://github.com/hharcolezi/anonymity-impact-fairness.", "date": "2025-05-12", "categories": []}
{"id": "2510.26940", "title": "Mind the Gaps: Auditing and Reducing Group Inequity in Large-Scale Mobility Prediction", "abstract": "Next location prediction underpins a growing number of mobility, retail, and public-health applications, yet its societal impacts remain largely unexplored. In this paper, we audit state-of-the-art mobility prediction models trained on a large-scale dataset, highlighting hidden disparities based on user demographics. Drawing from aggregate census data, we compute the difference in predictive performance on racial and ethnic user groups and show a systematic disparity resulting from the underlying dataset, resulting in large differences in accuracy based on location and user groups. To address this, we propose Fairness-Guided Incremental Sampling (FGIS), a group-aware sampling strategy designed for incremental data collection settings. Because individual-level demographic labels are unavailable, we introduce Size-Aware K-Means (SAKM), a clustering method that partitions users in latent mobility space while enforcing census-derived group proportions. This yields proxy racial labels for the four largest groups in the state: Asian, Black, Hispanic, and White. Built on these labels, our sampling algorithm prioritizes users based on expected performance gains and current group representation. This method incrementally constructs training datasets that reduce demographic performance gaps while preserving overall accuracy. Our method reduces total disparity between groups by up to 40\\% with minimal accuracy trade-offs, as evaluated on a state-of-art MetaPath2Vec model and a transformer-encoder model. Improvements are most significant in early sampling stages, highlighting the potential for fairness-aware strategies to deliver meaningful gains even in low-resource settings. Our findings expose structural inequities in mobility prediction pipelines and demonstrate how lightweight, data-centric interventions can improve fairness with little added complexity, especially for low-data applications.", "date": "2025-10-30", "categories": []}
{"id": "2510.27087", "title": "Characterizing Selective Refusal Bias in Large Language Models", "abstract": "Safety guardrails in large language models(LLMs) are developed to prevent malicious users from generating toxic content at a large scale. However, these measures can inadvertently introduce or reflect new biases, as LLMs may refuse to generate harmful content targeting some demographic groups and not others. We explore this selective refusal bias in LLM guardrails through the lens of refusal rates of targeted individual and intersectional demographic groups, types of LLM responses, and length of generated refusals. Our results show evidence of selective refusal bias across gender, sexual orientation, nationality, and religion attributes. This leads us to investigate additional safety implications via an indirect attack, where we target previously refused groups. Our findings emphasize the need for more equitable and robust performance in safety guardrails across demographic groups.", "date": "2025-10-31", "categories": []}
{"id": "2510.27136", "title": "FairAD: Computationally Efficient Fair Graph Clustering via Algebraic Distance", "abstract": "Due to the growing concern about unsavory behaviors of machine learning models toward certain demographic groups, the notion of 'fairness' has recently drawn much attention from the community, thereby motivating the study of fairness in graph clustering. Fair graph clustering aims to partition the set of nodes in a graph into $k$ disjoint clusters such that the proportion of each protected group within each cluster is consistent with the proportion of that group in the entire dataset. It is, however, computationally challenging to incorporate fairness constraints into existing graph clustering algorithms, particularly for large graphs. To address this problem, we propose FairAD, a computationally efficient fair graph clustering method. It first constructs a new affinity matrix based on the notion of algebraic distance such that fairness constraints are imposed. A graph coarsening process is then performed on this affinity matrix to find representative nodes that correspond to $k$ clusters. Finally, a constrained minimization problem is solved to obtain the solution of fair clustering. Experiment results on the modified stochastic block model and six public datasets show that FairAD can achieve fair clustering while being up to 40 times faster compared to state-of-the-art fair graph clustering algorithms.", "date": "2025-10-31", "categories": []}
{"id": "2504.01223", "title": "Explainable post-training bias mitigation with distribution-based fairness metrics", "abstract": "We develop a novel bias mitigation framework with distribution-based fairness constraints suitable for producing demographically blind and explainable machine-learning models across a wide range of fairness levels. This is accomplished through post-processing, allowing fairer models to be generated efficiently without retraining the underlying model. Our framework, which is based on stochastic gradient descent, can be applied to a wide range of model types, with a particular emphasis on the post-processing of gradient-boosted decision trees. Additionally, we design a broad family of global fairness metrics, along with differentiable and consistent estimators compatible with our framework, building on previous work. We empirically test our methodology on a variety of datasets and compare it with alternative post-processing approaches, including Bayesian search, optimal transport projection, and direct neural network training.", "date": "2025-04-01", "categories": []}
{"id": "2510.25805", "title": "Ideology-Based LLMs for Content Moderation", "abstract": "Large language models (LLMs) are increasingly used in content moderation systems, where ensuring fairness and neutrality is essential. In this study, we examine how persona adoption influences the consistency and fairness of harmful content classification across different LLM architectures, model sizes, and content modalities (language vs. vision). At first glance, headline performance metrics suggest that personas have little impact on overall classification accuracy. However, a closer analysis reveals important behavioral shifts. Personas with different ideological leanings display distinct propensities to label content as harmful, showing that the lens through which a model \"views\" input can subtly shape its judgments. Further agreement analyses highlight that models, particularly larger ones, tend to align more closely with personas from the same political ideology, strengthening within-ideology consistency while widening divergence across ideological groups. To show this effect more directly, we conducted an additional study on a politically targeted task, which confirmed that personas not only behave more coherently within their own ideology but also exhibit a tendency to defend their perspective while downplaying harmfulness in opposing views. Together, these findings highlight how persona conditioning can introduce subtle ideological biases into LLM outputs, raising concerns about the use of AI systems that may reinforce partisan perspectives under the guise of neutrality.", "date": "2025-10-29", "categories": []}
{"id": "2505.09868", "title": "Which Demographic Features Are Relevant for Individual Fairness Evaluation of U.S. Recidivism Risk Assessment Tools?", "abstract": "Despite its constitutional relevance, the technical ``individual fairness'' criterion has not been operationalized in U.S. state or federal statutes/regulations. We conduct a human subjects experiment to address this gap, evaluating which demographic features are relevant for individual fairness evaluation of recidivism risk assessment (RRA) tools. Our analyses conclude that the individual similarity function should consider age and sex, but it should ignore race.", "date": "2025-05-15", "categories": []}
{"id": "2510.13852", "title": "ConsistencyAI: A Benchmark to Assess LLMs' Factual Consistency When Responding to Different Demographic Groups", "abstract": "Is an LLM telling you different facts than it's telling me? This paper introduces ConsistencyAI, an independent benchmark for measuring the factual consistency of large language models (LLMs) for different personas. ConsistencyAI tests whether, when users of different demographics ask identical questions, the model responds with factually inconsistent answers. Designed without involvement from LLM providers, this benchmark offers impartial evaluation and accountability. In our experiment, we queried 19 LLMs with prompts that requested 5 facts for each of 15 topics. We repeated this query 100 times for each LLM, each time adding prompt context from a different persona selected from a subset of personas modeling the general population. We processed the responses into sentence embeddings, computed cross-persona cosine similarity, and computed the weighted average of cross-persona cosine similarity to calculate factual consistency scores. In 100-persona experiments, scores ranged from 0.9065 to 0.7896, and the mean was 0.8656, which we adopt as a benchmark threshold. xAI's Grok-3 is most consistent, while several lightweight models rank lowest. Consistency varies by topic: the job market is least consistent, G7 world leaders most consistent, and issues like vaccines or the Israeli-Palestinian conflict diverge by provider. These results show that both the provider and the topic shape the factual consistency. We release our code and interactive demo to support reproducible evaluation and encourage persona-invariant prompting strategies.", "date": "2025-10-11", "categories": []}
{"id": "2510.25096", "title": "Learning Fair Graph Representations with Multi-view Information Bottleneck", "abstract": "Graph neural networks (GNNs) excel on relational data by passing messages over node features and structure, but they can amplify training data biases, propagating discriminatory attributes and structural imbalances into unfair outcomes. Many fairness methods treat bias as a single source, ignoring distinct attribute and structure effects and leading to suboptimal fairness and utility trade-offs. To overcome this challenge, we propose FairMIB, a multi-view information bottleneck framework designed to decompose graphs into feature, structural, and diffusion views for mitigating complexity biases in GNNs. Especially, the proposed FairMIB employs contrastive learning to maximize cross-view mutual information for bias-free representation learning. It further integrates multi-perspective conditional information bottleneck objectives to balance task utility and fairness by minimizing mutual information with sensitive attributes. Additionally, FairMIB introduces an inverse probability-weighted (IPW) adjacency correction in the diffusion view, which reduces the spread of bias propagation during message passing. Experiments on five real-world benchmark datasets demonstrate that FairMIB achieves state-of-the-art performance across both utility and fairness metrics.", "date": "2025-10-29", "categories": []}
{"id": "2510.23693", "title": "On the Societal Impact of Machine Learning", "abstract": "This PhD thesis investigates the societal impact of machine learning (ML). ML increasingly informs consequential decisions and recommendations, significantly affecting many aspects of our lives. As these data-driven systems are often developed without explicit fairness considerations, they carry the risk of discriminatory effects. The contributions in this thesis enable more appropriate measurement of fairness in ML systems, systematic decomposition of ML systems to anticipate bias dynamics, and effective interventions that reduce algorithmic discrimination while maintaining system utility. I conclude by discussing ongoing challenges and future research directions as ML systems, including generative artificial intelligence, become increasingly integrated into society. This work offers a foundation for ensuring that ML's societal impact aligns with broader social values.", "date": "2025-10-27", "categories": []}
{"id": "2510.23921", "title": "Breaking the Benchmark: Revealing LLM Bias via Minimal Contextual Augmentation", "abstract": "Large Language Models have been shown to demonstrate stereotypical biases in their representations and behavior due to the discriminative nature of the data that they have been trained on. Despite significant progress in the development of methods and models that refrain from using stereotypical information in their decision-making, recent work has shown that approaches used for bias alignment are brittle. In this work, we introduce a novel and general augmentation framework that involves three plug-and-play steps and is applicable to a number of fairness evaluation benchmarks. Through application of augmentation to a fairness evaluation dataset (Bias Benchmark for Question Answering (BBQ)), we find that Large Language Models (LLMs), including state-of-the-art open and closed weight models, are susceptible to perturbations to their inputs, showcasing a higher likelihood to behave stereotypically. Furthermore, we find that such models are more likely to have biased behavior in cases where the target demographic belongs to a community less studied by the literature, underlining the need to expand the fairness and safety research to include more diverse communities.", "date": "2025-10-27", "categories": []}
{"id": "2406.08222", "title": "Refusal as Silence: Gendered Disparities in Vision-Language Model Responses", "abstract": "Refusal behavior by Large Language Models is increasingly visible in content moderation, yet little is known about how refusals vary by the identity of the user making the request. This study investigates refusal as a sociotechnical outcome through a counterfactual persona design that varies gender identity--including male, female, non-binary, and transgender personas--while keeping the classification task and visual input constant. Focusing on a vision-language model (GPT-4V), we examine how identity-based language cues influence refusal in binary gender classification tasks. We find that transgender and non-binary personas experience significantly higher refusal rates, even in non-harmful contexts. Our findings also provide methodological implications for equity audits and content analysis using LLMs. Our findings underscore the importance of modeling identity-driven disparities and caution against uncritical use of AI systems for content coding. This study advances algorithmic fairness by reframing refusal as a communicative act that may unevenly regulate epistemic access and participation.", "date": "2024-06-12", "categories": []}
{"id": "2410.02777", "title": "Secure and Confidential Certificates of Online Fairness", "abstract": "The black-box service model enables ML service providers to serve clients while keeping their intellectual property and client data confidential. Confidentiality is critical for delivering ML services legally and responsibly, but makes it difficult for outside parties to verify important model properties such as fairness. Existing methods that assess model fairness confidentially lack either (i) reliability because they certify fairness with respect to a static set of data, and therefore fail to guarantee fairness in the presence of distribution shift or service provider malfeasance; and/or (ii) scalability due to the computational overhead of confidentiality-preserving cryptographic primitives. We address these problems by introducing online fairness certificates, which verify that a model is fair with respect to data received by the service provider online during deployment. We then present OATH, a deployably efficient and scalable zero-knowledge proof protocol for confidential online group fairness certification. OATH exploits statistical properties of group fairness via a cut-and-choose style protocol, enabling scalability improvements over baselines.", "date": "2024-09-17", "categories": []}
{"id": "2412.00382", "title": "Toward Fair Graph Neural Networks Via Dual-Teacher Knowledge Distillation", "abstract": "Graph Neural Networks (GNNs) have demonstrated strong performance in graph representation learning across various real-world applications. However, they often produce biased predictions caused by sensitive attributes, such as religion or gender, an issue that has been largely overlooked in existing methods. Recently, numerous studies have focused on reducing biases in GNNs. However, these approaches often rely on training with partial data (e.g., using either node features or graph structure alone), which can enhance fairness but frequently compromises model utility due to the limited utilization of available graph information. To address this tradeoff, we propose an effective strategy to balance fairness and utility in knowledge distillation. Specifically, we introduce FairDTD, a novel Fair representation learning framework built on Dual-Teacher Distillation, leveraging a causal graph model to guide and optimize the design of the distillation process. Specifically, FairDTD employs two fairness-oriented teacher models: a feature teacher and a structure teacher, to facilitate dual distillation, with the student model learning fairness knowledge from the teachers while also leveraging full data to mitigate utility loss. To enhance information transfer, we incorporate graph-level distillation to provide an indirect supplement of graph information during training, as well as a node-specific temperature module to improve the comprehensive transfer of fair knowledge. Experiments on diverse benchmark datasets demonstrate that FairDTD achieves optimal fairness while preserving high model utility, showcasing its effectiveness in fair representation learning for GNNs.", "date": "2024-11-30", "categories": []}
{"id": "2505.16291", "title": "Fairness under Competition", "abstract": "Algorithmic fairness has emerged as a central issue in ML, and it has become standard practice to adjust ML algorithms so that they will satisfy fairness requirements such as Equal Opportunity. In this paper we consider the effects of adopting such fair classifiers on the overall level of ecosystem fairness. Specifically, we introduce the study of fairness with competing firms, and demonstrate the failure of fair classifiers in yielding fair ecosystems. Our results quantify the loss of fairness in systems, under a variety of conditions, based on classifiers' correlation and the level of their data overlap. We show that even if competing classifiers are individually fair, the ecosystem's outcome may be unfair; and that adjusting biased algorithms to improve their individual fairness may lead to an overall decline in ecosystem fairness. In addition to these theoretical results, we also provide supporting experimental evidence. Together, our model and results provide a novel and essential call for action.", "date": "2025-05-22", "categories": []}
{"id": "2510.12822", "title": "Evidence Without Injustice: A New Counterfactual Test for Fair Algorithms", "abstract": "The growing philosophical literature on algorithmic fairness has examined statistical criteria such as equalized odds and calibration, causal and counterfactual approaches, and the role of structural and compounding injustices. Yet an important dimension has been overlooked: whether the evidential value of an algorithmic output itself depends on structural injustice. We contrast a predictive policing algorithm, which relies on historical crime data, with a camera-based system that records ongoing offenses, where both are designed to guide police deployment. In evaluating the moral acceptability of acting on a piece of evidence, we must ask not only whether the evidence is probative in the actual world, but also whether it would remain probative in nearby worlds without the relevant injustices. The predictive policing algorithm fails this test, but the camera-based system passes it. When evidence fails the test, it is morally problematic to use it punitively, more so than evidence that passes the test.", "date": "2025-10-10", "categories": []}
{"id": "2510.22084", "title": "Compositional Bias Control in Large Language Models: Preference Learning Fails, Supervision Succeeds", "abstract": "Large Language Models (LLMs) still produce gender-stereotyped language even in occupation-neutral contexts that reflect deep societal biases (Rudinger et al., 2018). To address this, prior work has proposed prompting, constrained decoding (Dathathri et al., 2020; Zhou et al., 2024), post-processing, and fine-tuning-based alignment (Rafailov et al., 2023; Ravfogel et al., 2022). However, the comparative efficacy and learning dynamics remain little understood. We report a comparative analysis of six control techniques for bias mitigation: prompt-only, generate-and-filter, DFA-based Ctrl-G decoding, Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Iterative Nullspace Projection (INLP). We evaluate each method on a compositional constraint task. This task requires generating sentences that contain at least one agentic and one communal descriptor for each of the twenty Winogender-derived occupations. We quantify trade-offs between control strength and naturalness with evaluations of constraint compliance, lexical diversity, and fluency. Our results reveal key contrasts among the methods: SFT achieves 99.87 +- 0.15% compliance and high lexical diversity, while DPO, despite similar training stability, fails at 4.53 +- 0.82%. Ctrl-G guarantees perfect compliance, but at the cost of severely reduced fluency and diversity. Preference-based learning fundamentally differs: it cannot satisfy compositional constraints, as binary preference signals encode ranking, not logical conjunctions. Only explicit positive supervision enables mitigation of compositional biases; preference-based alignment fails to generalize logical structures, underscoring the limitations of preference learning and the necessity of explicit supervision for fair and fluent controlled generation.", "date": "2025-10-24", "categories": []}
{"id": "2510.22209", "title": "Visual Model Selection using Feature Importance Clusters in Fairness-Performance Similarity Optimized Space", "abstract": "In the context of algorithmic decision-making, fair machine learning methods often yield multiple models that balance predictive fairness and performance in varying degrees. This diversity introduces a challenge for stakeholders who must select a model that aligns with their specific requirements and values. To address this, we propose an interactive framework that assists in navigating and interpreting the trade-offs across a portfolio of models. Our approach leverages weakly supervised metric learning to learn a Mahalanobis distance that reflects similarity in fairness and performance outcomes, effectively structuring the feature importance space of the models according to stakeholder-relevant criteria. We then apply clustering technique (k-means) to group models based on their transformed representations of feature importances, allowing users to explore clusters of models with similar predictive behaviors and fairness characteristics. This facilitates informed decision-making by helping users understand how models differ not only in their fairness-performance balance but also in the features that drive their predictions.", "date": "2025-10-25", "categories": []}
{"id": "2510.22363", "title": "Bias Begins with Data: The FairGround Corpus for Robust and Reproducible Research on Algorithmic Fairness", "abstract": "As machine learning (ML) systems are increasingly adopted in high-stakes decision-making domains, ensuring fairness in their outputs has become a central challenge. At the core of fair ML research are the datasets used to investigate bias and develop mitigation strategies. Yet, much of the existing work relies on a narrow selection of datasets--often arbitrarily chosen, inconsistently processed, and lacking in diversity--undermining the generalizability and reproducibility of results.\n  To address these limitations, we present FairGround: a unified framework, data corpus, and Python package aimed at advancing reproducible research and critical data studies in fair ML classification. FairGround currently comprises 44 tabular datasets, each annotated with rich fairness-relevant metadata. Our accompanying Python package standardizes dataset loading, preprocessing, transformation, and splitting, streamlining experimental workflows. By providing a diverse and well-documented dataset corpus along with robust tooling, FairGround enables the development of fairer, more reliable, and more reproducible ML models. All resources are publicly available to support open and collaborative research.", "date": "2025-10-25", "categories": []}
{"id": "2409.17643", "title": "Efficient Fairness-Performance Pareto Front Computation", "abstract": "There is a well known intrinsic trade-off between the fairness of a representation and the performance of classifiers derived from the representation. Due to the complexity of optimisation algorithms in most modern representation learning approaches, for a given method it may be non-trivial to decide whether the obtained fairness-performance curve of the method is optimal, i.e., whether it is close to the true Pareto front for these quantities for the underlying data distribution.\n  In this paper we propose a new method to compute the optimal Pareto front, which does not require the training of complex representation models. We show that optimal fair representations possess several useful structural properties, and that these properties enable a reduction of the computation of the Pareto Front to a compact discrete problem. We then also show that these compact approximating problems can be efficiently solved via off-the shelf concave-convex programming methods.\n  Since our approach is independent of the specific model of representations, it may be used as the benchmark to which representation learning algorithms may be compared. We experimentally evaluate the approach on a number of real world benchmark datasets.", "date": "2024-09-26", "categories": []}
{"id": "2502.09564", "title": "Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing", "abstract": "Deep learning model effectiveness in classification tasks is often challenged by the quality and quantity of training data whenever they are affected by strong spurious correlations between specific attributes and target labels. This results in a form of bias affecting training data, which typically leads to unrecoverable weak generalization in prediction. This paper aims at facing this problem by leveraging bias amplification with generated synthetic data: we introduce Diffusing DeBias (DDB), a novel approach acting as a plug-in for common methods of unsupervised model debiasing exploiting the inherent bias-learning tendency of diffusion models in data generation. Specifically, our approach adopts conditional diffusion models to generate synthetic bias-aligned images, which replace the original training set for learning an effective bias amplifier model that we subsequently incorporate into an end-to-end and a two-step unsupervised debiasing approach. By tackling the fundamental issue of bias-conflicting training samples memorization in learning auxiliary models, typical of this type of techniques, our proposed method beats current state-of-the-art in multiple benchmark datasets, demonstrating its potential as a versatile and effective tool for tackling bias in deep learning models. Code is available at https://github.com/Malga-Vision/DiffusingDeBias", "date": "2025-02-13", "categories": []}
{"id": "2504.14882", "title": "Some Optimizers are More Equal: Understanding the Role of Optimizers in Group Fairness", "abstract": "We study whether and how the choice of optimization algorithm can impact group fairness in deep neural networks. Through stochastic differential equation analysis of optimization dynamics in an analytically tractable setup, we demonstrate that the choice of optimization algorithm indeed influences fairness outcomes, particularly under severe imbalance. Furthermore, we show that when comparing two categories of optimizers, adaptive methods and stochastic methods, RMSProp (from the adaptive category) has a higher likelihood of converging to fairer minima than SGD (from the stochastic category). Building on this insight, we derive two new theoretical guarantees showing that, under appropriate conditions, RMSProp exhibits fairer parameter updates and improved fairness in a single optimization step compared to SGD. We then validate these findings through extensive experiments on three publicly available datasets, namely CelebA, FairFace, and MS-COCO, across different tasks as facial expression recognition, gender classification, and multi-label classification, using various backbones. Considering multiple fairness definitions including equalized odds, equal opportunity, and demographic parity, adaptive optimizers like RMSProp and Adam consistently outperform SGD in terms of group fairness, while maintaining comparable predictive accuracy. Our results highlight the role of adaptive updates as a crucial yet overlooked mechanism for promoting fair outcomes. We release the source code at: https://github.com/Mkolahdoozi/Some-Optimizers-Are-More-Equal.", "date": "2025-04-21", "categories": []}
{"id": "2506.04193", "title": "Understanding challenges to the interpretation of disaggregated evaluations of algorithmic fairness", "abstract": "Disaggregated evaluation across subgroups is critical for assessing the fairness of machine learning models, but its uncritical use can mislead practitioners. We show that equal performance across subgroups is an unreliable measure of fairness when data are representative of the relevant populations but reflective of real-world disparities. Furthermore, when data are not representative due to selection bias, both disaggregated evaluation and alternative approaches based on conditional independence testing may be invalid without explicit assumptions regarding the bias mechanism. We use causal graphical models to characterize fairness properties and metric stability across subgroups under different data generating processes. Our framework suggests complementing disaggregated evaluations with explicit causal assumptions and analysis to control for confounding and distribution shift, including conditional independence testing and weighted performance estimation. These findings have broad implications for how practitioners design and interpret model assessments given the ubiquity of disaggregated evaluation.", "date": "2025-06-04", "categories": []}
{"id": "2509.15759", "title": "On Optimal Steering to Achieve Exact Fairness", "abstract": "To fix the 'bias in, bias out' problem in fair machine learning, it is important to steer feature distributions of data or internal representations of Large Language Models (LLMs) to ideal ones that guarantee group-fair outcomes. Previous work on fair generative models and representation steering could greatly benefit from provable fairness guarantees on the model output. We define a distribution as ideal if the minimizer of any cost-sensitive risk on it is guaranteed to have exact group-fair outcomes (e.g., demographic parity, equal opportunity)-in other words, it has no fairness-utility trade-off. We formulate an optimization program for optimal steering by finding the nearest ideal distribution in KL-divergence, and provide efficient algorithms for it when the underlying distributions come from well-known parametric families (e.g., normal, log-normal). Empirically, our optimal steering techniques on both synthetic and real-world datasets improve fairness without diminishing utility (and sometimes even improve utility). We demonstrate affine steering of LLM representations to reduce bias in multi-class classification, e.g., occupation prediction from a short biography in Bios dataset (De-Arteaga et al.). Furthermore, we steer internal representations of LLMs towards desired outputs so that it works equally well across different groups.", "date": "2025-09-19", "categories": []}
{"id": "2510.21363", "title": "FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models", "abstract": "Text-to-image diffusion models, such as Stable Diffusion, have demonstrated remarkable capabilities in generating high-quality and diverse images from natural language prompts. However, recent studies reveal that these models often replicate and amplify societal biases, particularly along demographic attributes like gender and race. In this paper, we introduce FairImagen (https://github.com/fuzihaofzh/FairImagen), a post-hoc debiasing framework that operates on prompt embeddings to mitigate such biases without retraining or modifying the underlying diffusion model. Our method integrates Fair Principal Component Analysis to project CLIP-based input embeddings into a subspace that minimizes group-specific information while preserving semantic content. We further enhance debiasing effectiveness through empirical noise injection and propose a unified cross-demographic projection method that enables simultaneous debiasing across multiple demographic attributes. Extensive experiments across gender, race, and intersectional settings demonstrate that FairImagen significantly improves fairness with a moderate trade-off in image quality and prompt fidelity. Our framework outperforms existing post-hoc methods and offers a simple, scalable, and model-agnostic solution for equitable text-to-image generation.", "date": "2025-10-24", "categories": []}
{"id": "2412.18810", "title": "FairGen: Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions", "abstract": "While Diffusion Models (DM) exhibit remarkable performance across various image generative tasks, they nonetheless reflect the inherent bias presented in the training set. As DMs are now widely used in real-world applications, these biases could perpetuate a distorted worldview and hinder opportunities for minority groups. Existing methods on debiasing DMs usually requires model retraining with a human-crafted reference dataset or additional classifiers, which suffer from two major limitations: (1) collecting reference datasets causes expensive annotation cost; (2) the debiasing performance is heavily constrained by the quality of the reference dataset or the additional classifier. To address the above limitations, we propose FairGen, a plug-and-play method that learns attribute latent directions in a self-discovering manner, thus eliminating the reliance on such reference dataset. Specifically, FairGen consists of two parts: a set of attribute adapters and a distribution indicator. Each adapter in the set aims to learn an attribute latent direction, and is optimized via noise composition through a self-discovering process. Then, the distribution indicator is multiplied by the set of adapters to guide the generation process towards the prescribed distribution. Our method enables debiasing multiple attributes in DMs simultaneously, while remaining lightweight and easily integrable with other DMs, eliminating the need for retraining. Extensive experiments on debiasing gender, racial, and their intersectional biases show that our method outperforms previous SOTA by a large margin.", "date": "2024-12-25", "categories": []}
{"id": "2505.09131", "title": "Fair Clustering via Alignment", "abstract": "Algorithmic fairness in clustering aims to balance the proportions of instances assigned to each cluster with respect to a given sensitive attribute. While recently developed fair clustering algorithms optimize clustering objectives under specific fairness constraints, their inherent complexity or approximation often results in suboptimal clustering utility or numerical instability in practice. To resolve these limitations, we propose a new fair clustering algorithm based on a novel decomposition of the fair $K$-means clustering objective function. The proposed algorithm, called Fair Clustering via Alignment (FCA), operates by alternately (i) finding a joint probability distribution to align the data from different protected groups, and (ii) optimizing cluster centers in the aligned space. A key advantage of FCA is that it theoretically guarantees approximately optimal clustering utility for any given fairness level without complex constraints, thereby enabling high-utility fair clustering in practice. Experiments show that FCA outperforms existing methods by (i) attaining a superior trade-off between fairness level and clustering utility, and (ii) achieving near-perfect fairness without numerical instability.", "date": "2025-05-14", "categories": []}
{"id": "2510.19257", "title": "FnRGNN: Distribution-aware Fairness in Graph Neural Network", "abstract": "Graph Neural Networks (GNNs) excel at learning from structured data, yet fairness in regression tasks remains underexplored. Existing approaches mainly target classification and representation-level debiasing, which cannot fully address the continuous nature of node-level regression. We propose FnRGNN, a fairness-aware in-processing framework for GNN-based node regression that applies interventions at three levels: (i) structure-level edge reweighting, (ii) representation-level alignment via MMD, and (iii) prediction-level normalization through Sinkhorn-based distribution matching. This multi-level strategy ensures robust fairness under complex graph topologies. Experiments on four real-world datasets demonstrate that FnRGNN reduces group disparities without sacrificing performance. Code is available at https://github.com/sybeam27/FnRGNN.", "date": "2025-10-22", "categories": []}
{"id": "2510.20332", "title": "Bias by Design? How Data Practices Shape Fairness in AI Healthcare Systems", "abstract": "Artificial intelligence (AI) holds great promise for transforming healthcare. However, despite significant advances, the integration of AI solutions into real-world clinical practice remains limited. A major barrier is the quality and fairness of training data, which is often compromised by biased data collection practices. This paper draws on insights from the AI4HealthyAging project, part of Spain's national R&D initiative, where our task was to detect biases during clinical data collection. We identify several types of bias across multiple use cases, including historical, representation, and measurement biases. These biases manifest in variables such as sex, gender, age, habitat, socioeconomic status, equipment, and labeling. We conclude with practical recommendations for improving the fairness and robustness of clinical problem design and data collection. We hope that our findings and experience contribute to guiding future projects in the development of fairer AI systems in healthcare.", "date": "2025-10-23", "categories": []}
{"id": "2510.20482", "title": "Reliable and Reproducible Demographic Inference for Fairness in Face Analysis", "abstract": "Fairness evaluation in face analysis systems (FAS) typically depends on automatic demographic attribute inference (DAI), which itself relies on predefined demographic segmentation. However, the validity of fairness auditing hinges on the reliability of the DAI process. We begin by providing a theoretical motivation for this dependency, showing that improved DAI reliability leads to less biased and lower-variance estimates of FAS fairness. To address this, we propose a fully reproducible DAI pipeline that replaces conventional end-to-end training with a modular transfer learning approach. Our design integrates pretrained face recognition encoders with non-linear classification heads. We audit this pipeline across three dimensions: accuracy, fairness, and a newly introduced notion of robustness, defined via intra-identity consistency. The proposed robustness metric is applicable to any demographic segmentation scheme. We benchmark the pipeline on gender and ethnicity inference across multiple datasets and training setups. Our results show that the proposed method outperforms strong baselines, particularly on ethnicity, which is the more challenging attribute. To promote transparency and reproducibility, we will publicly release the training dataset metadata, full codebase, pretrained models, and evaluation toolkit. This work contributes a reliable foundation for demographic inference in fairness auditing.", "date": "2025-10-23", "categories": []}
{"id": "2510.20629", "title": "Equitable Survival Prediction: A Fairness-Aware Survival Modeling (FASM) Approach", "abstract": "As machine learning models become increasingly integrated into healthcare, structural inequities and social biases embedded in clinical data can be perpetuated or even amplified by data-driven models. In survival analysis, censoring and time dynamics can further add complexity to fair model development. Additionally, algorithmic fairness approaches often overlook disparities in cross-group rankings, e.g., high-risk Black patients may be ranked below lower-risk White patients who do not experience the event of mortality. Such misranking can reinforce biological essentialism and undermine equitable care. We propose a Fairness-Aware Survival Modeling (FASM), designed to mitigate algorithmic bias regarding both intra-group and cross-group risk rankings over time. Using breast cancer prognosis as a representative case and applying FASM to SEER breast cancer data, we show that FASM substantially improves fairness while preserving discrimination performance comparable to fairness-unaware survival models. Time-stratified evaluations show that FASM maintains stable fairness over a 10-year horizon, with the greatest improvements observed during the mid-term of follow-up. Our approach enables the development of survival models that prioritize both accuracy and equity in clinical decision-making, advancing fairness as a core principle in clinical care.", "date": "2025-10-23", "categories": []}
{"id": "2505.15788", "title": "Fair Supervised Learning Through Constraints on Smooth Nonconvex Unfairness-Measure Surrogates", "abstract": "A new strategy for fair supervised machine learning is proposed. The main advantages of the proposed strategy as compared to others in the literature are as follows. (a) We introduce a new smooth nonconvex surrogate to approximate the Heaviside functions involved in discontinuous unfairness measures. The surrogate is based on smoothing methods from the optimization literature, and is new for the fair supervised learning literature. The surrogate is a tight approximation which ensures the trained prediction models are fair, as opposed to other (e.g., convex) surrogates that can fail to lead to a fair prediction model in practice. (b) Rather than rely on regularizers (that lead to optimization problems that are difficult to solve) and corresponding regularization parameters (that can be expensive to tune), we propose a strategy that employs hard constraints so that specific tolerances for unfairness can be enforced without the complications associated with the use of regularization. (c) Our proposed strategy readily allows for constraints on multiple (potentially conflicting) unfairness measures at the same time. Multiple measures can be considered with a regularization approach, but at the cost of having even more difficult optimization problems to solve and further expense for tuning. By contrast, through hard constraints, our strategy leads to optimization models that can be solved tractably with minimal tuning.", "date": "2025-05-21", "categories": []}
{"id": "2510.19331", "title": "Algorithmic Fairness in NLP: Persona-Infused LLMs for Human-Centric Hate Speech Detection", "abstract": "In this paper, we investigate how personalising Large Language Models (Persona-LLMs) with annotator personas affects their sensitivity to hate speech, particularly regarding biases linked to shared or differing identities between annotators and targets. To this end, we employ Google's Gemini and OpenAI's GPT-4.1-mini models and two persona-prompting methods: shallow persona prompting and a deeply contextualised persona development based on Retrieval-Augmented Generation (RAG) to incorporate richer persona profiles. We analyse the impact of using in-group and out-group annotator personas on the models' detection performance and fairness across diverse social groups. This work bridges psychological insights on group identity with advanced NLP techniques, demonstrating that incorporating socio-demographic attributes into LLMs can address bias in automated hate speech detection. Our results highlight both the potential and limitations of persona-based approaches in reducing bias, offering valuable insights for developing more equitable hate speech detection systems.", "date": "2025-10-22", "categories": []}
{"id": "2510.19421", "title": "FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRA", "abstract": "Ensuring fairness in machine learning models is a critical challenge. Existing debiasing methods often compromise performance, rely on static correction strategies, and struggle with data sparsity, particularly within minority groups. Furthermore, their utilization of sensitive attributes is often suboptimal, either depending excessively on complete attribute labeling or disregarding these attributes entirely. To overcome these limitations, we propose FairNet, a novel framework for dynamic, instance-level fairness correction. FairNet integrates a bias detector with conditional low-rank adaptation (LoRA), which enables selective activation of the fairness correction mechanism exclusively for instances identified as biased, and thereby preserve performance on unbiased instances. A key contribution is a new contrastive loss function for training the LoRA module, specifically designed to minimize intra-class representation disparities across different sensitive groups and effectively address underfitting in minority groups. The FairNet framework can flexibly handle scenarios with complete, partial, or entirely absent sensitive attribute labels. Theoretical analysis confirms that, under moderate TPR/FPR for the bias detector, FairNet can enhance the performance of the worst group without diminishing overall model performance, and potentially yield slight performance improvements. Comprehensive empirical evaluations across diverse vision and language benchmarks validate the effectiveness of FairNet.", "date": "2025-10-22", "categories": []}
{"id": "2404.18134", "title": "Learning Fairer Representations with FairVIC", "abstract": "Mitigating bias in automated decision-making systems, particularly in deep learning models, is a critical challenge due to nuanced definitions of fairness, dataset-specific biases, and the inherent trade-off between fairness and accuracy. To address these issues, we introduce FairVIC, an innovative approach that enhances fairness in neural networks by integrating variance, invariance, and covariance terms into the loss function during training. Unlike methods that rely on predefined fairness criteria, FairVIC abstracts fairness concepts to minimise dependency on protected characteristics. We evaluate FairVIC against comparable bias mitigation techniques on benchmark datasets, considering both group and individual fairness, and conduct an ablation study on the accuracy-fairness trade-off. FairVIC demonstrates significant improvements ($\\approx70\\%$) in fairness across all tested metrics without compromising accuracy, thus offering a robust, generalisable solution for fair deep learning across diverse tasks and datasets.", "date": "2024-04-28", "categories": []}
{"id": "2510.18473", "title": "Benchmarking Fairness-aware Graph Neural Networks in Knowledge Graphs", "abstract": "Graph neural networks (GNNs) are powerful tools for learning from graph-structured data but often produce biased predictions with respect to sensitive attributes. Fairness-aware GNNs have been actively studied for mitigating biased predictions. However, no prior studies have evaluated fairness-aware GNNs on knowledge graphs, which are one of the most important graphs in many applications, such as recommender systems. Therefore, we introduce a benchmarking study on knowledge graphs. We generate new graphs from three knowledge graphs, YAGO, DBpedia, and Wikidata, that are significantly larger than the existing graph datasets used in fairness studies. We benchmark inprocessing and preprocessing methods in different GNN backbones and early stopping conditions. We find several key insights: (i) knowledge graphs show different trends from existing datasets; clearer trade-offs between prediction accuracy and fairness metrics than other graphs in fairness-aware GNNs, (ii) the performance is largely affected by not only fairness-aware GNN methods but also GNN backbones and early stopping conditions, and (iii) preprocessing methods often improve fairness metrics, while inprocessing methods improve prediction accuracy.", "date": "2025-10-21", "categories": []}
{"id": "2510.18556", "title": "Building Trust in Clinical LLMs: Bias Analysis and Dataset Transparency", "abstract": "Large language models offer transformative potential for healthcare, yet their responsible and equitable development depends critically on a deeper understanding of how training data characteristics influence model behavior, including the potential for bias. Current practices in dataset curation and bias assessment often lack the necessary transparency, creating an urgent need for comprehensive evaluation frameworks to foster trust and guide improvements. In this study, we present an in-depth analysis of potential downstream biases in clinical language models, with a focus on differential opioid prescription tendencies across diverse demographic groups, such as ethnicity, gender, and age. As part of this investigation, we introduce HC4: Healthcare Comprehensive Commons Corpus, a novel and extensively curated pretraining dataset exceeding 89 billion tokens. Our evaluation leverages both established general benchmarks and a novel, healthcare-specific methodology, offering crucial insights to support fairness and safety in clinical AI applications.", "date": "2025-10-21", "categories": []}
{"id": "2510.18719", "title": "Causally Perturbed Fairness Testing", "abstract": "To mitigate unfair and unethical discrimination over sensitive features (e.g., gender, age, or race), fairness testing plays an integral role in engineering systems that leverage AI models to handle tabular data. A key challenge therein is how to effectively reveal fairness bugs under an intractable sample size using perturbation. Much current work has been focusing on designing the test sample generators, ignoring the valuable knowledge about data characteristics that can help guide the perturbation and hence limiting their full potential. In this paper, we seek to bridge such a gap by proposing a generic framework of causally perturbed fairness testing, dubbed CausalFT. Through causal inference, the key idea of CausalFT is to extract the most directly and causally relevant non-sensitive feature to its sensitive counterpart, which can jointly influence the prediction of the label. Such a causal relationship is then seamlessly injected into the perturbation to guide a test sample generator. Unlike existing generator-level work, CausalFT serves as a higher-level framework that can be paired with diverse base generators. Extensive experiments on 1296 cases confirm that CausalFT can considerably improve arbitrary base generators in revealing fairness bugs over 93% of the cases with acceptable extra runtime overhead. Compared with a state-of-the-art approach that ranks the non-sensitive features solely based on correlation, CausalFT performs significantly better on 64% cases while being much more efficient. Further, CausalFT can better improve bias resilience in nearly all cases.", "date": "2025-10-21", "categories": []}
{"id": "2411.09178", "title": "SAFES: Sequential Privacy and Fairness Enhancing Data Synthesis for Responsible AI", "abstract": "As data-driven and AI-based decision making gains widespread adoption across disciplines, it is crucial that both data privacy and decision fairness are appropriately addressed. Although differential privacy (DP) provides a robust framework for guaranteeing privacy and methods are available to improve fairness, most prior work treats the two concerns separately. Even though there are existing approaches that consider privacy and fairness simultaneously, they typically focus on a single specific learning task, limiting their generalizability. In response, we introduce SAFES, a Sequential PrivAcy and Fairness Enhancing data Synthesis procedure that sequentially combines DP data synthesis with a fairness-aware data preprocessing step. SAFES allows users flexibility in navigating the privacy-fairness-utility trade-offs. We illustrate SAFES with different DP synthesizers and fairness-aware data preprocessing methods and run extensive experiments on multiple real datasets to examine the privacy-fairness-utility trade-offs of synthetic data generated by SAFES. Empirical evaluations demonstrate that for reasonable privacy loss, SAFES-generated synthetic data can achieve significantly improved fairness metrics with relatively low utility loss.", "date": "2024-11-14", "categories": []}
{"id": "2505.13122", "title": "When majority rules, minority loses: bias amplification of gradient descent", "abstract": "Despite growing empirical evidence of bias amplification in machine learning, its theoretical foundations remain poorly understood. We develop a formal framework for majority-minority learning tasks, showing how standard training can favor majority groups and produce stereotypical predictors that neglect minority-specific features. Assuming population and variance imbalance, our analysis reveals three key findings: (i) the close proximity between ``full-data'' and stereotypical predictors, (ii) the dominance of a region where training the entire model tends to merely learn the majority traits, and (iii) a lower bound on the additional training required. Our results are illustrated through experiments in deep learning for tabular and image classification tasks.", "date": "2025-05-19", "categories": []}
{"id": "2505.16115", "title": "A Generic Framework for Conformal Fairness", "abstract": "Conformal Prediction (CP) is a popular method for uncertainty quantification with machine learning models. While conformal prediction provides probabilistic guarantees regarding the coverage of the true label, these guarantees are agnostic to the presence of sensitive attributes within the dataset. In this work, we formalize \\textit{Conformal Fairness}, a notion of fairness using conformal predictors, and provide a theoretically well-founded algorithm and associated framework to control for the gaps in coverage between different sensitive groups. Our framework leverages the exchangeability assumption (implicit to CP) rather than the typical IID assumption, allowing us to apply the notion of Conformal Fairness to data types and tasks that are not IID, such as graph data. Experiments were conducted on graph and tabular datasets to demonstrate that the algorithm can control fairness-related gaps in addition to coverage aligned with theoretical expectations.", "date": "2025-05-22", "categories": []}
{"id": "2510.16056", "title": "Algorithmic Fairness in AI Surrogates for End-of-Life Decision-Making", "abstract": "Artificial intelligence surrogates are systems designed to infer preferences when individuals lose decision-making capacity. Fairness in such systems is a domain that has been insufficiently explored. Traditional algorithmic fairness frameworks are insufficient for contexts where decisions are relational, existential, and culturally diverse. This paper explores an ethical framework for algorithmic fairness in AI surrogates by mapping major fairness notions onto potential real-world end-of-life scenarios. It then examines fairness across moral traditions. The authors argue that fairness in this domain extends beyond parity of outcomes to encompass moral representation, fidelity to the patient's values, relationships, and worldview.", "date": "2025-10-16", "categories": []}
{"id": "2510.16072", "title": "Data-Driven Analysis of Intersectional Bias in Image Classification: A Framework with Bias-Weighted Augmentation", "abstract": "Machine learning models trained on imbalanced datasets often exhibit intersectional biases-systematic errors arising from the interaction of multiple attributes such as object class and environmental conditions. This paper presents a data-driven framework for analyzing and mitigating such biases in image classification. We introduce the Intersectional Fairness Evaluation Framework (IFEF), which combines quantitative fairness metrics with interpretability tools to systematically identify bias patterns in model predictions. Building on this analysis, we propose Bias-Weighted Augmentation (BWA), a novel data augmentation strategy that adapts transformation intensities based on subgroup distribution statistics. Experiments on the Open Images V7 dataset with five object classes demonstrate that BWA improves accuracy for underrepresented class-environment intersections by up to 24 percentage points while reducing fairness metric disparities by 35%. Statistical analysis across multiple independent runs confirms the significance of improvements (p < 0.05). Our methodology provides a replicable approach for analyzing and addressing intersectional biases in image classification systems.", "date": "2025-10-17", "categories": []}
{"id": "2408.09570", "title": "Say My Name: a Model's Bias Discovery Framework", "abstract": "In the last few years, due to the broad applicability of deep learning to downstream tasks and end-to-end training capabilities, increasingly more concerns about potential biases to specific, non-representative patterns have been raised. Many works focusing on unsupervised debiasing usually leverage the tendency of deep models to learn ``easier'' samples, for example by clustering the latent space to obtain bias pseudo-labels. However, the interpretation of such pseudo-labels is not trivial, especially for a non-expert end user, as it does not provide semantic information about the bias features. To address this issue, we introduce ``Say My Name'' (SaMyNa), the first tool to identify biases within deep models semantically. Unlike existing methods, our approach focuses on biases learned by the model. Our text-based pipeline enhances explainability and supports debiasing efforts: applicable during either training or post-hoc validation, our method can disentangle task-related information and proposes itself as a tool to analyze biases. Evaluation on traditional benchmarks demonstrates its effectiveness in detecting biases and even disclaiming them, showcasing its broad applicability for model diagnosis.", "date": "2024-08-18", "categories": []}
{"id": "2510.14669", "title": "Machine Learning and Public Health: Identifying and Mitigating Algorithmic Bias through a Systematic Review", "abstract": "Machine learning (ML) promises to revolutionize public health through improved surveillance, risk stratification, and resource allocation. However, without systematic attention to algorithmic bias, ML may inadvertently reinforce existing health disparities. We present a systematic literature review of algorithmic bias identification, discussion, and reporting in Dutch public health ML research from 2021 to 2025. To this end, we developed the Risk of Algorithmic Bias Assessment Tool (RABAT) by integrating elements from established frameworks (Cochrane Risk of Bias, PROBAST, Microsoft Responsible AI checklist) and applied it to 35 peer-reviewed studies. Our analysis reveals pervasive gaps: although data sampling and missing data practices are well documented, most studies omit explicit fairness framing, subgroup analyses, and transparent discussion of potential harms. In response, we introduce a four-stage fairness-oriented framework called ACAR (Awareness, Conceptualization, Application, Reporting), with guiding questions derived from our systematic literature review to help researchers address fairness across the ML lifecycle. We conclude with actionable recommendations for public health ML practitioners to consistently consider algorithmic bias and foster transparency, ensuring that algorithmic innovations advance health equity rather than undermine it.", "date": "2025-10-16", "categories": []}
{"id": "2510.13202", "title": "LLM-Guided Synthetic Augmentation (LGSA) for Mitigating Bias in AI Systems", "abstract": "Bias in AI systems, especially those relying on natural language data, raises ethical and practical concerns. Underrepresentation of certain groups often leads to uneven performance across demographics. Traditional fairness methods, such as pre-processing, in-processing, and post-processing, depend on protected-attribute labels, involve accuracy-fairness trade-offs, and may not generalize across datasets. To address these challenges, we propose LLM-Guided Synthetic Augmentation (LGSA), which uses large language models to generate counterfactual examples for underrepresented groups while preserving label integrity. We evaluated LGSA on a controlled dataset of short English sentences with gendered pronouns, professions, and binary classification labels. Structured prompts were used to produce gender-swapped paraphrases, followed by quality control including semantic similarity checks, attribute verification, toxicity screening, and human spot checks. The augmented dataset expanded training coverage and was used to train a classifier under consistent conditions. Results show that LGSA reduces performance disparities without compromising accuracy. The baseline model achieved 96.7 percent accuracy with a 7.2 percent gender bias gap. Simple swap augmentation reduced the gap to 0.7 percent but lowered accuracy to 95.6 percent. LGSA achieved 99.1 percent accuracy with a 1.9 percent bias gap, improving performance on female-labeled examples. These findings demonstrate that LGSA is an effective strategy for bias mitigation, enhancing subgroup balance while maintaining high task accuracy and label fidelity.", "date": "2025-10-15", "categories": []}
{"id": "2411.19623", "title": "FairDD: Fair Dataset Distillation", "abstract": "Condensing large datasets into smaller synthetic counterparts has demonstrated its promise for image classification. However, previous research has overlooked a crucial concern in image recognition: ensuring that models trained on condensed datasets are unbiased towards protected attributes (PA), such as gender and race. Our investigation reveals that dataset distillation fails to alleviate the unfairness towards minority groups within original datasets. Moreover, this bias typically worsens in the condensed datasets due to their smaller size. To bridge the research gap, we propose a novel fair dataset distillation (FDD) framework, namely FairDD, which can be seamlessly applied to diverse matching-based DD approaches (DDs), requiring no modifications to their original architectures. The key innovation of FairDD lies in synchronously matching synthetic datasets to PA-wise groups of original datasets, rather than indiscriminate alignment to the whole distributions in vanilla DDs, dominated by majority groups. This synchronized matching allows synthetic datasets to avoid collapsing into majority groups and bootstrap their balanced generation to all PA groups. Consequently, FairDD could effectively regularize vanilla DDs to favor biased generation toward minority groups while maintaining the accuracy of target attributes. Theoretical analyses and extensive experimental evaluations demonstrate that FairDD significantly improves fairness compared to vanilla DDs, with a promising trade-off between fairness and accuracy. Its consistent superiority across diverse DDs, spanning Distribution and Gradient Matching, establishes it as a versatile FDD approach. Code is available at https://github.com/zqhang/FairDD.", "date": "2024-11-29", "categories": []}
{"id": "2505.11740", "title": "Simple and Effective Specialized Representations for Fair Classifiers", "abstract": "Fair classification is a critical challenge that has gained increasing importance due to international regulations and its growing use in high-stakes decision-making settings. Existing methods often rely on adversarial learning or distribution matching across sensitive groups; however, adversarial learning can be unstable, and distribution matching can be computationally intensive. To address these limitations, we propose a novel approach based on the characteristic function distance. Our method ensures that the learned representation contains minimal sensitive information while maintaining high effectiveness for downstream tasks. By utilizing characteristic functions, we achieve a more stable and efficient solution compared to traditional methods. Additionally, we introduce a simple relaxation of the objective function that guarantees fairness in common classification models with no performance degradation. Experimental results on benchmark datasets demonstrate that our approach consistently matches or achieves better fairness and predictive accuracy than existing methods. Moreover, our method maintains robustness and computational efficiency, making it a practical solution for real-world applications.", "date": "2025-05-16", "categories": []}
{"id": "2506.06243", "title": "Fairmetrics: An R package for group fairness evaluation", "abstract": "Fairness is a growing area of machine learning (ML) that focuses on ensuring models do not produce systematically biased outcomes for specific groups, particularly those defined by protected attributes such as race, gender, or age. Evaluating fairness is a critical aspect of ML model development, as biased models can perpetuate structural inequalities. The {fairmetrics} R package offers a user-friendly framework for rigorously evaluating numerous group-based fairness criteria, including metrics based on independence (e.g., statistical parity), separation (e.g., equalized odds), and sufficiency (e.g., predictive parity). Group-based fairness criteria assess whether a model is equally accurate or well-calibrated across a set of predefined groups so that appropriate bias mitigation strategies can be implemented. {fairmetrics} provides both point and interval estimates for multiple metrics through a convenient wrapper function and includes an example dataset derived from the Medical Information Mart for Intensive Care, version II (MIMIC-II) database (Goldberger et al., 2000; Raffa, 2016).", "date": "2025-06-06", "categories": []}
{"id": "2509.08163", "title": "Machine Learning with Multitype Protected Attributes: Intersectional Fairness through Regularisation", "abstract": "Ensuring equitable treatment (fairness) across protected attributes (such as gender or ethnicity) is a critical issue in machine learning. Most existing literature focuses on binary classification, but achieving fairness in regression tasks-such as insurance pricing or hiring score assessments-is equally important. Moreover, anti-discrimination laws also apply to continuous attributes, such as age, for which many existing methods are not applicable. In practice, multiple protected attributes can exist simultaneously; however, methods targeting fairness across several attributes often overlook so-called \"fairness gerrymandering\", thereby ignoring disparities among intersectional subgroups (e.g., African-American women or Hispanic men). In this paper, we propose a distance covariance regularisation framework that mitigates the association between model predictions and protected attributes, in line with the fairness definition of demographic parity, and that captures both linear and nonlinear dependencies. To enhance applicability in the presence of multiple protected attributes, we extend our framework by incorporating two multivariate dependence measures based on distance covariance: the previously proposed joint distance covariance (JdCov) and our novel concatenated distance covariance (CCdCov), which effectively address fairness gerrymandering in both regression and classification tasks involving protected attributes of various types. We discuss and illustrate how to calibrate regularisation strength, including a method based on Jensen-Shannon divergence, which quantifies dissimilarities in prediction distributions across groups. We apply our framework to the COMPAS recidivism dataset and a large motor insurance claims dataset.", "date": "2025-09-09", "categories": []}
{"id": "2510.09705", "title": "A Multi-Component Reward Function with Policy Gradient for Automated Feature Selection with Dynamic Regularization and Bias Mitigation", "abstract": "Static feature exclusion strategies often fail to prevent bias when hidden dependencies influence the model predictions. To address this issue, we explore a reinforcement learning (RL) framework that integrates bias mitigation and automated feature selection within a single learning process. Unlike traditional heuristic-driven filter or wrapper approaches, our RL agent adaptively selects features using a reward signal that explicitly integrates predictive performance with fairness considerations. This dynamic formulation allows the model to balance generalization, accuracy, and equity throughout the training process, rather than rely exclusively on pre-processing adjustments or post hoc correction mechanisms. In this paper, we describe the construction of a multi-component reward function, the specification of the agents action space over feature subsets, and the integration of this system with ensemble learning. We aim to provide a flexible and generalizable way to select features in environments where predictors are correlated and biases can inadvertently re-emerge.", "date": "2025-10-09", "categories": []}
{"id": "2510.10191", "title": "Fairness Without Labels: Pseudo-Balancing for Bias Mitigation in Face Gender Classification", "abstract": "Face gender classification models often reflect and amplify demographic biases present in their training data, leading to uneven performance across gender and racial subgroups. We introduce pseudo-balancing, a simple and effective strategy for mitigating such biases in semi-supervised learning. Our method enforces demographic balance during pseudo-label selection, using only unlabeled images from a race-balanced dataset without requiring access to ground-truth annotations.\n  We evaluate pseudo-balancing under two conditions: (1) fine-tuning a biased gender classifier using unlabeled images from the FairFace dataset, and (2) stress-testing the method with intentionally imbalanced training data to simulate controlled bias scenarios. In both cases, models are evaluated on the All-Age-Faces (AAF) benchmark, which contains a predominantly East Asian population. Our results show that pseudo-balancing consistently improves fairness while preserving or enhancing accuracy. The method achieves 79.81% overall accuracy - a 6.53% improvement over the baseline - and reduces the gender accuracy gap by 44.17%. In the East Asian subgroup, where baseline disparities exceeded 49%, the gap is narrowed to just 5.01%. These findings suggest that even in the absence of label supervision, access to a demographically balanced or moderately skewed unlabeled dataset can serve as a powerful resource for debiasing existing computer vision models.", "date": "2025-10-11", "categories": []}
{"id": "2510.10998", "title": "ABLEIST: Intersectional Disability Bias in LLM-Generated Hiring Scenarios", "abstract": "Large language models (LLMs) are increasingly under scrutiny for perpetuating identity-based discrimination in high-stakes domains such as hiring, particularly against people with disabilities (PwD). However, existing research remains largely Western-centric, overlooking how intersecting forms of marginalization--such as gender and caste--shape experiences of PwD in the Global South. We conduct a comprehensive audit of six LLMs across 2,820 hiring scenarios spanning diverse disability, gender, nationality, and caste profiles. To capture subtle intersectional harms and biases, we introduce ABLEIST (Ableism, Inspiration, Superhumanization, and Tokenism), a set of five ableism-specific and three intersectional harm metrics grounded in disability studies literature. Our results reveal significant increases in ABLEIST harms towards disabled candidates--harms that many state-of-the-art models failed to detect. These harms were further amplified by sharp increases in intersectional harms (e.g., Tokenism) for gender and caste-marginalized disabled candidates, highlighting critical blind spots in current safety tools and the need for intersectional safety evaluations of frontier models in high-stakes domains like hiring.", "date": "2025-10-13", "categories": []}
{"id": "2510.11222", "title": "Fairness Metric Design Exploration in Multi-Domain Moral Sentiment Classification using Transformer-Based Models", "abstract": "Ensuring fairness in natural language processing for moral sentiment classification is challenging, particularly under cross-domain shifts where transformer models are increasingly deployed. Using the Moral Foundations Twitter Corpus (MFTC) and Moral Foundations Reddit Corpus (MFRC), this work evaluates BERT and DistilBERT in a multi-label setting with in-domain and cross-domain protocols. Aggregate performance can mask disparities: we observe pronounced asymmetry in transfer, with Twitter->Reddit degrading micro-F1 by 14.9% versus only 1.5% for Reddit->Twitter. Per-label analysis reveals fairness violations hidden by overall scores; notably, the authority label exhibits Demographic Parity Differences of 0.22-0.23 and Equalized Odds Differences of 0.40-0.41. To address this gap, we introduce the Moral Fairness Consistency (MFC) metric, which quantifies the cross-domain stability of moral foundation detection. MFC shows strong empirical validity, achieving a perfect negative correlation with Demographic Parity Difference (rho = -1.000, p < 0.001) while remaining independent of standard performance metrics. Across labels, loyalty demonstrates the highest consistency (MFC = 0.96) and authority the lowest (MFC = 0.78). These findings establish MFC as a complementary, diagnosis-oriented metric for fairness-aware evaluation of moral reasoning models, enabling more reliable deployment across heterogeneous linguistic contexts. .", "date": "2025-10-13", "categories": []}
{"id": "2506.19028", "title": "Quantifying Fairness in LLMs Beyond Tokens: A Semantic and Statistical Perspective", "abstract": "Large Language Models (LLMs) often generate responses with inherent biases, undermining their reliability in real-world applications. Existing evaluation methods often overlook biases in long-form responses and the intrinsic variability of LLM outputs. To address these challenges, we propose FiSCo (Fine-grained Semantic Comparison), a novel statistical framework to evaluate group-level fairness in LLMs by detecting subtle semantic differences in long-form responses across demographic groups. Unlike prior work focusing on sentiment or token-level comparisons, FiSCo goes beyond surface-level analysis by operating at the claim level, leveraging entailment checks to assess the consistency of meaning across responses. We decompose model outputs into semantically distinct claims and apply statistical hypothesis testing to compare inter- and intra-group similarities, enabling robust detection of subtle biases. We formalize a new group counterfactual fairness definition and validate FiSCo on both synthetic and human-annotated datasets spanning gender, race, and age. Experiments show that FiSCo more reliably identifies nuanced biases while reducing the impact of stochastic LLM variability, outperforming various evaluation metrics.", "date": "2025-06-23", "categories": []}
{"id": "2510.08628", "title": "The Digital Mirror: Gender Bias and Occupational Stereotypes in AI-Generated Images", "abstract": "Generative AI offers vast opportunities for creating visualisations, such as graphics, videos, and images. However, recent studies around AI-generated visualisations have primarily focused on the creation process and image quality, overlooking representational biases. This study addresses this gap by testing representation biases in AI-generated pictures in an occupational setting and evaluating how two AI image generator tools, DALL-E 3 and Ideogram, compare. Additionally, the study discusses topics such as ageing and emotions in AI-generated images. As AI image tools are becoming more widely used, addressing and mitigating harmful gender biases becomes essential to ensure diverse representation in media and professional settings. In this study, over 750 AI-generated images of occupations were prompted. The thematic analysis results revealed that both DALL-E 3 and Ideogram reinforce traditional gender stereotypes in AI-generated images, although to varying degrees. These findings emphasise that AI visualisation tools risk reinforcing narrow representations. In our discussion section, we propose suggestions for practitioners, individuals and researchers to increase representation when generating images with visible genders.", "date": "2025-10-08", "categories": []}
{"id": "2510.08840", "title": "The Boundaries of Fair AI in Medical Image Prognosis: A Causal Perspective", "abstract": "As machine learning (ML) algorithms are increasingly used in medical image analysis, concerns have emerged about their potential biases against certain social groups. Although many approaches have been proposed to ensure the fairness of ML models, most existing works focus only on medical image diagnosis tasks, such as image classification and segmentation, and overlooked prognosis scenarios, which involve predicting the likely outcome or progression of a medical condition over time. To address this gap, we introduce FairTTE, the first comprehensive framework for assessing fairness in time-to-event (TTE) prediction in medical imaging. FairTTE encompasses a diverse range of imaging modalities and TTE outcomes, integrating cutting-edge TTE prediction and fairness algorithms to enable systematic and fine-grained analysis of fairness in medical image prognosis. Leveraging causal analysis techniques, FairTTE uncovers and quantifies distinct sources of bias embedded within medical imaging datasets. Our large-scale evaluation reveals that bias is pervasive across different imaging modalities and that current fairness methods offer limited mitigation. We further demonstrate a strong association between underlying bias sources and model disparities, emphasizing the need for holistic approaches that target all forms of bias. Notably, we find that fairness becomes increasingly difficult to maintain under distribution shifts, underscoring the limitations of existing solutions and the pressing need for more robust, equitable prognostic models.", "date": "2025-10-09", "categories": []}
{"id": "2206.03234", "title": "Disparate Conditional Prediction in Multiclass Classifiers", "abstract": "We propose methods for auditing multiclass classifiers for fairness under multiclass equalized odds,by estimating the deviation from equalized odds when the classifier is not completely fair. We generalize to multiclass classifiers the measure of Disparate Conditional Prediction (DCP), originally suggested by Sabato & Yom-Tov (2020) for binary classifiers. DCP is defined as the fraction of the population for which the classifier predicts with conditional prediction probabilities that differ from the closest common baseline. We provide new local-optimization methods for estimating the multiclass DCPunder two different regimes,one in which the conditional confusion matrices for each protected sub-population are known, and one in which these cannot be estimated, for instance, because the classifier is inaccessible or because good-quality individual-level data is not available. These methods can be used to detect classifiers that likely treat a significant fraction of the population unfairly. Experiments demonstrate the accuracy of the methods. Code is provided at https://github.com/sivansabato/DCPmulticlass.", "date": "2022-06-07", "categories": []}
{"id": "2508.13169", "title": "Fair Play in the Newsroom: Actor-Based Filtering Gender Discrimination in Text Corpora", "abstract": "Language corpora are the foundation of most natural language processing research, yet they often reproduce structural inequalities. One such inequality is gender discrimination in how actors are represented, which can distort analyses and perpetuate discriminatory outcomes. This paper introduces a user-centric, actor-level pipeline for detecting and mitigating gender discrimination in large-scale text corpora. By combining discourse-aware analysis with metrics for sentiment, syntactic agency, and quotation styles, our method enables both fine-grained auditing and exclusion-based balancing. Applied to the taz2024full corpus of German newspaper articles (1980-2024), the pipeline yields a more gender-balanced dataset while preserving core dynamics of the source material. Our findings show that structural asymmetries can be reduced through systematic filtering, though subtler biases in sentiment and framing remain. We release the tools and reports to support further research in discourse-based fairness auditing and equitable corpus construction.", "date": "2025-08-07", "categories": []}
{"id": "2510.07328", "title": "MultiFair: Multimodal Balanced Fairness-Aware Medical Classification with Dual-Level Gradient Modulation", "abstract": "Medical decision systems increasingly rely on data from multiple sources to ensure reliable and unbiased diagnosis. However, existing multimodal learning models fail to achieve this goal because they often ignore two critical challenges. First, various data modalities may learn unevenly, thereby converging to a model biased towards certain modalities. Second, the model may emphasize learning on certain demographic groups causing unfair performances. The two aspects can influence each other, as different data modalities may favor respective groups during optimization, leading to both imbalanced and unfair multimodal learning. This paper proposes a novel approach called MultiFair for multimodal medical classification, which addresses these challenges with a dual-level gradient modulation process. MultiFair dynamically modulates training gradients regarding the optimization direction and magnitude at both data modality and group levels. We conduct extensive experiments on two multimodal medical datasets with different demographic groups. The results show that MultiFair outperforms state-of-the-art multimodal learning and fairness learning methods.", "date": "2025-09-30", "categories": []}
{"id": "2510.08086", "title": "From Ethical Declarations to Provable Independence: An Ontology-Driven Optimal-Transport Framework for Certifiably Fair AI Systems", "abstract": "This paper presents a framework for provably fair AI that overcomes the limits of current bias mitigation methods by systematically removing all sensitive information and its proxies. Using ontology engineering in OWL 2 QL, it formally defines sensitive attributes and infers their proxies through logical reasoning, constructing a sigma algebra G that captures the full structure of biased patterns. Fair representations are then obtained via Delbaen Majumdar optimal transport, which generates variables independent of G while minimizing L2 distance to preserve accuracy. This guarantees true independence rather than mere decorrelation. By modeling bias as dependence between sigma algebras, compiling ontological knowledge into measurable structures, and using optimal transport as the unique fair transformation, the approach ensures complete fairness in tasks like loan approval, where proxies such as ZIP code reveal race. The result is a certifiable and mathematically grounded method for trustworthy AI.", "date": "2025-10-09", "categories": []}
{"id": "2510.06354", "title": "LLM Bias Detection and Mitigation through the Lens of Desired Distributions", "abstract": "Although prior work on bias mitigation has focused on promoting social equality and demographic parity, less attention has been given to aligning LLM's outputs to desired distributions. For example, we might want to align a model with real-world distributions to support factual grounding. Thus, we define bias as deviation from a desired distribution, which may be an equal or real-world distribution, depending on application goals. We propose a weighted adaptive loss based fine-tuning method that aligns LLM's gender-profession output distribution with the desired distribution, while preserving language modeling capability. Using 3 profession sets -- male-dominated, female-dominated, and gender-balanced -- derived from U.S. labor statistics (2024), we assess both our adaptive method for reflecting reality and a non-adaptive variant for equality. Across three masked language models, bias is observed under both distributions. We achieve near-complete mitigation under equality and 30-75% reduction under real-world settings. Autoregressive LLMs show no bias under equality but notable bias under real-world settings, with the Llama Instruct models (3.2-3B, 3.1-8B) achieving a 50-62% reduction.", "date": "2025-10-07", "categories": []}
{"id": "2504.06303", "title": "On the Effectiveness and Generalization of Race Representations for Debiasing High-Stakes Decisions", "abstract": "Understanding and mitigating biases is critical for the adoption of large language models (LLMs) in high-stakes decision-making. We introduce Admissions and Hiring, decision tasks with hypothetical applicant profiles where a person's race can be inferred from their name, as simplified test beds for racial bias. We show that Gemma 2B Instruct and LLaMA 3.2 3B Instruct exhibit strong biases. Gemma grants admission to 26% more White than Black applicants, and LLaMA hires 60% more Asian than White applicants. We demonstrate that these biases are resistant to prompt engineering: multiple prompting strategies all fail to promote fairness. In contrast, using distributed alignment search, we can identify \"race subspaces\" within model activations and intervene on them to debias model decisions. Averaging the representation across all races within the subspaces reduces Gemma's bias by 37-57%. Finally, we examine the generalizability of Gemma's race subspaces, and find limited evidence for generalization, where changing the prompt format can affect the race representation. Our work suggests mechanistic approaches may provide a promising venue for improving the fairness of LLMs, but a universal race representation remains elusive.", "date": "2025-04-07", "categories": []}
{"id": "2510.04174", "title": "BLADE: Bias-Linked Adaptive DEbiasing", "abstract": "Neural networks have revolutionized numerous fields, yet they remain vulnerable to a critical flaw: the tendency to learn implicit biases, spurious correlations between certain attributes and target labels in training data. These biases are often more prevalent and easier to learn, causing models to rely on superficial patterns rather than task-relevant features necessary for generalization. Existing methods typically rely on strong assumptions, such as prior knowledge of these biases or access to bias-conflicting samples, i.e., samples that contradict spurious correlations and counterbalance bias-aligned samples, samples that conform to these spurious correlations. However, such assumptions are often impractical in real-world settings. We propose BLADE ({B}ias-{L}inked {A}daptive {DE}biasing), a generative debiasing framework that requires no prior knowledge of bias or bias-conflicting samples. BLADE first trains a generative model to translate images across bias domains while preserving task-relevant features. Then, it adaptively refines each image with its synthetic counterpart based on the image's susceptibility to bias. To encourage robust representations, BLADE aligns an image with its bias-translated synthetic counterpart that shares task-relevant features but differs in bias, while misaligning it with samples sharing the same bias. We evaluate BLADE on multiple benchmark datasets and show that it significantly outperforms state-of-the-art methods. Notably, it exceeds the closest baseline by an absolute margin of around 18% on the corrupted CIFAR-10 dataset under the worst group setting, establishing a new benchmark in bias mitigation and demonstrating its potential for developing more robust deep learning models without explicit supervision.", "date": "2025-10-05", "categories": []}
{"id": "2510.04317", "title": "FairAgent: Democratizing Fairness-Aware Machine Learning with LLM-Powered Agents", "abstract": "Training fair and unbiased machine learning models is crucial for high-stakes applications, yet it presents significant challenges. Effective bias mitigation requires deep expertise in fairness definitions, metrics, data preprocessing, and machine learning techniques. In addition, the complex process of balancing model performance with fairness requirements while properly handling sensitive attributes makes fairness-aware model development inaccessible to many practitioners. To address these challenges, we introduce FairAgent, an LLM-powered automated system that significantly simplifies fairness-aware model development. FairAgent eliminates the need for deep technical expertise by automatically analyzing datasets for potential biases, handling data preprocessing and feature engineering, and implementing appropriate bias mitigation strategies based on user requirements. Our experiments demonstrate that FairAgent achieves significant performance improvements while significantly reducing development time and expertise requirements, making fairness-aware machine learning more accessible to practitioners.", "date": "2025-10-05", "categories": []}
{"id": "2510.04624", "title": "Fairness in Repeated Matching: A Maximin Perspective", "abstract": "We study a sequential decision-making model where a set of items is repeatedly matched to the same set of agents over multiple rounds. The objective is to determine a sequence of matchings that either maximizes the utility of the least advantaged agent at the end of all rounds (optimal) or at the end of every individual round (anytime optimal). We investigate the computational challenges associated with finding (anytime) optimal outcomes and demonstrate that these problems are generally computationally intractable. However, we provide approximation algorithms, fixed-parameter tractable algorithms, and identify several special cases whereby the problem(s) can be solved efficiently. Along the way, we also establish characterizations of Pareto-optimal/maximum matchings, which may be of independent interest to works in matching theory and house allocation.", "date": "2025-10-06", "categories": []}
{"id": "2510.04926", "title": "Set to Be Fair: Demographic Parity Constraints for Set-Valued Classification", "abstract": "Set-valued classification is used in multiclass settings where confusion between classes can occur and lead to misleading predictions. However, its application may amplify discriminatory bias motivating the development of set-valued approaches under fairness constraints. In this paper, we address the problem of set-valued classification under demographic parity and expected size constraints. We propose two complementary strategies: an oracle-based method that minimizes classification risk while satisfying both constraints, and a computationally efficient proxy that prioritizes constraint satisfaction. For both strategies, we derive closed-form expressions for the (optimal) fair set-valued classifiers and use these to build plug-in, data-driven procedures for empirical predictions. We establish distribution-free convergence rates for violations of the size and fairness constraints for both methods, and under mild assumptions we also provide excess-risk bounds for the oracle-based approach. Empirical results demonstrate the effectiveness of both strategies and highlight the efficiency of our proxy method.", "date": "2025-10-06", "categories": []}
{"id": "2207.00108", "title": "Discrimination in machine learning algorithms", "abstract": "Machine learning algorithms are routinely used for business decisions that may directly affect individuals, for example, because a credit scoring algorithm refuses them a loan. It is then relevant from an ethical (and legal) point of view to ensure that these algorithms do not discriminate based on sensitive attributes (like sex or race), which may occur unwittingly and unknowingly by the operator and the management. Statistical tools and methods are then required to detect and eliminate such potential biases.", "date": "2022-06-30", "categories": []}
{"id": "2510.02352", "title": "Evaluating Bias in Spoken Dialogue LLMs for Real-World Decisions and Recommendations", "abstract": "While biases in large language models (LLMs), such as stereotypes and cultural tendencies in outputs, have been examined and identified, their presence and characteristics in spoken dialogue models (SDMs) with audio input and output remain largely unexplored. Paralinguistic features, such as age, gender, and accent, can affect model outputs; when compounded by multi-turn conversations, these effects may exacerbate biases, with potential implications for fairness in decision-making and recommendation tasks. In this paper, we systematically evaluate biases in speech LLMs and study the impact of multi-turn dialogues with repeated negative feedback. Bias is measured using Group Unfairness Score (GUS) for decisions and similarity-based normalized statistics rate (SNSR) for recommendations, across both open-source models like Qwen2.5-Omni and GLM-4-Voice, as well as closed-source APIs such as GPT-4o Audio and Gemini-2.5-Flash. Our analysis reveals that closed-source models generally exhibit lower bias, while open-source models are more sensitive to age and gender, and recommendation tasks tend to amplify cross-group disparities. We found that biased decisions may persist in multi-turn conversations. This work provides the first systematic study of biases in end-to-end spoken dialogue models, offering insights towards fair and reliable audio-based interactive systems. To facilitate further research, we release the FairDialogue dataset and evaluation code.", "date": "2025-09-27", "categories": []}
{"id": "2510.02742", "title": "IndiCASA: A Dataset and Bias Evaluation Framework in LLMs Using Contrastive Embedding Similarity in the Indian Context", "abstract": "Large Language Models (LLMs) have gained significant traction across critical domains owing to their impressive contextual understanding and generative capabilities. However, their increasing deployment in high stakes applications necessitates rigorous evaluation of embedded biases, particularly in culturally diverse contexts like India where existing embedding-based bias assessment methods often fall short in capturing nuanced stereotypes. We propose an evaluation framework based on a encoder trained using contrastive learning that captures fine-grained bias through embedding similarity. We also introduce a novel dataset - IndiCASA (IndiBias-based Contextually Aligned Stereotypes and Anti-stereotypes) comprising 2,575 human-validated sentences spanning five demographic axes: caste, gender, religion, disability, and socioeconomic status. Our evaluation of multiple open-weight LLMs reveals that all models exhibit some degree of stereotypical bias, with disability related biases being notably persistent, and religion bias generally lower likely due to global debiasing efforts demonstrating the need for fairer model development.", "date": "2025-10-03", "categories": []}
{"id": "2310.17687", "title": "Consistent End-to-End Estimation for Counterfactual Fairness", "abstract": "Fairness in predictions is of direct importance in practice due to legal, ethical, and societal reasons. This is often accomplished through counterfactual fairness, which ensures that the prediction for an individual is the same as that in a counterfactual world under a different sensitive attribute. However, achieving counterfactual fairness is challenging as counterfactuals are unobservable, and, because of that, existing baselines for counterfactual fairness do not have theoretical guarantees. In this paper, we propose a novel counterfactual fairness predictor for making predictions under counterfactual fairness. Here, we follow the standard counterfactual fairness setting and directly learn the counterfactual distribution of the descendants of the sensitive attribute via tailored neural networks, which we then use to enforce fair predictions through a novel counterfactual mediator regularization. Unique to our work is that we provide theoretical guarantees that our method is effective in ensuring the notion of counterfactual fairness. We further compare the performance across various datasets, where our method achieves state-of-the-art performance.", "date": "2023-10-26", "categories": []}
{"id": "2510.01281", "title": "An Analysis of the New EU AI Act and A Proposed Standardization Framework for Machine Learning Fairness", "abstract": "The European Union's AI Act represents a crucial step towards regulating ethical and responsible AI systems. However, we find an absence of quantifiable fairness metrics and the ambiguity in terminology, particularly the interchangeable use of the keywords transparency, explainability, and interpretability in the new EU AI Act and no reference of transparency of ethical compliance. We argue that this ambiguity creates substantial liability risk that would deter investment. Fairness transparency is strategically important. We recommend a more tailored regulatory framework to enhance the new EU AI regulation. Further-more, we propose a public system framework to assess the fairness and transparency of AI systems. Drawing from past work, we advocate for the standardization of industry best practices as a necessary addition to broad regulations to achieve the level of details required in industry, while preventing stifling innovation and investment in the AI sector. The proposals are exemplified with the case of ASR and speech synthesizers.", "date": "2025-09-30", "categories": []}
{"id": "2510.02017", "title": "FairContrast: Enhancing Fairness through Contrastive learning and Customized Augmenting Methods on Tabular Data", "abstract": "As AI systems become more embedded in everyday life, the development of fair and unbiased models becomes more critical. Considering the social impact of AI systems is not merely a technical challenge but a moral imperative. As evidenced in numerous research studies, learning fair and robust representations has proven to be a powerful approach to effectively debiasing algorithms and improving fairness while maintaining essential information for prediction tasks. Representation learning frameworks, particularly those that utilize self-supervised and contrastive learning, have demonstrated superior robustness and generalizability across various domains. Despite the growing interest in applying these approaches to tabular data, the issue of fairness in these learned representations remains underexplored. In this study, we introduce a contrastive learning framework specifically designed to address bias and learn fair representations in tabular datasets. By strategically selecting positive pair samples and employing supervised and self-supervised contrastive learning, we significantly reduce bias compared to existing state-of-the-art contrastive learning models for tabular data. Our results demonstrate the efficacy of our approach in mitigating bias with minimum trade-off in accuracy and leveraging the learned fair representations in various downstream tasks.", "date": "2025-10-02", "categories": []}
{"id": "2510.02036", "title": "The Current State of AI Bias Bounties: An Overview of Existing Programmes and Research", "abstract": "Current bias evaluation methods rarely engage with communities impacted by AI systems. Inspired by bug bounties, bias bounties have been proposed as a reward-based method that involves communities in AI bias detection by asking users of AI systems to report biases they encounter when interacting with such systems. In the absence of a state-of-the-art review, this survey aimed to identify and analyse existing AI bias bounty programmes and to present academic literature on bias bounties. Google, Google Scholar, PhilPapers, and IEEE Xplore were searched, and five bias bounty programmes, as well as five research publications, were identified. All bias bounties were organised by U.S.-based organisations as time-limited contests, with public participation in four programmes and prize pools ranging from 7,000 to 24,000 USD. The five research publications included a report on the application of bug bounties to algorithmic harms, an article addressing Twitter's bias bounty, a proposal for bias bounties as an institutional mechanism to increase AI scrutiny, a workshop discussing bias bounties from queer perspectives, and an algorithmic framework for bias bounties. We argue that reducing the technical requirements to enter bounty programmes is important to include those without coding experience. Given the limited adoption of bias bounties, future efforts should explore the transferability of the best practices from bug bounties and examine how such programmes can be designed to be sensitive to underrepresented groups while lowering adoption barriers for organisations.", "date": "2025-10-02", "categories": []}
{"id": "2510.00088", "title": "Judging by Appearances? Auditing and Intervening Vision-Language Models for Bail Prediction", "abstract": "Large language models (LLMs) have been extensively used for legal judgment prediction tasks based on case reports and crime history. However, with a surge in the availability of large vision language models (VLMs), legal judgment prediction systems can now be made to leverage the images of the criminals in addition to the textual case reports/crime history. Applications built in this way could lead to inadvertent consequences and be used with malicious intent. In this work, we run an audit to investigate the efficiency of standalone VLMs in the bail decision prediction task. We observe that the performance is poor across multiple intersectional groups and models \\textit{wrongly deny bail to deserving individuals with very high confidence}. We design different intervention algorithms by first including legal precedents through a RAG pipeline and then fine-tuning the VLMs using innovative schemes. We demonstrate that these interventions substantially improve the performance of bail prediction. Our work paves the way for the design of smarter interventions on VLMs in the future, before they can be deployed for real-world legal judgment prediction.", "date": "2025-09-30", "categories": []}
{"id": "2510.00163", "title": "Partial Identification Approach to Counterfactual Fairness Assessment", "abstract": "The wide adoption of AI decision-making systems in critical domains such as criminal justice, loan approval, and hiring processes has heightened concerns about algorithmic fairness. As we often only have access to the output of algorithms without insights into their internal mechanisms, it was natural to examine how decisions would alter when auxiliary sensitive attributes (such as race) change. This led the research community to come up with counterfactual fairness measures, but how to evaluate the measure from available data remains a challenging task. In many practical applications, the target counterfactual measure is not identifiable, i.e., it cannot be uniquely determined from the combination of quantitative data and qualitative knowledge. This paper addresses this challenge using partial identification, which derives informative bounds over counterfactual fairness measures from observational data. We introduce a Bayesian approach to bound unknown counterfactual fairness measures with high confidence. We demonstrate our algorithm on the COMPAS dataset, examining fairness in recidivism risk scores with respect to race, age, and sex. Our results reveal a positive (spurious) effect on the COMPAS score when changing race to African-American (from all others) and a negative (direct causal) effect when transitioning from young to old age.", "date": "2025-09-30", "categories": []}
{"id": "2502.05849", "title": "Where Fact Ends and Fairness Begins: Redefining AI Bias Evaluation through Cognitive Biases", "abstract": "Recent failures such as Google Gemini generating people of color in Nazi-era uniforms illustrate how AI outputs can be factually plausible yet socially harmful. AI models are increasingly evaluated for \"fairness,\" yet existing benchmarks often conflate two fundamentally different dimensions: factual correctness and normative fairness. A model may generate responses that are factually accurate but socially unfair, or conversely, appear fair while distorting factual reality. We argue that identifying the boundary between fact and fair is essential for meaningful fairness evaluation. We introduce Fact-or-Fair, a benchmark with (i) objective queries aligned with descriptive, fact-based judgments, and (ii) subjective queries aligned with normative, fairness-based judgments. Our queries are constructed from 19 statistics and are grounded in cognitive psychology, drawing on representativeness bias, attribution bias, and ingroup-outgroup bias to explain why models often misalign fact and fairness. Experiments across ten frontier models reveal different levels of fact-fair trade-offs. By reframing fairness evaluation, we provide both a new theoretical lens and a practical benchmark to advance the responsible model assessments. Our test suite is publicly available at https://github.com/uclanlp/Fact-or-Fair.", "date": "2025-02-09", "categories": []}
{"id": "2505.04931", "title": "Fair Uncertainty Quantification for Depression Prediction", "abstract": "Trustworthy depression prediction based on deep learning, incorporating both predictive reliability and algorithmic fairness across diverse demographic groups, is crucial for clinical application. Recently, achieving reliable depression predictions through uncertainty quantification has attracted increasing attention. However, few studies have focused on the fairness of uncertainty quantification (UQ) in depression prediction. In this work, we investigate the algorithmic fairness of UQ, namely Equal Opportunity Coverage (EOC) fairness, and propose Fair Uncertainty Quantification (FUQ) for depression prediction. FUQ pursues reliable and fair depression predictions through group-based analysis. Specifically, we first group all the participants by different sensitive attributes and leverage conformal prediction to quantify uncertainty within each demographic group, which provides a theoretically guaranteed and valid way to quantify uncertainty for depression prediction and facilitates the investigation of fairness across different demographic groups. Furthermore, we propose a fairness-aware optimization strategy that formulates fairness as a constrained optimization problem under EOC constraints. This enables the model to preserve predictive reliability while adapting to the heterogeneous uncertainty levels across demographic groups, thereby achieving optimal fairness. Through extensive evaluations on several visual and audio depression datasets, our approach demonstrates its effectiveness.", "date": "2025-05-08", "categories": []}
{"id": "2509.25435", "title": "GESA: Graph-Enhanced Semantic Allocation for Generalized, Fair, and Explainable Candidate-Role Matching", "abstract": "Accurate, fair, and explainable allocation of candidates to roles represents a fundamental challenge across multiple domains including corporate hiring, academic admissions, fellowship awards, and volunteer placement systems. Current state-of-the-art approaches suffer from semantic inflexibility, persistent demographic bias, opacity in decision-making processes, and poor scalability under dynamic policy constraints. We present GESA (Graph-Enhanced Semantic Allocation), a comprehensive framework that addresses these limitations through the integration of domain-adaptive transformer embeddings, heterogeneous self-supervised graph neural networks, adversarial debiasing mechanisms, multi-objective genetic optimization, and explainable AI components. Our experimental evaluation on large-scale international benchmarks comprising 20,000 candidate profiles and 3,000 role specifications demonstrates superior performance with 94.5% top-3 allocation accuracy, 37% improvement in diversity representation, 0.98 fairness score across demographic categories, and sub-second end-to-end latency. Additionally, GESA incorporates hybrid recommendation capabilities and glass-box explainability, making it suitable for deployment across diverse international contexts in industry, academia, and non-profit sectors.", "date": "2025-09-29", "categories": []}
{"id": "2509.25481", "title": "Fair Classification by Direct Intervention on Operating Characteristics", "abstract": "We develop new classifiers under group fairness in the attribute-aware setting for binary classification with multiple group fairness constraints (e.g., demographic parity (DP), equalized odds (EO), and predictive parity (PP)). We propose a novel approach, applicable to linear fractional constraints, based on directly intervening on the operating characteristics of a pre-trained base classifier, by (i) identifying optimal operating characteristics using the base classifier's group-wise ROC convex hulls and (ii) post-processing the base classifier to match those targets. As practical post-processors, we consider randomizing a mixture of group-wise thresholding rules subject to minimizing the expected number of interventions. We further extend our approach to handle multiple protected attributes and multiple linear fractional constraints. On standard datasets (COMPAS and ACSIncome), our methods simultaneously satisfy approximate DP, EO, and PP with few interventions and a near-oracle drop in accuracy; comparing favorably to previous methods.", "date": "2025-09-29", "categories": []}
{"id": "2509.25673", "title": "Mitigating Biases in Language Models via Bias Unlearning", "abstract": "Many studies have shown various biases targeting different demographic groups in language models, amplifying discrimination and harming fairness. Recent parameter modification debiasing approaches significantly degrade core capabilities such as text coherence and task accuracy. And Prompt-based debiasing methods, only effective for predefined trigger words, fail to address deeply embedded stereotypical associations in model parameters. In this paper, we propose BiasUnlearn, a novel model debiasing framework which achieves targeted debiasing via dual-pathway unlearning mechanisms coordinating stereotype forgetting with anti-stereotype retention, while preventing bias polarity reversal through adversarial forget set and dynamic dataset swapping. We conducted extensive experiments with multiple language models across various evaluation benchmarks. The results show that BiasUnlearn outperforms existing methods in mitigating bias in language models while retaining language modeling capabilities. Further experiments reveal that debiasing weights are transferable across model variants, confirming that bias representations become entrenched during pre-training and persist through fine-tuning phases.", "date": "2025-09-30", "categories": []}
{"id": "2509.26241", "title": "From Fragile to Certified: Wasserstein Audits of Group Fairness Under Distribution Shift", "abstract": "Group-fairness metrics (e.g., equalized odds) can vary sharply across resamples and are especially brittle under distribution shift, undermining reliable audits. We propose a Wasserstein distributionally robust framework that certifies worst-case group fairness over a ball of plausible test distributions centered at the empirical law. Our formulation unifies common group fairness notions via a generic conditional-probability functional and defines $\\varepsilon$-Wasserstein Distributional Fairness ($\\varepsilon$-WDF) as the audit target. Leveraging strong duality, we derive tractable reformulations and an efficient estimator (DRUNE) for $\\varepsilon$-WDF. We prove feasibility and consistency and establish finite-sample certification guarantees for auditing fairness, along with quantitative bounds under smoothness and margin conditions. Across standard benchmarks and classifiers, $\\varepsilon$-WDF delivers stable fairness assessments under distribution shift, providing a principled basis for auditing and certifying group fairness beyond observational data.", "date": "2025-09-30", "categories": []}
{"id": "2509.22660", "title": "Fairness for niche users and providers: algorithmic choice and profile portability", "abstract": "Ensuring fair outcomes for multiple stakeholders in recommender systems has been studied mostly in terms of algorithmic interventions: building new models with better fairness properties, or using reranking to improve outcomes from an existing algorithm. What has rarely been studied is structural changes in the recommendation ecosystem itself. Our work explores the fairness impact of algorithmic pluralism, the idea that the recommendation algorithm is decoupled from the platform through which users access content, enabling user choice in algorithms. Prior work using a simulation approach has shown that niche consumers and (especially) niche providers benefit from algorithmic choice. In this paper, we use simulation to explore the question of profile portability, to understand how different policies regarding the handling of user profiles interact with fairness outcomes for consumers and providers.", "date": "2025-08-28", "categories": []}
{"id": "2509.22703", "title": "AccessEval: Benchmarking Disability Bias in Large Language Models", "abstract": "Large Language Models (LLMs) are increasingly deployed across diverse domains but often exhibit disparities in how they handle real-life queries. To systematically investigate these effects within various disability contexts, we introduce \\textbf{AccessEval (Accessibility Evaluation)}, a benchmark evaluating 21 closed- and open-source LLMs across 6 real-world domains and 9 disability types using paired Neutral and Disability-Aware Queries. We evaluated model outputs with metrics for sentiment, social perception, and factual accuracy.\n  Our analysis reveals that responses to disability-aware queries tend to have a more negative tone, increased stereotyping, and higher factual error compared to neutral queries. These effects show notable variation by domain and disability type, with disabilities affecting hearing, speech, and mobility disproportionately impacted. These disparities reflect persistent forms of ableism embedded in model behavior.\n  By examining model performance in real-world decision-making contexts, we better illuminate how such biases can translate into tangible harms for disabled users. This framing helps bridges the gap between technical evaluation and user impact, reinforcing importance of bias mitigation in day-to-day applications. Our dataset is publicly available at: https://huggingface.co/datasets/Srikant86/AccessEval", "date": "2025-09-22", "categories": []}
{"id": "2509.22712", "title": "Achieving Fair Skin Lesion Detection through Skin Tone Normalization and Channel Pruning", "abstract": "Recent works have shown that deep learning based skin lesion image classification models trained on unbalanced dataset can exhibit bias toward protected demographic attributes such as race, age,and gender. Current bias mitigation methods usually either achieve high level of fairness with the degradation of accuracy, or only improve the model fairness on a single attribute. Additionally usually most bias mitigation strategies are either pre hoc through data processing or post hoc through fairness evaluation, instead of being integrated into the model learning itself. To solve these existing drawbacks, we propose a new Individual Typology Angle (ITA) Loss-based skin tone normalization and data augmentation method that directly feeds into an adaptable meta learning-based joint channel pruning framework. In skin tone normalization, ITA is used to estimate skin tone type and adjust automatically to target tones for dataset balancing. In the joint channel pruning framework, two nested optimization loops are used to find critical channels.The inner optimization loop finds and prunes the local critical channels by weighted soft nearest neighbor loss, and the outer optimization loop updates the weight of each attribute using group wise variance loss on meta-set. Experiments conducted in the ISIC2019 dataset validate the effectiveness of our method in simultaneously improving the fairness of the model on multiple sensitive attributes without significant degradation of accuracy. Finally, although the pruning mechanism adds some computational cost during training phase, usually training is done off line. More importantly,", "date": "2025-09-24", "categories": []}
{"id": "2509.23482", "title": "GeoBS: Information-Theoretic Quantification of Geographic Bias in AI Models", "abstract": "The widespread adoption of AI models, especially foundation models (FMs), has made a profound impact on numerous domains. However, it also raises significant ethical concerns, including bias issues. Although numerous efforts have been made to quantify and mitigate social bias in AI models, geographic bias (in short, geo-bias) receives much less attention, which presents unique challenges. While previous work has explored ways to quantify geo-bias, these measures are model-specific (e.g., mean absolute deviation of LLM ratings) or spatially implicit (e.g., average fairness scores of all spatial partitions). We lack a model-agnostic, universally applicable, and spatially explicit geo-bias evaluation framework that allows researchers to fairly compare the geo-bias of different AI models and to understand what spatial factors contribute to the geo-bias. In this paper, we establish an information-theoretic framework for geo-bias evaluation, called GeoBS (Geo-Bias Scores). We demonstrate the generalizability of the proposed framework by showing how to interpret and analyze existing geo-bias measures under this framework. Then, we propose three novel geo-bias scores that explicitly take intricate spatial factors (multi-scalability, distance decay, and anisotropy) into consideration. Finally, we conduct extensive experiments on 3 tasks, 8 datasets, and 8 models to demonstrate that both task-specific GeoAI models and general-purpose foundation models may suffer from various types of geo-bias. This framework will not only advance the technical understanding of geographic bias but will also establish a foundation for integrating spatial fairness into the design, deployment, and evaluation of AI systems.", "date": "2025-09-27", "categories": []}
{"id": "2509.23805", "title": "Open-DeBias: Toward Mitigating Open-Set Bias in Language Models", "abstract": "Large Language Models (LLMs) have achieved remarkable success on question answering (QA) tasks, yet they often encode harmful biases that compromise fairness and trustworthiness. Most existing bias mitigation approaches are restricted to predefined categories, limiting their ability to address novel or context-specific emergent biases. To bridge this gap, we tackle the novel problem of open-set bias detection and mitigation in text-based QA. We introduce OpenBiasBench, a comprehensive benchmark designed to evaluate biases across a wide range of categories and subgroups, encompassing both known and previously unseen biases. Additionally, we propose Open-DeBias, a novel, data-efficient, and parameter-efficient debiasing method that leverages adapter modules to mitigate existing social and stereotypical biases while generalizing to unseen ones. Compared to the state-of-the-art BMBI method, Open-DeBias improves QA accuracy on BBQ dataset by nearly $48\\%$ on ambiguous subsets and $6\\%$ on disambiguated ones, using adapters fine-tuned on just a small fraction of the training data. Remarkably, the same adapters, in a zero-shot transfer to Korean BBQ, achieve $84\\%$ accuracy, demonstrating robust language-agnostic generalization. Through extensive evaluation, we also validate the effectiveness of Open-DeBias across a broad range of NLP tasks, including StereoSet and CrowS-Pairs, highlighting its robustness, multilingual strength, and suitability for general-purpose, open-domain bias mitigation. The project page is available at: https://sites.google.com/view/open-debias25", "date": "2025-09-28", "categories": []}
{"id": "2509.24077", "title": "Demographic-Agnostic Fairness without Harm", "abstract": "As machine learning (ML) algorithms are increasingly used in social domains to make predictions about humans, there is a growing concern that these algorithms may exhibit biases against certain social groups. Numerous notions of fairness have been proposed in the literature to measure the unfairness of ML. Among them, one class that receives the most attention is \\textit{parity-based}, i.e., achieving fairness by equalizing treatment or outcomes for different social groups. However, achieving parity-based fairness often comes at the cost of lowering model accuracy and is undesirable for many high-stakes domains like healthcare. To avoid inferior accuracy, a line of research focuses on \\textit{preference-based} fairness, under which any group of individuals would experience the highest accuracy and collectively prefer the ML outcomes assigned to them if they were given the choice between various sets of outcomes. However, these works assume individual demographic information is known and fully accessible during training. In this paper, we relax this requirement and propose a novel \\textit{demographic-agnostic fairness without harm (DAFH)} optimization algorithm, which jointly learns a group classifier that partitions the population into multiple groups and a set of decoupled classifiers associated with these groups. Theoretically, we conduct sample complexity analysis and show that our method can outperform the baselines when demographic information is known and used to train decoupled classifiers. Experiments on both synthetic and real data validate the proposed method.", "date": "2025-09-28", "categories": []}
{"id": "2509.24468", "title": "Bias Mitigation or Cultural Commonsense? Evaluating LLMs with a Japanese Dataset", "abstract": "Large language models (LLMs) exhibit social biases, prompting the development of various debiasing methods. However, debiasing methods may degrade the capabilities of LLMs. Previous research has evaluated the impact of bias mitigation primarily through tasks measuring general language understanding, which are often unrelated to social biases. In contrast, cultural commonsense is closely related to social biases, as both are rooted in social norms and values. The impact of bias mitigation on cultural commonsense in LLMs has not been well investigated. Considering this gap, we propose SOBACO (SOcial BiAs and Cultural cOmmonsense benchmark), a Japanese benchmark designed to evaluate social biases and cultural commonsense in LLMs in a unified format. We evaluate several LLMs on SOBACO to examine how debiasing methods affect cultural commonsense in LLMs. Our results reveal that the debiasing methods degrade the performance of the LLMs on the cultural commonsense task (up to 75% accuracy deterioration). These results highlight the importance of developing debiasing methods that consider the trade-off with cultural commonsense to improve fairness and utility of LLMs.", "date": "2025-09-29", "categories": []}
{"id": "2502.13319", "title": "Elucidating Mechanisms of Demographic Bias in LLMs for Healthcare", "abstract": "We know from prior work that LLMs encode social biases, and that this manifests in clinical tasks. In this work we adopt tools from mechanistic interpretability to unveil sociodemographic representations and biases within LLMs in the context of healthcare. Specifically, we ask: Can we identify activations within LLMs that encode sociodemographic information (e.g., gender, race)? We find that gender information is highly localized in MLP layers and can be reliably manipulated at inference time via patching. Such interventions can surgically alter generated clinical vignettes for specific conditions, and also influence downstream clinical predictions which correlate with gender, e.g., patient risk of depression. We find that representation of patient race is somewhat more distributed, but can also be intervened upon, to a degree. To our knowledge, this is the first application of mechanistic interpretability methods to LLMs for healthcare.", "date": "2025-02-18", "categories": []}
{"id": "2509.21466", "title": "Gender Stereotypes in Professional Roles Among Saudis: An Analytical Study of AI-Generated Images Using Language Models", "abstract": "This study investigates the extent to which contemporary Text-to-Image artificial intelligence (AI) models perpetuate gender stereotypes and cultural inaccuracies when generating depictions of professionals in Saudi Arabia. We analyzed 1,006 images produced by ImageFX, DALL-E V3, and Grok for 56 diverse Saudi professions using neutral prompts. Two trained Saudi annotators evaluated each image on five dimensions: perceived gender, clothing and appearance, background and setting, activities and interactions, and age. A third senior researcher adjudicated whenever the two primary raters disagreed, yielding 10,100 individual judgements. The results reveal a strong gender imbalance, with ImageFX outputs being 85\\% male, Grok 86.6\\% male, and DALL-E V3 96\\% male, indicating that DALL-E V3 exhibited the strongest overall gender stereotyping. This imbalance was most evident in leadership and technical roles. Moreover, cultural inaccuracies in clothing, settings, and depicted activities were frequently observed across all three models. Counter-stereotypical images often arise from cultural misinterpretations rather than genuinely progressive portrayals. We conclude that current models mirror societal biases embedded in their training data, generated by humans, offering only a limited reflection of the Saudi labour market's gender dynamics and cultural nuances. These findings underscore the urgent need for more diverse training data, fairer algorithms, and culturally sensitive evaluation frameworks to ensure equitable and authentic visual outputs.", "date": "2025-09-25", "categories": []}
{"id": "2509.21785", "title": "Unbiased Binning: Fairness-aware Attribute Representation", "abstract": "Discretizing raw features into bucketized attribute representations is a popular step before sharing a dataset. It is, however, evident that this step can cause significant bias in data and amplify unfairness in downstream tasks.\n  In this paper, we address this issue by introducing the unbiased binning problem that, given an attribute to bucketize, finds its closest discretization to equal-size binning that satisfies group parity across different buckets. Defining a small set of boundary candidates, we prove that unbiased binning must select its boundaries from this set. We then develop an efficient dynamic programming algorithm on top of the boundary candidates to solve the unbiased binning problem.\n  Finding an unbiased binning may sometimes result in a high price of fairness, or it may not even exist, especially when group values follow different distributions. Considering that a small bias in the group ratios may be tolerable in such settings, we introduce the epsilon-biased binning problem that bounds the group disparities across buckets to a small value epsilon. We first develop a dynamic programming solution, DP, that finds the optimal binning in quadratic time. The DP algorithm, while polynomial, does not scale to very large settings. Therefore, we propose a practically scalable algorithm, based on local search (LS), for epsilon-biased binning. The key component of the LS algorithm is a divide-and-conquer (D&C) algorithm that finds a near-optimal solution for the problem in near-linear time. We prove that D&C finds a valid solution for the problem unless none exists. The LS algorithm then initiates a local search, using the D&C solution as the upper bound, to find the optimal solution.", "date": "2025-09-26", "categories": []}
{"id": "2509.22232", "title": "Fairness-Aware Reinforcement Learning (FAReL): A Framework for Transparent and Balanced Sequential Decision-Making", "abstract": "Equity in real-world sequential decision problems can be enforced using fairness-aware methods. Therefore, we require algorithms that can make suitable and transparent trade-offs between performance and the desired fairness notions. As the desired performance-fairness trade-off is hard to specify a priori, we propose a framework where multiple trade-offs can be explored. Insights provided by the reinforcement learning algorithm regarding the obtainable performance-fairness trade-offs can then guide stakeholders in selecting the most appropriate policy. To capture fairness, we propose an extended Markov decision process, $f$MDP, that explicitly encodes individuals and groups. Given this $f$MDP, we formalise fairness notions in the context of sequential decision problems and formulate a fairness framework that computes fairness measures over time. We evaluate our framework in two scenarios with distinct fairness requirements: job hiring, where strong teams must be composed while treating applicants equally, and fraud detection, where fraudulent transactions must be detected while ensuring the burden on customers is fairly distributed. We show that our framework learns policies that are more fair across multiple scenarios, with only minor loss in performance reward. Moreover, we observe that group and individual fairness notions do not necessarily imply one another, highlighting the benefit of our framework in settings where both fairness types are desired. Finally, we provide guidelines on how to apply this framework across different problem settings.", "date": "2025-09-26", "categories": []}
{"id": "2409.02963", "title": "Fair Clustering with Minimum Representation Constraints", "abstract": "Clustering is a well-studied unsupervised learning task that aims to partition data points into a number of clusters. In many applications, these clusters correspond to real-world constructs (e.g., electoral districts, playlists, TV channels), where a group (e.g., social or demographic) benefits only if it reaches a minimum level of representation in the cluster (e.g., 50% to elect their preferred candidate). In this paper, we study the k-means and k-medians clustering problems under the additional fairness constraint that each group must attain a minimum level of representation in at least a specified number of clusters. We formulate this problem as a mixed-integer (nonlinear) optimization problem and propose an alternating minimization algorithm, called MiniReL, to solve it. Although incorporating fairness constraints results in an NP-hard assignment problem within the MiniReL algorithm, we present several heuristic strategies that make the approach practical even for large datasets. Numerical results demonstrate that our method yields fair clusters without increasing clustering cost across standard benchmark datasets.", "date": "2024-09-04", "categories": []}
{"id": "2501.05926", "title": "LLMs Reproduce Stereotypes of Sexual and Gender Minorities", "abstract": "A large body of research has found substantial gender bias in NLP systems. Most of this research takes a binary, essentialist view of gender: limiting its variation to the categories _men_ and _women_, conflating gender with sex, and ignoring different sexual identities. But gender and sexuality exist on a spectrum, so in this paper we study the biases of large language models (LLMs) towards sexual and gender minorities beyond binary categories. Grounding our study in a widely used social psychology model -- the Stereotype Content Model -- we demonstrate that English-language survey questions about social perceptions elicit more negative stereotypes of sexual and gender minorities from both humans and LLMs. We then extend this framework to a more realistic use case: text generation. Our analysis shows that LLMs generate stereotyped representations of sexual and gender minorities in this setting, showing that they amplify representational harms in creative writing, a widely advertised use for LLMs.", "date": "2025-01-10", "categories": []}
{"id": "2505.04038", "title": "Identities are not Interchangeable: The Problem of Overgeneralization in Fair Machine Learning", "abstract": "A key value proposition of machine learning is generalizability: the same methods and model architecture should be able to work across different domains and different contexts. While powerful, this generalization can sometimes go too far, and miss the importance of the specifics. In this work, we look at how fair machine learning has often treated as interchangeable the identity axis along which discrimination occurs. In other words, racism is measured and mitigated the same way as sexism, as ableism, as ageism. Disciplines outside of computer science have pointed out both the similarities and differences between these different forms of oppression, and in this work we draw out the implications for fair machine learning. While certainly not all aspects of fair machine learning need to be tailored to the specific form of oppression, there is a pressing need for greater attention to such specificity than is currently evident. Ultimately, context specificity can deepen our understanding of how to build more fair systems, widen our scope to include currently overlooked harms, and, almost paradoxically, also help to narrow our scope and counter the fear of an infinite number of group-specific methods of analysis.", "date": "2025-05-07", "categories": []}
{"id": "2509.19927", "title": "TABFAIRGDT: A Fast Fair Tabular Data Generator using Autoregressive Decision Trees", "abstract": "Ensuring fairness in machine learning remains a significant challenge, as models often inherit biases from their training data. Generative models have recently emerged as a promising approach to mitigate bias at the data level while preserving utility. However, many rely on deep architectures, despite evidence that simpler models can be highly effective for tabular data. In this work, we introduce TABFAIRGDT, a novel method for generating fair synthetic tabular data using autoregressive decision trees. To enforce fairness, we propose a soft leaf resampling technique that adjusts decision tree outputs to reduce bias while preserving predictive performance. Our approach is non-parametric, effectively capturing complex relationships between mixed feature types, without relying on assumptions about the underlying data distributions. We evaluate TABFAIRGDT on benchmark fairness datasets and demonstrate that it outperforms state-of-the-art (SOTA) deep generative models, achieving better fairness-utility trade-off for downstream tasks, as well as higher synthetic data quality. Moreover, our method is lightweight, highly efficient, and CPU-compatible, requiring no data pre-processing. Remarkably, TABFAIRGDT achieves a 72% average speedup over the fastest SOTA baseline across various dataset sizes, and can generate fair synthetic data for medium-sized datasets (10 features, 10K samples) in just one second on a standard CPU, making it an ideal solution for real-world fairness-sensitive applications.", "date": "2025-09-24", "categories": []}
{"id": "2509.13908", "title": "APFEx: Adaptive Pareto Front Explorer for Intersectional Fairness", "abstract": "Ensuring fairness in machine learning models is critical, especially when biases compound across intersecting protected attributes like race, gender, and age. While existing methods address fairness for single attributes, they fail to capture the nuanced, multiplicative biases faced by intersectional subgroups. We introduce Adaptive Pareto Front Explorer (APFEx), the first framework to explicitly model intersectional fairness as a joint optimization problem over the Cartesian product of sensitive attributes. APFEx combines three key innovations- (1) an adaptive multi-objective optimizer that dynamically switches between Pareto cone projection, gradient weighting, and exploration strategies to navigate fairness-accuracy trade-offs, (2) differentiable intersectional fairness metrics enabling gradient-based optimization of non-smooth subgroup disparities, and (3) theoretical guarantees of convergence to Pareto-optimal solutions. Experiments on four real-world datasets demonstrate APFEx's superiority, reducing fairness violations while maintaining competitive accuracy. Our work bridges a critical gap in fair ML, providing a scalable, model-agnostic solution for intersectional fairness.", "date": "2025-09-17", "categories": []}
{"id": "2506.23411", "title": "Datasets for Fairness in Language Models: An In-Depth Survey", "abstract": "Despite the growing reliance on fairness benchmarks to evaluate language models, the datasets that underpin these benchmarks remain critically underexamined. This survey addresses that overlooked foundation by offering a comprehensive analysis of the most widely used fairness datasets in language model research. To ground this analysis, we characterize each dataset across key dimensions, including provenance, demographic scope, annotation design, and intended use, revealing the assumptions and limitations baked into current evaluation practices. Building on this foundation, we propose a unified evaluation framework that surfaces consistent patterns of demographic disparities across benchmarks and scoring metrics. Applying this framework to sixteen popular datasets, we uncover overlooked biases that may distort conclusions about model fairness and offer guidance on selecting, combining, and interpreting these resources more effectively and responsibly. Our findings highlight an urgent need for new benchmarks that capture a broader range of social contexts and fairness notions. To support future research, we release all data, code, and results at https://github.com/vanbanTruong/Fairness-in-Large-Language-Models/tree/main/datasets, fostering transparency and reproducibility in the evaluation of language model fairness.", "date": "2025-06-29", "categories": []}
{"id": "2509.16400", "title": "'Rich Dad, Poor Lad': How do Large Language Models Contextualize Socioeconomic Factors in College Admission ?", "abstract": "Large Language Models (LLMs) are increasingly involved in high-stakes domains, yet how they reason about socially sensitive decisions remains underexplored. We present a large-scale audit of LLMs' treatment of socioeconomic status (SES) in college admissions decisions using a novel dual-process framework inspired by cognitive science. Leveraging a synthetic dataset of 30,000 applicant profiles grounded in real-world correlations, we prompt 4 open-source LLMs (Qwen 2, Mistral v0.3, Gemma 2, Llama 3.1) under 2 modes: a fast, decision-only setup (System 1) and a slower, explanation-based setup (System 2). Results from 5 million prompts reveal that LLMs consistently favor low-SES applicants -- even when controlling for academic performance -- and that System 2 amplifies this tendency by explicitly invoking SES as compensatory justification, highlighting both their potential and volatility as decision-makers. We then propose DPAF, a dual-process audit framework to probe LLMs' reasoning behaviors in sensitive applications.", "date": "2025-09-19", "categories": []}
{"id": "2509.16462", "title": "Intrinsic Meets Extrinsic Fairness: Assessing the Downstream Impact of Bias Mitigation in Large Language Models", "abstract": "Large Language Models (LLMs) exhibit socio-economic biases that can propagate into downstream tasks. While prior studies have questioned whether intrinsic bias in LLMs affects fairness at the downstream task level, this work empirically investigates the connection. We present a unified evaluation framework to compare intrinsic bias mitigation via concept unlearning with extrinsic bias mitigation via counterfactual data augmentation (CDA). We examine this relationship through real-world financial classification tasks, including salary prediction, employment status, and creditworthiness assessment. Using three open-source LLMs, we evaluate models both as frozen embedding extractors and as fine-tuned classifiers. Our results show that intrinsic bias mitigation through unlearning reduces intrinsic gender bias by up to 94.9%, while also improving downstream task fairness metrics, such as demographic parity by up to 82%, without compromising accuracy. Our framework offers practical guidance on where mitigation efforts can be most effective and highlights the importance of applying early-stage mitigation before downstream deployment.", "date": "2025-09-19", "categories": []}
{"id": "2509.16475", "title": "Towards Universal Debiasing for Language Models-based Tabular Data Generation", "abstract": "Large language models (LLMs) have achieved promising results in tabular data generation. However, inherent historical biases in tabular datasets often cause LLMs to exacerbate fairness issues, particularly when multiple advantaged and protected features are involved. In this work, we introduce a universal debiasing framework that minimizes group-level dependencies by simultaneously reducing the mutual information between advantaged and protected attributes. By leveraging the autoregressive structure and analytic sampling distributions of LLM-based tabular data generators, our approach efficiently computes mutual information, reducing the need for cumbersome numerical estimations. Building on this foundation, we propose two complementary methods: a direct preference optimization (DPO)-based strategy, namely UDF-DPO, that integrates seamlessly with existing models, and a targeted debiasing technique, namely UDF-MIX, that achieves debiasing without tuning the parameters of LLMs. Extensive experiments demonstrate that our framework effectively balances fairness and utility, offering a scalable and practical solution for debiasing in high-stakes applications.", "date": "2025-09-20", "categories": []}
{"id": "2509.16491", "title": "FairTune: A Bias-Aware Fine-Tuning Framework Towards Fair Heart Rate Prediction from PPG", "abstract": "Foundation models pretrained on physiological data such as photoplethysmography (PPG) signals are increasingly used to improve heart rate (HR) prediction across diverse settings. Fine-tuning these models for local deployment is often seen as a practical and scalable strategy. However, its impact on demographic fairness particularly under domain shifts remains underexplored. We fine-tune PPG-GPT a transformer-based foundation model pretrained on intensive care unit (ICU) data across three heterogeneous datasets (ICU, wearable, smartphone) and systematically evaluate the effects on HR prediction accuracy and gender fairness. While fine-tuning substantially reduces mean absolute error (up to 80%), it can simultaneously widen fairness gaps, especially in larger models and under significant distributional characteristics shifts. To address this, we introduce FairTune, a bias-aware fine-tuning framework in which we benchmark three mitigation strategies: class weighting based on inverse group frequency (IF), Group Distributionally Robust Optimization (GroupDRO), and adversarial debiasing (ADV). We find that IF and GroupDRO significantly reduce fairness gaps without compromising accuracy, with effectiveness varying by deployment domain. Representation analyses further reveal that mitigation techniques reshape internal embeddings to reduce demographic clustering. Our findings highlight that fairness does not emerge as a natural byproduct of fine-tuning and that explicit mitigation is essential for equitable deployment of physiological foundation models.", "date": "2025-09-20", "categories": []}
{"id": "2509.17411", "title": "Robust Mixture Models for Algorithmic Fairness Under Latent Heterogeneity", "abstract": "Standard machine learning models optimized for average performance often fail on minority subgroups and lack robustness to distribution shifts. This challenge worsens when subgroups are latent and affected by complex interactions among continuous and discrete features. We introduce ROME (RObust Mixture Ensemble), a framework that learns latent group structure from data while optimizing for worst-group performance. ROME employs two approaches: an Expectation-Maximization algorithm for linear models and a neural Mixture-of-Experts for nonlinear settings. Through simulations and experiments on real-world datasets, we demonstrate that ROME significantly improves algorithmic fairness compared to standard methods while maintaining competitive average performance. Importantly, our method requires no predefined group labels, making it practical when sources of disparities are unknown or evolving.", "date": "2025-09-22", "categories": []}
{"id": "2509.18067", "title": "Learning to Rank with Top-$K$ Fairness", "abstract": "Fairness in ranking models is crucial, as disparities in exposure can disproportionately affect protected groups. Most fairness-aware ranking systems focus on ensuring comparable average exposure for groups across the entire ranked list, which may not fully address real-world concerns. For example, when a ranking model is used for allocating resources among candidates or disaster hotspots, decision-makers often prioritize only the top-$K$ ranked items, while the ranking beyond top-$K$ becomes less relevant. In this paper, we propose a list-wise learning-to-rank framework that addresses the issues of inequalities in top-$K$ rankings at training time. Specifically, we propose a top-$K$ exposure disparity measure that extends the classic exposure disparity metric in a ranked list. We then learn a ranker to balance relevance and fairness in top-$K$ rankings. Since direct top-$K$ selection is computationally expensive for a large number of items, we transform the non-differentiable selection process into a differentiable objective function and develop efficient stochastic optimization algorithms to achieve both high accuracy and sufficient fairness. Extensive experiments demonstrate that our method outperforms existing methods.", "date": "2025-09-22", "categories": []}
{"id": "2506.12556", "title": "Algorithmic Fairness: Not a Purely Technical but Socio-Technical Property", "abstract": "The rapid trend of deploying artificial intelligence (AI) and machine learning (ML) systems in socially consequential domains has raised growing concerns about their trustworthiness, including potential discriminatory behaviours. Research in algorithmic fairness has generated a proliferation of mathematical definitions and metrics, yet persistent misconceptions and limitations -- both within and beyond the fairness community -- limit their effectiveness, such as an unreached consensus on its understanding, prevailing measures primarily tailored to binary group settings, and superficial handling for intersectional contexts. Here we critically remark on these misconceptions and argue that fairness cannot be reduced to purely technical constraints on models; we also examine the limitations of existing fairness measures through conceptual analysis and empirical illustrations, showing their limited applicability in the face of complex real-world scenarios, challenging prevailing views on the incompatibility between accuracy and fairness as well as that among fairness measures themselves, and outlining three worth-considering principles in the design of fairness measures. We believe these findings will help bridge the gap between technical formalisation and social realities and meet the challenges of real-world AI/ML deployment.", "date": "2025-06-14", "categories": []}
{"id": "2509.14438", "title": "Simulating a Bias Mitigation Scenario in Large Language Models", "abstract": "Large Language Models (LLMs) have fundamentally transformed the field of natural language processing; however, their vulnerability to biases presents a notable obstacle that threatens both fairness and trust. This review offers an extensive analysis of the bias landscape in LLMs, tracing its roots and expressions across various NLP tasks. Biases are classified into implicit and explicit types, with particular attention given to their emergence from data sources, architectural designs, and contextual deployments. This study advances beyond theoretical analysis by implementing a simulation framework designed to evaluate bias mitigation strategies in practice. The framework integrates multiple approaches including data curation, debiasing during model training, and post-hoc output calibration and assesses their impact in controlled experimental settings. In summary, this work not only synthesizes existing knowledge on bias in LLMs but also contributes original empirical validation through simulation of mitigation strategies.", "date": "2025-09-17", "categories": []}
{"id": "2509.13496", "title": "BiasMap: Leveraging Cross-Attentions to Discover and Mitigate Hidden Social Biases in Text-to-Image Generation", "abstract": "Bias discovery is critical for black-box generative models, especiall text-to-image (TTI) models. Existing works predominantly focus on output-level demographic distributions, which do not necessarily guarantee concept representations to be disentangled post-mitigation. We propose BiasMap, a model-agnostic framework for uncovering latent concept-level representational biases in stable diffusion models. BiasMap leverages cross-attention attribution maps to reveal structural entanglements between demographics (e.g., gender, race) and semantics (e.g., professions), going deeper into representational bias during the image generation. Using attribution maps of these concepts, we quantify the spatial demographics-semantics concept entanglement via Intersection over Union (IoU), offering a lens into bias that remains hidden in existing fairness discovery approaches. In addition, we further utilize BiasMap for bias mitigation through energy-guided diffusion sampling that directly modifies latent noise space and minimizes the expected SoftIoU during the denoising process. Our findings show that existing fairness interventions may reduce the output distributional gap but often fail to disentangle concept-level coupling, whereas our mitigation method can mitigate concept entanglement in image generation while complementing distributional bias mitigation.", "date": "2025-09-16", "categories": []}
{"id": "2509.13651", "title": "Controllable Pareto Trade-off between Fairness and Accuracy", "abstract": "The fairness-accuracy trade-off is a key challenge in NLP tasks. Current work focuses on finding a single \"optimal\" solution to balance the two objectives, which is limited considering the diverse solutions on the Pareto front. This work intends to provide controllable trade-offs according to the user's preference of the two objectives, which is defined as a reference vector. To achieve this goal, we apply multi-objective optimization (MOO), which can find solutions from various regions of the Pareto front. However, it is challenging to precisely control the trade-off due to the stochasticity of the training process and the high dimentional gradient vectors. Thus, we propose Controllable Pareto Trade-off (CPT) that can effectively train models to perform different trade-offs according to users' preferences. CPT 1) stabilizes the fairness update with a moving average of stochastic gradients to determine the update direction, and 2) prunes the gradients by only keeping the gradients of the critical parameters. We evaluate CPT on hate speech detection and occupation classification tasks. Experiments show that CPT can achieve a higher-quality set of solutions on the Pareto front than the baseline methods. It also exhibits better controllability and can precisely follow the human-defined reference vectors.", "date": "2025-09-17", "categories": []}
{"id": "2509.13803", "title": "Measuring Gender Bias in Job Title Matching for Grammatical Gender Languages", "abstract": "This work sets the ground for studying how explicit grammatical gender assignment in job titles can affect the results of automatic job ranking systems. We propose the usage of metrics for ranking comparison controlling for gender to evaluate gender bias in job title ranking systems, in particular RBO (Rank-Biased Overlap). We generate and share test sets for a job title matching task in four grammatical gender languages, including occupations in masculine and feminine form and annotated by gender and matching relevance. We use the new test sets and the proposed methodology to evaluate the gender bias of several out-of-the-box multilingual models to set as baselines, showing that all of them exhibit varying degrees of gender bias.", "date": "2025-09-17", "categories": []}
{"id": "2505.16467", "title": "Reading Between the Prompts: How Stereotypes Shape LLM's Implicit Personalization", "abstract": "Generative Large Language Models (LLMs) infer user's demographic information from subtle cues in the conversation -- a phenomenon called implicit personalization. Prior work has shown that such inferences can lead to lower quality responses for users assumed to be from minority groups, even when no demographic information is explicitly provided. In this work, we systematically explore how LLMs respond to stereotypical cues using controlled synthetic conversations, by analyzing the models' latent user representations through both model internals and generated answers to targeted user questions. Our findings reveal that LLMs do infer demographic attributes based on these stereotypical signals, which for a number of groups even persists when the user explicitly identifies with a different demographic group. Finally, we show that this form of stereotype-driven implicit personalization can be effectively mitigated by intervening on the model's internal representations using a trained linear probe to steer them toward the explicitly stated identity. Our results highlight the need for greater transparency and control in how LLMs represent user identity.", "date": "2025-05-22", "categories": []}
{"id": "2509.13165", "title": "On the Correlation between Individual Fairness and Predictive Accuracy in Probabilistic Models", "abstract": "We investigate individual fairness in generative probabilistic classifiers by analysing the robustness of posterior inferences to perturbations in private features. Building on established results in robustness analysis, we hypothesise a correlation between robustness and predictive accuracy, specifically, instances exhibiting greater robustness are more likely to be classified accurately. We empirically assess this hypothesis using a benchmark of fourteen datasets with fairness concerns, employing Bayesian networks as the underlying generative models. To address the computational complexity associated with robustness analysis over multiple private features with Bayesian networks, we reformulate the problem as a most probable explanation task in an auxiliary Markov random field. Our experiments confirm the hypothesis about the correlation, suggesting novel directions to mitigate the traditional trade-off between fairness and accuracy.", "date": "2025-09-16", "categories": []}
{"id": "2406.09070", "title": "FairCoT: Enhancing Fairness in Text-to-Image Generation via Chain of Thought Reasoning with Multimodal Large Language Models", "abstract": "In the domain of text-to-image generative models, biases inherent in training datasets often propagate into generated content, posing significant ethical challenges, particularly in socially sensitive contexts. We introduce FairCoT, a novel framework that enhances fairness in text to image models through Chain of Thought (CoT) reasoning within multimodal generative large language models. FairCoT employs iterative CoT refinement to systematically mitigate biases, and dynamically adjusts textual prompts in real time, ensuring diverse and equitable representation in generated images. By integrating iterative reasoning processes, FairCoT addresses the limitations of zero shot CoT in sensitive scenarios, balancing creativity with ethical responsibility. Experimental evaluations across popular text-to-image systems including DALLE and various Stable Diffusion variants, demonstrate that FairCoT significantly enhances fairness and diversity without sacrificing image quality or semantic fidelity. By combining robust reasoning, lightweight deployment, and extensibility to multiple models, FairCoT represents a promising step toward more socially responsible and transparent AI driven content generation.", "date": "2024-06-13", "categories": []}
{"id": "2508.17610", "title": "Less Is More? Examining Fairness in Pruned Large Language Models for Summarising Opinions", "abstract": "Model compression through post-training pruning offers a way to reduce model size and computational requirements without significantly impacting model performance. However, the effect of pruning on the fairness of LLM-generated summaries remains unexplored, particularly for opinion summarisation where biased outputs could influence public views.In this paper, we present a comprehensive empirical analysis of opinion summarisation, examining three state-of-the-art pruning methods and various calibration sets across three open-source LLMs using four fairness metrics. Our systematic analysis reveals that pruning methods have a greater impact on fairness than calibration sets. Building on these insights, we propose High Gradient Low Activation (HGLA) pruning, which identifies and removes parameters that are redundant for input processing but influential in output generation. Our experiments demonstrate that HGLA can better maintain or even improve fairness compared to existing methods, showing promise across models and tasks where traditional methods have limitations. Our human evaluation shows HGLA-generated outputs are fairer than existing state-of-the-art pruning methods. Code is available at: https://github.com/amberhuang01/HGLA.", "date": "2025-08-25", "categories": []}
{"id": "2509.11184", "title": "The Impact of Skin Tone Label Granularity on the Performance and Fairness of AI Based Dermatology Image Classification Models", "abstract": "Artificial intelligence (AI) models to automatically classify skin lesions from dermatology images have shown promising performance but also susceptibility to bias by skin tone. The most common way of representing skin tone information is the Fitzpatrick Skin Tone (FST) scale. The FST scale has been criticised for having greater granularity in its skin tone categories for lighter-skinned subjects. This paper conducts an investigation of the impact (on performance and bias) on AI classification models of granularity in the FST scale. By training multiple AI models to classify benign vs. malignant lesions using FST-specific data of differing granularity, we show that: (i) when training models using FST-specific data based on three groups (FST 1/2, 3/4 and 5/6), performance is generally better for models trained on FST-specific data compared to a general model trained on FST-balanced data; (ii) reducing the granularity of FST scale information (from 1/2 and 3/4 to 1/2/3/4) can have a detrimental effect on performance. Our results highlight the importance of the granularity of FST groups when training lesion classification models. Given the question marks over possible human biases in the choice of categories in the FST scale, this paper provides evidence for a move away from the FST scale in fair AI research and a transition to an alternative scale that better represents the diversity of human skin tones.", "date": "2025-09-14", "categories": []}
{"id": "2509.11847", "title": "Transparent and Fair Profiling in Employment Services: Evidence from Switzerland", "abstract": "Long-term unemployment (LTU) is a challenge for both jobseekers and public employment services. Statistical profiling tools are increasingly used to predict LTU risk. Some profiling tools are opaque, black-box machine learning models, which raise issues of transparency and fairness. This paper investigates whether interpretable models could serve as an alternative, using administrative data from Switzerland. Traditional statistical, interpretable, and black-box models are compared in terms of predictive performance, interpretability, and fairness. It is shown that explainable boosting machines, a recent interpretable model, perform nearly as well as the best black-box models. It is also shown how model sparsity, feature smoothing, and fairness mitigation can enhance transparency and fairness with only minor losses in performance. These findings suggest that interpretable profiling provides an accountable and trustworthy alternative to black-box models without compromising performance.", "date": "2025-09-15", "categories": []}
{"id": "2402.07586", "title": "Unveiling Group-Specific Distributed Concept Drift: A Fairness Imperative in Federated Learning", "abstract": "In the evolving field of machine learning, ensuring group fairness has become a critical concern, prompting the development of algorithms designed to mitigate bias in decision-making processes. Group fairness refers to the principle that a model's decisions should be equitable across different groups defined by sensitive attributes such as gender or race, ensuring that individuals from privileged groups and unprivileged groups are treated fairly and receive similar outcomes. However, achieving fairness in the presence of group-specific concept drift remains an unexplored frontier, and our research represents pioneering efforts in this regard. Group-specific concept drift refers to situations where one group experiences concept drift over time while another does not, leading to a decrease in fairness even if accuracy remains fairly stable. Within the framework of Federated Learning, where clients collaboratively train models, its distributed nature further amplifies these challenges since each client can experience group-specific concept drift independently while still sharing the same underlying concept, creating a complex and dynamic environment for maintaining fairness. The most significant contribution of our research is the formalization and introduction of the problem of group-specific concept drift and its distributed counterpart, shedding light on its critical importance in the field of fairness. Additionally, leveraging insights from prior research, we adapt an existing distributed concept drift adaptation algorithm to tackle group-specific distributed concept drift which uses a multi-model approach, a local group-specific drift detection mechanism, and continuous clustering of models over time. The findings from our experiments highlight the importance of addressing group-specific concept drift and its distributed counterpart to advance fairness in machine learning.", "date": "2024-02-12", "categories": []}
{"id": "2410.03855", "title": "A Survey on Group Fairness in Federated Learning: Challenges, Taxonomy of Solutions and Directions for Future Research", "abstract": "Group fairness in machine learning is an important area of research focused on achieving equitable outcomes across different groups defined by sensitive attributes such as race or gender. Federated Learning, a decentralized approach to training machine learning models across multiple clients, amplifies the need for fairness methodologies due to its inherent heterogeneous data distributions that can exacerbate biases. The intersection of Federated Learning and group fairness has attracted significant interest, with 48 research works specifically dedicated to addressing this issue. However, no comprehensive survey has specifically focused on group fairness in Federated Learning. In this work, we analyze the key challenges of this topic, propose practices for its identification and benchmarking, and create a novel taxonomy based on criteria such as data partitioning, location, and strategy. Furthermore, we analyze broader concerns, review how different approaches handle the complexities of various sensitive attributes, examine common datasets and applications, and discuss the ethical, legal, and policy implications of group fairness in FL. We conclude by highlighting key areas for future research, emphasizing the need for more methods to address the complexities of achieving group fairness in federated systems.", "date": "2024-10-04", "categories": []}
{"id": "2505.19317", "title": "Effort-aware Fairness: Incorporating a Philosophy-informed, Human-centered Notion of Effort into Algorithmic Fairness Metrics", "abstract": "Although popularized AI fairness metrics, e.g., demographic parity, have uncovered bias in AI-assisted decision-making outcomes, they do not consider how much effort one has spent to get to where one is today in the input feature space. However, the notion of effort is important in how Philosophy and humans understand fairness. We propose a philosophy-informed approach to conceptualize and evaluate Effort-aware Fairness (EaF), grounded in the concept of Force, which represents the temporal trajectory of predictive features coupled with inertia. Besides theoretical formulation, our empirical contributions include: (1) a pre-registered human subjects experiment, which shows that for both stages of the (individual) fairness evaluation process, people consider the temporal trajectory of a predictive feature more than its aggregate value; (2) pipelines to compute Effort-aware Individual/Group Fairness in the criminal justice and personal finance contexts. Our work may enable AI model auditors to uncover and potentially correct unfair decisions against individuals who have spent significant efforts to improve but are still stuck with systemic disadvantages outside their control.", "date": "2025-05-25", "categories": []}
{"id": "2509.08004", "title": "Evaluating and comparing gender bias across four text-to-image models", "abstract": "As we increasingly use Artificial Intelligence (AI) in decision-making for industries like healthcare, finance, e-commerce, and even entertainment, it is crucial to also reflect on the ethical aspects of AI, for example the inclusivity and fairness of the information it provides. In this work, we aimed to evaluate different text-to-image AI models and compare the degree of gender bias they present. The evaluated models were Stable Diffusion XL (SDXL), Stable Diffusion Cascade (SC), DALL-E and Emu. We hypothesized that DALL-E and Stable Diffusion, which are comparatively older models, would exhibit a noticeable degree of gender bias towards men, while Emu, which was recently released by Meta AI, would have more balanced results. As hypothesized, we found that both Stable Diffusion models exhibit a noticeable degree of gender bias while Emu demonstrated more balanced results (i.e. less gender bias). However, interestingly, Open AI's DALL-E exhibited almost opposite results, such that the ratio of women to men was significantly higher in most cases tested. Here, although we still observed a bias, the bias favored females over males. This bias may be explained by the fact that OpenAI changed the prompts at its backend, as observed during our experiment. We also observed that Emu from Meta AI utilized user information while generating images via WhatsApp. We also proposed some potential solutions to avoid such biases, including ensuring diversity across AI research teams and having diverse datasets.", "date": "2025-09-07", "categories": []}
{"id": "2509.08156", "title": "MMM-fair: An Interactive Toolkit for Exploring and Operationalizing Multi-Fairness Trade-offs", "abstract": "Fairness-aware classification requires balancing performance and fairness, often intensified by intersectional biases. Conflicting fairness definitions further complicate the task, making it difficult to identify universally fair solutions. Despite growing regulatory and societal demands for equitable AI, popular toolkits offer limited support for exploring multi-dimensional fairness and related trade-offs. To address this, we present mmm-fair, an open-source toolkit leveraging boosting-based ensemble approaches that dynamically optimizes model weights to jointly minimize classification errors and diverse fairness violations, enabling flexible multi-objective optimization. The system empowers users to deploy models that align with their context-specific needs while reliably uncovering intersectional biases often missed by state-of-the-art methods. In a nutshell, mmm-fair uniquely combines in-depth multi-attribute fairness, multi-objective optimization, a no-code, chat-based interface, LLM-powered explanations, interactive Pareto exploration for model selection, custom fairness constraint definition, and deployment-ready models in a single open-source toolkit, a combination rarely found in existing fairness tools. Demo walkthrough available at: https://youtu.be/_rcpjlXFqkw.", "date": "2025-09-09", "categories": []}
{"id": "2509.07050", "title": "Automated Evaluation of Gender Bias Across 13 Large Multimodal Models", "abstract": "Large multimodal models (LMMs) have revolutionized text-to-image generation, but they risk perpetuating the harmful social biases in their training data. Prior work has identified gender bias in these models, but methodological limitations prevented large-scale, comparable, cross-model analysis. To address this gap, we introduce the Aymara Image Fairness Evaluation, a benchmark for assessing social bias in AI-generated images. We test 13 commercially available LMMs using 75 procedurally-generated, gender-neutral prompts to generate people in stereotypically-male, stereotypically-female, and non-stereotypical professions. We then use a validated LLM-as-a-judge system to score the 965 resulting images for gender representation. Our results reveal (p < .001 for all): 1) LMMs systematically not only reproduce but actually amplify occupational gender stereotypes relative to real-world labor data, generating men in 93.0% of images for male-stereotyped professions but only 22.5% for female-stereotyped professions; 2) Models exhibit a strong default-male bias, generating men in 68.3% of the time for non-stereotyped professions; and 3) The extent of bias varies dramatically across models, with overall male representation ranging from 46.7% to 73.3%. Notably, the top-performing model de-amplified gender stereotypes and approached gender parity, achieving the highest fairness scores. This variation suggests high bias is not an inevitable outcome but a consequence of design choices. Our work provides the most comprehensive cross-model benchmark of gender bias to date and underscores the necessity of standardized, automated evaluation tools for promoting accountability and fairness in AI development.", "date": "2025-09-08", "categories": []}
{"id": "2509.07222", "title": "Explaining How Quantization Disparately Skews a Model", "abstract": "Post Training Quantization (PTQ) is widely adopted due to its high compression capacity and speed with minimal impact on accuracy. However, we observed that disparate impacts are exacerbated by quantization, especially for minority groups. Our analysis explains that in the course of quantization there is a chain of factors attributed to a disparate impact across groups during forward and backward passes. We explore how the changes in weights and activations induced by quantization cause cascaded impacts in the network, resulting in logits with lower variance, increased loss, and compromised group accuracies. We extend our study to verify the influence of these impacts on group gradient norms and eigenvalues of the Hessian matrix, providing insights into the state of the network from an optimization point of view. To mitigate these effects, we propose integrating mixed precision Quantization Aware Training (QAT) with dataset sampling methods and weighted loss functions, therefore providing fair deployment of quantized neural networks.", "date": "2025-09-08", "categories": []}
{"id": "2509.07889", "title": "From Detection to Mitigation: Addressing Gender Bias in Chinese Texts via Efficient Tuning and Voting-Based Rebalancing", "abstract": "This paper presents our team's solution to Shared Task 7 of NLPCC-2025, which focuses on sentence-level gender bias detection and mitigation in Chinese. The task aims to promote fairness and controllability in natural language generation by automatically detecting, classifying, and mitigating gender bias. To address this challenge, we adopt a fine-tuning approach based on large language models (LLMs), efficiently adapt to the bias detection task via Low-Rank Adaptation (LoRA). In terms of data processing, we construct a more balanced training set to alleviate class imbalance and introduce heterogeneous samples from multiple sources to enhance model generalization. For the detection and classification sub-tasks, we employ a majority voting strategy that integrates outputs from multiple expert models to boost performance. Additionally, to improve bias generation detection and mitigation, we design a multi-temperature sampling mechanism to capture potential variations in bias expression styles. Experimental results demonstrate the effectiveness of our approach in bias detection, classification, and mitigation. Our method ultimately achieves an average score of 47.90%, ranking fourth in the shared task.", "date": "2025-09-09", "categories": []}
{"id": "2410.22591", "title": "FACEGroup: Feasible and Actionable Counterfactual Explanations for Group Fairness", "abstract": "Counterfactual explanations assess unfairness by revealing how inputs must change to achieve a desired outcome. This paper introduces the first graph-based framework for generating group counterfactual explanations to audit group fairness, a key aspect of trustworthy machine learning. Our framework, FACEGroup (Feasible and Actionable Counterfactual Explanations for Group Fairness), models real-world feasibility constraints, identifies subgroups with similar counterfactuals, and captures key trade-offs in counterfactual generation, distinguishing it from existing methods. To evaluate fairness, we introduce novel metrics for both group and subgroup level analysis that explicitly account for these trade-offs. Experiments on benchmark datasets show that FACEGroup effectively generates feasible group counterfactuals while accounting for trade-offs, and that our metrics capture and quantify fairness disparities.", "date": "2024-10-29", "categories": []}
{"id": "2508.20401", "title": "Revealing Potential Biases in LLM-Based Recommender Systems in the Cold Start Setting", "abstract": "Large Language Models (LLMs) are increasingly used for recommendation tasks due to their general-purpose capabilities. While LLMs perform well in rich-context settings, their behavior in cold-start scenarios, where only limited signals such as age, gender, or language are available, raises fairness concerns because they may rely on societal biases encoded during pretraining. We introduce a benchmark specifically designed to evaluate fairness in zero-context recommendation. Our modular pipeline supports configurable recommendation domains and sensitive attributes, enabling systematic and flexible audits of any open-source LLM. Through evaluations of state-of-the-art models (Gemma 3 and Llama 3.2), we uncover consistent biases across recommendation domains (music, movies, and colleges) including gendered and cultural stereotypes. We also reveal a non-linear relationship between model size and fairness, highlighting the need for nuanced analysis.", "date": "2025-08-28", "categories": []}
{"id": "2509.05627", "title": "Audits Under Resource, Data, and Access Constraints: Scaling Laws For Less Discriminatory Alternatives", "abstract": "AI audits play a critical role in AI accountability and safety. One branch of the law for which AI audits are particularly salient is anti-discrimination law. Several areas of anti-discrimination law implicate the \"less discriminatory alternative\" (LDA) requirement, in which a protocol (e.g., model) is defensible if no less discriminatory protocol that achieves comparable performance can be found with a reasonable amount of effort. Notably, the burden of proving an LDA exists typically falls on the claimant (the party alleging discrimination). This creates a significant hurdle in AI cases, as the claimant would seemingly need to train a less discriminatory yet high-performing model, a task requiring resources and expertise beyond most litigants. Moreover, developers often shield information about and access to their model and training data as trade secrets, making it difficult to reproduce a similar model from scratch.\n  In this work, we present a procedure enabling claimants to determine if an LDA exists, even when they have limited compute, data, information, and model access. We focus on the setting in which fairness is given by demographic parity and performance by binary cross-entropy loss. As our main result, we provide a novel closed-form upper bound for the loss-fairness Pareto frontier (PF). We show how the claimant can use it to fit a PF in the \"low-resource regime,\" then extrapolate the PF that applies to the (large) model being contested, all without training a single large model. The expression thus serves as a scaling law for loss-fairness PFs. To use this scaling law, the claimant would require a small subsample of the train/test data. Then, the claimant can fit the context-specific PF by training as few as 7 (small) models. We stress test our main result in simulations, finding that our scaling law holds even when the exact conditions of our theory do not.", "date": "2025-09-06", "categories": []}
{"id": "2509.04515", "title": "Mitigation of Gender and Ethnicity Bias in AI-Generated Stories through Model Explanations", "abstract": "Language models have been shown to propagate social bias through their output, particularly in the representation of gender and ethnicity. This paper investigates gender and ethnicity biases in AI-generated occupational stories. Representation biases are measured before and after applying our proposed mitigation strategy, Bias Analysis and Mitigation through Explanation (BAME), revealing improvements in demographic representation ranging from 2% to 20%. BAME leverages model-generated explanations to inform targeted prompt engineering, effectively reducing biases without modifying model parameters. By analyzing stories generated across 25 occupational groups, three large language models (Claude 3.5 Sonnet, Llama 3.1 70B Instruct, and GPT-4 Turbo), and multiple demographic dimensions, we identify persistent patterns of overrepresentation and underrepresentation linked to training data stereotypes. Our findings demonstrate that guiding models with their own internal reasoning mechanisms can significantly enhance demographic parity, thereby contributing to the development of more transparent generative AI systems.", "date": "2025-09-03", "categories": []}
{"id": "2305.06166", "title": "Mitigating Bias in Text Classification via Prompt-Based Text Transformation", "abstract": "The presence of specific linguistic signals particular to a certain sub-group can become highly salient to language models during training. In automated decision-making settings, this may lead to biased outcomes when models rely on cues that correlate with protected characteristics. We investigate whether prompting ChatGPT to rewrite text using simplification, neutralisation, localisation, and formalisation can reduce demographic signals while preserving meaning. Experimental results show a statistically significant drop in location classification accuracy across multiple models after transformation, suggesting reduced reliance on group-specific language. At the same time, sentiment analysis and rating prediction tasks confirm that the core meaning of the reviews remains greatly intact. These results suggest that prompt-based rewriting offers a practical and generalisable approach for mitigating bias in text classification.", "date": "2023-05-09", "categories": []}
{"id": "2501.04316", "title": "Small Changes, Large Consequences: Analyzing the Allocational Fairness of LLMs in Hiring Contexts", "abstract": "Large language models (LLMs) are increasingly being deployed in high-stakes applications like hiring, yet their potential for unfair decision-making remains understudied in generative and retrieval settings. In this work, we examine the allocational fairness of LLM-based hiring systems through two tasks that reflect actual HR usage: resume summarization and applicant ranking. By constructing a synthetic resume dataset with controlled perturbations and curating job postings, we investigate whether model behavior differs across demographic groups. Our findings reveal that generated summaries exhibit meaningful differences more frequently for race than for gender perturbations. Models also display non-uniform retrieval selection patterns across demographic groups and exhibit high ranking sensitivity to both gender and race perturbations. Surprisingly, retrieval models can show comparable sensitivity to both demographic and non-demographic changes, suggesting that fairness issues may stem from broader model brittleness. Overall, our results indicate that LLM-based hiring systems, especially in the retrieval stage, can exhibit notable biases that lead to discriminatory outcomes in real-world contexts.", "date": "2025-01-08", "categories": []}
{"id": "2508.18826", "title": "SWiFT: Soft-Mask Weight Fine-tuning for Bias Mitigation", "abstract": "Recent studies have shown that Machine Learning (ML) models can exhibit bias in real-world scenarios, posing significant challenges in ethically sensitive domains such as healthcare. Such bias can negatively affect model fairness, model generalization abilities and further risks amplifying social discrimination. There is a need to remove biases from trained models. Existing debiasing approaches often necessitate access to original training data and need extensive model retraining; they also typically exhibit trade-offs between model fairness and discriminative performance. To address these challenges, we propose Soft-Mask Weight Fine-Tuning (SWiFT), a debiasing framework that efficiently improves fairness while preserving discriminative performance with much less debiasing costs. Notably, SWiFT requires only a small external dataset and only a few epochs of model fine-tuning. The idea behind SWiFT is to first find the relative, and yet distinct, contributions of model parameters to both bias and predictive performance. Then, a two-step fine-tuning process updates each parameter with different gradient flows defined by its contribution. Extensive experiments with three bias sensitive attributes (gender, skin tone, and age) across four dermatological and two chest X-ray datasets demonstrate that SWiFT can consistently reduce model bias while achieving competitive or even superior diagnostic accuracy under common fairness and accuracy metrics, compared to the state-of-the-art. Specifically, we demonstrate improved model generalization ability as evidenced by superior performance on several out-of-distribution (OOD) datasets.", "date": "2025-08-26", "categories": []}
{"id": "2509.03616", "title": "Multi Attribute Bias Mitigation via Representation Learning", "abstract": "Real world images frequently exhibit multiple overlapping biases, including textures, watermarks, gendered makeup, scene object pairings, etc. These biases collectively impair the performance of modern vision models, undermining both their robustness and fairness. Addressing these biases individually proves inadequate, as mitigating one bias often permits or intensifies others. We tackle this multi bias problem with Generalized Multi Bias Mitigation (GMBM), a lean two stage framework that needs group labels only while training and minimizes bias at test time. First, Adaptive Bias Integrated Learning (ABIL) deliberately identifies the influence of known shortcuts by training encoders for each attribute and integrating them with the main backbone, compelling the classifier to explicitly recognize these biases. Then Gradient Suppression Fine Tuning prunes those very bias directions from the backbone's gradients, leaving a single compact network that ignores all the shortcuts it just learned to recognize. Moreover we find that existing bias metrics break under subgroup imbalance and train test distribution shifts, so we introduce Scaled Bias Amplification (SBA): a test time measure that disentangles model induced bias amplification from distributional differences. We validate GMBM on FB CMNIST, CelebA, and COCO, where we boost worst group accuracy, halve multi attribute bias amplification, and set a new low in SBA even as bias complexity and distribution shifts intensify, making GMBM the first practical, end to end multibias solution for visual recognition. Project page: http://visdomlab.github.io/GMBM/", "date": "2025-09-03", "categories": []}
{"id": "2509.04295", "title": "A Primer on Causal and Statistical Dataset Biases for Fair and Robust Image Analysis", "abstract": "Machine learning methods often fail when deployed in the real world. Worse still, they fail in high-stakes situations and across socially sensitive lines. These issues have a chilling effect on the adoption of machine learning methods in settings such as medical diagnosis, where they are arguably best-placed to provide benefits if safely deployed. In this primer, we introduce the causal and statistical structures which induce failure in machine learning methods for image analysis. We highlight two previously overlooked problems, which we call the \\textit{no fair lunch} problem and the \\textit{subgroup separability} problem. We elucidate why today's fair representation learning methods fail to adequately solve them and propose potential paths forward for the field.", "date": "2025-09-04", "categories": []}
{"id": "2504.14212", "title": "Bias Analysis and Mitigation through Protected Attribute Detection and Regard Classification", "abstract": "Large language models (LLMs) acquire general linguistic knowledge from massive-scale pretraining. However, pretraining data mainly comprised of web-crawled texts contain undesirable social biases which can be perpetuated or even amplified by LLMs. In this study, we propose an efficient yet effective annotation pipeline to investigate social biases in the pretraining corpora. Our pipeline consists of protected attribute detection to identify diverse demographics, followed by regard classification to analyze the language polarity towards each attribute. Through our experiments, we demonstrate the effect of our bias analysis and mitigation measures, focusing on Common Crawl as the most representative pretraining corpus.", "date": "2025-04-19", "categories": []}
{"id": "2509.03204", "title": "Exploring the Design Space of Fair Tree Learning Algorithms", "abstract": "Decision trees have been studied extensively in the context of fairness, aiming to maximize prediction performance while ensuring non-discrimination against different groups. Techniques in this space usually focus on imposing constraints at training time, constraining the search space so that solutions which display unacceptable values of relevant metrics are not considered, discarded, or discouraged. If we assume one target variable y and one sensitive attribute s, the design space of tree learning algorithms can be spanned as follows: (i) One can have one tree T that is built using an objective function that is a function of y, s, and T. For instance, one can build a tree based on the weighted information gain regarding y (maximizing) and s (minimizing). (ii) The second option is to have one tree model T that uses an objective function in y and T and a constraint on s and T. Here, s is no longer part of the objective, but part of a constraint. This can be achieved greedily by aborting a further split as soon as the condition that optimizes the objective in y fails to satisfy the constraint on s. A simple way to explore other splits is to backtrack during tree construction once a fairness constraint is violated. (iii) The third option is to have two trees T_y and T_s, one for y and one for s, such that the tree structure for y and s does not have to be shared. In this way, information regarding y and regarding s can be used independently, without having to constrain the choices in tree construction by the mutual information between the two variables. Quite surprisingly, of the three options, only the first one and the greedy variant of the second have been studied in the literature so far. In this paper, we introduce the above two additional options from that design space and characterize them experimentally on multiple datasets.", "date": "2025-09-03", "categories": []}
{"id": "2312.10128", "title": "An Information-Flow Perspective on Algorithmic Fairness", "abstract": "This work presents insights gained by investigating the relationship between algorithmic fairness and the concept of secure information flow. The problem of enforcing secure information flow is well-studied in the context of information security: If secret information may \"flow\" through an algorithm or program in such a way that it can influence the program's output, then that is considered insecure information flow as attackers could potentially observe (parts of) the secret.\n  There is a strong correspondence between secure information flow and algorithmic fairness: if protected attributes such as race, gender, or age are treated as secret program inputs, then secure information flow means that these ``secret'' attributes cannot influence the result of a program. While most research in algorithmic fairness evaluation concentrates on studying the impact of algorithms (often treating the algorithm as a black-box), the concepts derived from information flow can be used both for the analysis of disparate treatment as well as disparate impact w.r.t. a structural causal model.\n  In this paper, we examine the relationship between quantitative as well as qualitative information-flow properties and fairness. Moreover, based on this duality, we derive a new quantitative notion of fairness called fairness spread, which can be easily analyzed using quantitative information flow and which strongly relates to counterfactual fairness. We demonstrate that off-the-shelf tools for information-flow properties can be used in order to formally analyze a program's algorithmic fairness properties, including the new notion of fairness spread as well as established notions such as demographic parity.", "date": "2023-12-15", "categories": []}
{"id": "2502.03953", "title": "Fairness Aware Reinforcement Learning via Proximal Policy Optimization", "abstract": "Fairness in multi-agent systems (MAS) focuses on equitable reward distribution among agents in scenarios involving sensitive attributes such as race, gender, or socioeconomic status. This paper introduces fairness in Proximal Policy Optimization (PPO) with a penalty term derived from a fairness definition such as demographic parity, counterfactual fairness, or conditional statistical parity. The proposed method, which we call Fair-PPO, balances reward maximisation with fairness by integrating two penalty components: a retrospective component that minimises disparities in past outcomes and a prospective component that ensures fairness in future decision-making. We evaluate our approach in two games: the Allelopathic Harvest, a cooperative and competitive MAS focused on resource collection, where some agents possess a sensitive attribute, and HospitalSim, a hospital simulation, in which agents coordinate the operations of hospital patients with different mobility and priority needs. Experiments show that Fair-PPO achieves fairer policies than PPO across the fairness metrics and, through the retrospective and prospective penalty components, reveals a wide spectrum of strategies to improve fairness; at the same time, its performance pairs with that of state-of-the-art fair reinforcement-learning algorithms. Fairness comes at the cost of reduced efficiency, but does not compromise equality among the overall population (Gini index). These findings underscore the potential of Fair-PPO to address fairness challenges in MAS.", "date": "2025-02-06", "categories": []}
{"id": "2509.00109", "title": "Bias Mitigation for AI-Feedback Loops in Recommender Systems: A Systematic Literature Review and Taxonomy", "abstract": "Recommender systems continually retrain on user reactions to their own predictions, creating AI feedback loops that amplify biases and diminish fairness over time. Despite this well-known risk, most bias mitigation techniques are tested only on static splits, so their long-term fairness across multiple retraining rounds remains unclear. We therefore present a systematic literature review of bias mitigation methods that explicitly consider AI feedback loops and are validated in multi-round simulations or live A/B tests. Screening 347 papers yields 24 primary studies published between 2019-2025. Each study is coded on six dimensions: mitigation technique, biases addressed, dynamic testing set-up, evaluation focus, application domain, and ML task, organising them into a reusable taxonomy. The taxonomy offers industry practitioners a quick checklist for selecting robust methods and gives researchers a clear roadmap to the field's most urgent gaps. Examples include the shortage of shared simulators, varying evaluation metrics, and the fact that most studies report either fairness or performance; only six use both.", "date": "2025-08-28", "categories": []}
{"id": "2509.00658", "title": "Face4FairShifts: A Large Image Benchmark for Fairness and Robust Learning across Visual Domains", "abstract": "Ensuring fairness and robustness in machine learning models remains a challenge, particularly under domain shifts. We present Face4FairShifts, a large-scale facial image benchmark designed to systematically evaluate fairness-aware learning and domain generalization. The dataset includes 100,000 images across four visually distinct domains with 39 annotations within 14 attributes covering demographic and facial features. Through extensive experiments, we analyze model performance under distribution shifts and identify significant gaps. Our findings emphasize the limitations of existing related datasets and the need for more effective fairness-aware domain adaptation techniques. Face4FairShifts provides a comprehensive testbed for advancing equitable and reliable AI systems. The dataset is available online at https://meviuslab.github.io/Face4FairShifts/.", "date": "2025-08-31", "categories": []}
{"id": "2509.00799", "title": "Fairness in Federated Learning: Trends, Challenges, and Opportunities", "abstract": "At the intersection of the cutting-edge technologies and privacy concerns, Federated Learning (FL) with its distributed architecture, stands at the forefront in a bid to facilitate collaborative model training across multiple clients while preserving data privacy. However, the applicability of FL systems is hindered by fairness concerns arising from numerous sources of heterogeneity that can result in biases and undermine a system's effectiveness, with skewed predictions, reduced accuracy, and inefficient model convergence. This survey thus explores the diverse sources of bias, including but not limited to, data, client, and model biases, and thoroughly discusses the strengths and limitations inherited within the array of the state-of-the-art techniques utilized in the literature to mitigate such disparities in the FL training process. We delineate a comprehensive overview of the several notions, theoretical underpinnings, and technical aspects associated with fairness and their adoption in FL-based multidisciplinary environments. Furthermore, we examine salient evaluation metrics leveraged to measure fairness quantitatively. Finally, we envisage exciting open research directions that have the potential to drive future advancements in achieving fairer FL frameworks, in turn, offering a strong foundation for future research in this pivotal area.", "date": "2025-08-31", "categories": []}
{"id": "2509.01730", "title": "BM-CL: Bias Mitigation through the lens of Continual Learning", "abstract": "Biases in machine learning pose significant challenges, particularly when models amplify disparities that affect disadvantaged groups. Traditional bias mitigation techniques often lead to a {\\itshape leveling-down effect}, whereby improving outcomes of disadvantaged groups comes at the expense of reduced performance for advantaged groups. This study introduces Bias Mitigation through Continual Learning (BM-CL), a novel framework that leverages the principles of continual learning to address this trade-off. We postulate that mitigating bias is conceptually similar to domain-incremental continual learning, where the model must adjust to changing fairness conditions, improving outcomes for disadvantaged groups without forgetting the knowledge that benefits advantaged groups. Drawing inspiration from techniques such as Learning without Forgetting and Elastic Weight Consolidation, we reinterpret bias mitigation as a continual learning problem. This perspective allows models to incrementally balance fairness objectives, enhancing outcomes for disadvantaged groups while preserving performance for advantaged groups. Experiments on synthetic and real-world image datasets, characterized by diverse sources of bias, demonstrate that the proposed framework mitigates biases while minimizing the loss of original knowledge. Our approach bridges the fields of fairness and continual learning, offering a promising pathway for developing machine learning systems that are both equitable and effective.", "date": "2025-09-01", "categories": []}
{"id": "2509.02007", "title": "mFARM: Towards Multi-Faceted Fairness Assessment based on HARMs in Clinical Decision Support", "abstract": "The deployment of Large Language Models (LLMs) in high-stakes medical settings poses a critical AI alignment challenge, as models can inherit and amplify societal biases, leading to significant disparities. Existing fairness evaluation methods fall short in these contexts as they typically use simplistic metrics that overlook the multi-dimensional nature of medical harms. This also promotes models that are fair only because they are clinically inert, defaulting to safe but potentially inaccurate outputs. To address this gap, our contributions are mainly two-fold: first, we construct two large-scale, controlled benchmarks (ED-Triage and Opioid Analgesic Recommendation) from MIMIC-IV, comprising over 50,000 prompts with twelve race x gender variants and three context tiers. Second, we propose a multi-metric framework - Multi-faceted Fairness Assessment based on hARMs ($mFARM$) to audit fairness for three distinct dimensions of disparity (Allocational, Stability, and Latent) and aggregate them into an $mFARM$ score. We also present an aggregated Fairness-Accuracy Balance (FAB) score to benchmark and observe trade-offs between fairness and prediction accuracy. We empirically evaluate four open-source LLMs (Mistral-7B, BioMistral-7B, Qwen-2.5-7B, Bio-LLaMA3-8B) and their finetuned versions under quantization and context variations. Our findings showcase that the proposed $mFARM$ metrics capture subtle biases more effectively under various settings. We find that most models maintain robust performance in terms of $mFARM$ score across varying levels of quantization but deteriorate significantly when the context is reduced. Our benchmarks and evaluation code are publicly released to enhance research in aligned AI for healthcare.", "date": "2025-09-02", "categories": []}
{"id": "2509.02133", "title": "AMBEDKAR-A Multi-level Bias Elimination through a Decoding Approach with Knowledge Augmentation for Robust Constitutional Alignment of Language Models", "abstract": "Large Language Models (LLMs) can inadvertently reflect societal biases present in their training data, leading to harmful or prejudiced outputs. In the Indian context, our empirical evaluations across a suite of models reveal that biases around caste and religion are particularly salient. Yet, most existing mitigation strategies are Western-centric and fail to address these local nuances. We propose AMBEDKAR, a framework inspired by the egalitarian vision of Dr B. R. Ambedkar, architect of the Indian Constitution, to guide LLM outputs toward fairness, neutrality, and inclusion in line with Articles 14 to 17. Our approach introduces a Constitution-Aware Decoding Layer, guided by the AI Constitution of India and applied only at inference time, without any parameter updates to the base model. We incorporate a speculative decoding algorithm that proactively reduces casteist and communal bias during generation. This mitigation layer operates directly within the decoding process, avoiding changes to model internals and lowering the computational and infrastructural costs associated with retraining. We reinterpret speculative decoding not merely as an efficiency tool but as a mechanism for fairness. In this framework, a Small Language Model (SLM) acts as a potentially biased generator, while a constitutionally guided Large Language Model (LLM) serves as the verifier. Rather than accelerating generation, the LLM enforces bias-robust trajectories in the SLM outputs. This inversion of roles gives rise to a fairness-by-speculation paradigm. Our approach yields an absolute reduction of bias up to 26.41 percent compared to baseline. Our source code, datasets, and results are available at https://anonymous.4open.science/r/AMBEDKAR-983B/", "date": "2025-09-02", "categories": []}
{"id": "2508.21179", "title": "Synthetic CVs To Build and Test Fairness-Aware Hiring Tools", "abstract": "Algorithmic hiring has become increasingly necessary in some sectors as it promises to deal with hundreds or even thousands of applicants. At the heart of these systems are algorithms designed to retrieve and rank candidate profiles, which are usually represented by Curricula Vitae (CVs). Research has shown, however, that such technologies can inadvertently introduce bias, leading to discrimination based on factors such as candidates' age, gender, or national origin. Developing methods to measure, mitigate, and explain bias in algorithmic hiring, as well as to evaluate and compare fairness techniques before deployment, requires sets of CVs that reflect the characteristics of people from diverse backgrounds.\n  However, datasets of these characteristics that can be used to conduct this research do not exist. To address this limitation, this paper introduces an approach for building a synthetic dataset of CVs with features modeled on real materials collected through a data donation campaign. Additionally, the resulting dataset of 1,730 CVs is presented, which we envision as a potential benchmarking standard for research on algorithmic hiring discrimination.", "date": "2025-08-28", "categories": []}
{"id": "2508.21334", "title": "Stairway to Fairness: Connecting Group and Individual Fairness", "abstract": "Fairness in recommender systems (RSs) is commonly categorised into group fairness and individual fairness. However, there is no established scientific understanding of the relationship between the two fairness types, as prior work on both types has used different evaluation measures or evaluation objectives for each fairness type, thereby not allowing for a proper comparison of the two. As a result, it is currently not known how increasing one type of fairness may affect the other. To fill this gap, we study the relationship of group and individual fairness through a comprehensive comparison of evaluation measures that can be used for both fairness types. Our experiments with 8 runs across 3 datasets show that recommendations that are highly fair for groups can be very unfair for individuals. Our finding is novel and useful for RS practitioners aiming to improve the fairness of their systems. Our code is available at: https://github.com/theresiavr/stairway-to-fairness.", "date": "2025-08-29", "categories": []}
{"id": "2508.21512", "title": "Accept or Deny? Evaluating LLM Fairness and Performance in Loan Approval across Table-to-Text Serialization Approaches", "abstract": "Large Language Models (LLMs) are increasingly employed in high-stakes decision-making tasks, such as loan approvals. While their applications expand across domains, LLMs struggle to process tabular data, ensuring fairness and delivering reliable predictions. In this work, we assess the performance and fairness of LLMs on serialized loan approval datasets from three geographically distinct regions: Ghana, Germany, and the United States. Our evaluation focuses on the model's zero-shot and in-context learning (ICL) capabilities. Our results reveal that the choice of serialization (Serialization refers to the process of converting tabular data into text formats suitable for processing by LLMs.) format significantly affects both performance and fairness in LLMs, with certain formats such as GReat and LIFT yielding higher F1 scores but exacerbating fairness disparities. Notably, while ICL improved model performance by 4.9-59.6% relative to zero-shot baselines, its effect on fairness varied considerably across datasets. Our work underscores the importance of effective tabular data representation methods and fairness-aware models to improve the reliability of LLMs in financial decision-making.", "date": "2025-08-29", "categories": []}
{"id": "2508.20881", "title": "Understanding and evaluating computer vision models through the lens of counterfactuals", "abstract": "Counterfactual reasoning -- the practice of asking ``what if'' by varying inputs and observing changes in model behavior -- has become central to interpretable and fair AI. This thesis develops frameworks that use counterfactuals to explain, audit, and mitigate bias in vision classifiers and generative models. By systematically altering semantically meaningful attributes while holding others fixed, these methods uncover spurious correlations, probe causal dependencies, and help build more robust systems.\n  The first part addresses vision classifiers. CAVLI integrates attribution (LIME) with concept-level analysis (TCAV) to quantify how strongly decisions rely on human-interpretable concepts. With localized heatmaps and a Concept Dependency Score, CAVLI shows when models depend on irrelevant cues like backgrounds. Extending this, ASAC introduces adversarial counterfactuals that perturb protected attributes while preserving semantics. Through curriculum learning, ASAC fine-tunes biased models for improved fairness and accuracy while avoiding stereotype-laden artifacts.\n  The second part targets generative Text-to-Image (TTI) models. TIBET provides a scalable pipeline for evaluating prompt-sensitive biases by varying identity-related terms, enabling causal auditing of how race, gender, and age affect image generation. To capture interactions, BiasConnect builds causal graphs diagnosing intersectional biases. Finally, InterMit offers a modular, training-free algorithm that mitigates intersectional bias via causal sensitivity scores and user-defined fairness goals.\n  Together, these contributions show counterfactuals as a unifying lens for interpretability, fairness, and causality in both discriminative and generative models, establishing principled, scalable methods for socially responsible bias evaluation and mitigation.", "date": "2025-08-28", "categories": []}
{"id": "2402.02817", "title": "Bayes-Optimal Fair Classification with Linear Disparity Constraints via Pre-, In-, and Post-processing", "abstract": "Machine learning algorithms may have disparate impacts on protected groups. To address this, we develop methods for Bayes-optimal fair classification, aiming to minimize classification error subject to given group fairness constraints. We introduce the notion of \\emph{linear disparity measures}, which are linear functions of a probabilistic classifier; and \\emph{bilinear disparity measures}, which are also linear in the group-wise regression functions. We show that several popular disparity measures -- the deviations from demographic parity, equality of opportunity, and predictive equality -- are bilinear.\n  We find the form of Bayes-optimal fair classifiers under a single linear disparity measure, by uncovering a connection with the Neyman-Pearson lemma. For bilinear disparity measures, we are able to find the explicit form of Bayes-optimal fair classifiers as group-wise thresholding rules with explicitly characterized thresholds. We develop similar algorithms for when protected attribute cannot be used at the prediction phase. Moreover, we obtain analogous theoretical characterizations of optimal classifiers for a multi-class protected attribute and for equalized odds.\n  Leveraging our theoretical results, we design methods that learn fair Bayes-optimal classifiers under bilinear disparity constraints. Our methods cover three popular approaches to fairness-aware classification, via pre-processing (Fair Up- and Down-Sampling), in-processing (Fair cost-sensitive Classification) and post-processing (a Fair Plug-In Rule). Our methods control disparity directly while achieving near-optimal fairness-accuracy tradeoffs. We show empirically that our methods have state-of-the-art performance compared to existing algorithms. In particular, our pre-processing method can a reach higher accuracy than prior pre-processing methods at low disparity levels.", "date": "2024-02-05", "categories": []}
{"id": "2411.17374", "title": "Understanding Fairness-Accuracy Trade-offs in Machine Learning Models: Does Promoting Fairness Undermine Performance?", "abstract": "Fairness in both Machine Learning (ML) predictions and human decision-making is essential, yet both are susceptible to different forms of bias, such as algorithmic and data-driven in ML, and cognitive or subjective in humans. In this study, we examine fairness using a real-world university admissions dataset comprising 870 applicant profiles, leveraging three ML models: XGB, Bi-LSTM, and KNN, alongside BERT embeddings for textual features. To evaluate individual fairness, we introduce a consistency metric that quantifies agreement in decisions among ML models and human experts with diverse backgrounds. Our analysis reveals that ML models surpass human evaluators in fairness consistency by margins ranging from 14.08\\% to 18.79\\%. Our findings highlight the potential of using ML to enhance fairness in admissions while maintaining high accuracy, advocating a hybrid approach combining human judgement and ML models.", "date": "2024-11-26", "categories": []}
{"id": "2508.19298", "title": "DemoBias: An Empirical Study to Trace Demographic Biases in Vision Foundation Models", "abstract": "Large Vision Language Models (LVLMs) have demonstrated remarkable capabilities across various downstream tasks, including biometric face recognition (FR) with description. However, demographic biases remain a critical concern in FR, as these foundation models often fail to perform equitably across diverse demographic groups, considering ethnicity/race, gender, and age. Therefore, through our work DemoBias, we conduct an empirical evaluation to investigate the extent of demographic biases in LVLMs for biometric FR with textual token generation tasks. We fine-tuned and evaluated three widely used pre-trained LVLMs: LLaVA, BLIP-2, and PaliGemma on our own generated demographic-balanced dataset. We utilize several evaluation metrics, like group-specific BERTScores and the Fairness Discrepancy Rate, to quantify and trace the performance disparities. The experimental results deliver compelling insights into the fairness and reliability of LVLMs across diverse demographic groups. Our empirical study uncovered demographic biases in LVLMs, with PaliGemma and LLaVA exhibiting higher disparities for Hispanic/Latino, Caucasian, and South Asian groups, whereas BLIP-2 demonstrated comparably consistent. Repository: https://github.com/Sufianlab/DemoBias.", "date": "2025-08-25", "categories": []}
{"id": "2508.20021", "title": "FairLoop: Software Support for Human-Centric Fairness in Predictive Business Process Monitoring", "abstract": "Sensitive attributes like gender or age can lead to unfair predictions in machine learning tasks such as predictive business process monitoring, particularly when used without considering context. We present FairLoop1, a tool for human-guided bias mitigation in neural network-based prediction models. FairLoop distills decision trees from neural networks, allowing users to inspect and modify unfair decision logic, which is then used to fine-tune the original model towards fairer predictions. Compared to other approaches to fairness, FairLoop enables context-aware bias removal through human involvement, addressing the influence of sensitive attributes selectively rather than excluding them uniformly.", "date": "2025-08-27", "categories": []}
{"id": "2508.18288", "title": "Toward Responsible ASR for African American English Speakers: A Scoping Review of Bias and Equity in Speech Technology", "abstract": "This scoping literature review examines how fairness, bias, and equity are conceptualized and operationalized in Automatic Speech Recognition (ASR) and adjacent speech and language technologies (SLT) for African American English (AAE) speakers and other linguistically diverse communities. Drawing from 44 peer-reviewed publications across Human-Computer Interaction (HCI), Machine Learning/Natural Language Processing (ML/NLP), and Sociolinguistics, we identify four major areas of inquiry: (1) how researchers understand ASR-related harms; (2) inclusive data practices spanning collection, curation, annotation, and model training; (3) methodological and theoretical approaches to linguistic inclusion; and (4) emerging practices and design recommendations for more equitable systems. While technical fairness interventions are growing, our review highlights a critical gap in governance-centered approaches that foreground community agency, linguistic justice, and participatory accountability. We propose a governance-centered ASR lifecycle as an emergent interdisciplinary framework for responsible ASR development and offer implications for researchers, practitioners, and policymakers seeking to address language marginalization in speech AI systems.", "date": "2025-08-20", "categories": []}
{"id": "2505.16638", "title": "Reconsidering Fairness Through Unawareness From the Perspective of Model Multiplicity", "abstract": "Fairness through Unawareness (FtU) describes the idea that discrimination against demographic groups can be avoided by not considering group membership in the decisions or predictions. This idea has long been criticized in the machine learning literature as not being sufficient to ensure fairness. In addition, the use of additional features is typically thought to increase the accuracy of the predictions for all groups, so that FtU is sometimes thought to be detrimental to all groups. In this paper, we show both theoretically and empirically that FtU can reduce algorithmic discrimination without necessarily reducing accuracy. We connect this insight with the literature on Model Multiplicity, to which we contribute with novel theoretical and empirical results. Furthermore, we illustrate how, in a real-life application, FtU can contribute to the deployment of more equitable policies without losing efficacy. Our findings suggest that FtU is worth considering in practical applications, particularly in high-risk scenarios, and that the use of protected attributes such as gender in predictive models should be accompanied by a clear and well-founded justification.", "date": "2025-05-22", "categories": []}
{"id": "2508.16673", "title": "Invisible Filters: Cultural Bias in Hiring Evaluations Using Large Language Models", "abstract": "Artificial Intelligence (AI) is increasingly used in hiring, with large language models (LLMs) having the potential to influence or even make hiring decisions. However, this raises pressing concerns about bias, fairness, and trust, particularly across diverse cultural contexts. Despite their growing role, few studies have systematically examined the potential biases in AI-driven hiring evaluation across cultures. In this study, we conduct a systematic analysis of how LLMs assess job interviews across cultural and identity dimensions. Using two datasets of interview transcripts, 100 from UK and 100 from Indian job seekers, we first examine cross-cultural differences in LLM-generated scores for hirability and related traits. Indian transcripts receive consistently lower scores than UK transcripts, even when they were anonymized, with disparities linked to linguistic features such as sentence complexity and lexical diversity. We then perform controlled identity substitutions (varying names by gender, caste, and region) within the Indian dataset to test for name-based bias. These substitutions do not yield statistically significant effects, indicating that names alone, when isolated from other contextual signals, may not influence LLM evaluations. Our findings underscore the importance of evaluating both linguistic and social dimensions in LLM-driven evaluations and highlight the need for culturally sensitive design and accountability in AI-assisted hiring.", "date": "2025-08-21", "categories": []}
{"id": "2508.16748", "title": "FAIRWELL: Fair Multimodal Self-Supervised Learning for Wellbeing Prediction", "abstract": "Early efforts on leveraging self-supervised learning (SSL) to improve machine learning (ML) fairness has proven promising. However, such an approach has yet to be explored within a multimodal context. Prior work has shown that, within a multimodal setting, different modalities contain modality-unique information that can complement information of other modalities. Leveraging on this, we propose a novel subject-level loss function to learn fairer representations via the following three mechanisms, adapting the variance-invariance-covariance regularization (VICReg) method: (i) the variance term, which reduces reliance on the protected attribute as a trivial solution; (ii) the invariance term, which ensures consistent predictions for similar individuals; and (iii) the covariance term, which minimizes correlational dependence on the protected attribute. Consequently, our loss function, coined as FAIRWELL, aims to obtain subject-independent representations, enforcing fairness in multimodal prediction tasks. We evaluate our method on three challenging real-world heterogeneous healthcare datasets (i.e. D-Vlog, MIMIC and MODMA) which contain different modalities of varying length and different prediction tasks. Our findings indicate that our framework improves overall fairness performance with minimal reduction in classification performance and significantly improves on the performance-fairness Pareto frontier.", "date": "2025-08-22", "categories": []}
{"id": "2508.16752", "title": "A Framework for Benchmarking Fairness-Utility Trade-offs in Text-to-Image Models via Pareto Frontiers", "abstract": "Achieving fairness in text-to-image generation demands mitigating social biases without compromising visual fidelity, a challenge critical to responsible AI. Current fairness evaluation procedures for text-to-image models rely on qualitative judgment or narrow comparisons, which limit the capacity to assess both fairness and utility in these models and prevent reproducible assessment of debiasing methods. Existing approaches typically employ ad-hoc, human-centered visual inspections that are both error-prone and difficult to replicate. We propose a method for evaluating fairness and utility in text-to-image models using Pareto-optimal frontiers across hyperparametrization of debiasing methods. Our method allows for comparison between distinct text-to-image models, outlining all configurations that optimize fairness for a given utility and vice-versa. To illustrate our evaluation method, we use Normalized Shannon Entropy and ClipScore for fairness and utility evaluation, respectively. We assess fairness and utility in Stable Diffusion, Fair Diffusion, SDXL, DeCoDi, and FLUX text-to-image models. Our method shows that most default hyperparameterizations of the text-to-image model are dominated solutions in the fairness-utility space, and it is straightforward to find better hyperparameters.", "date": "2025-08-22", "categories": []}
{"id": "2508.17465", "title": "Bias Amplification in Stable Diffusion's Representation of Stigma Through Skin Tones and Their Homogeneity", "abstract": "Text-to-image generators (T2Is) are liable to produce images that perpetuate social stereotypes, especially in regards to race or skin tone. We use a comprehensive set of 93 stigmatized identities to determine that three versions of Stable Diffusion (v1.5, v2.1, and XL) systematically associate stigmatized identities with certain skin tones in generated images. We find that SD XL produces skin tones that are 13.53% darker and 23.76% less red (both of which indicate higher likelihood of societal discrimination) than previous models and perpetuate societal stereotypes associating people of color with stigmatized identities. SD XL also shows approximately 30% less variability in skin tones when compared to previous models and 18.89-56.06% compared to human face datasets. Measuring variability through metrics which directly correspond to human perception suggest a similar pattern, where SD XL shows the least amount of variability in skin tones of people with stigmatized identities and depicts most (60.29%) stigmatized identities as being less diverse than non-stigmatized identities. Finally, SD shows more homogenization of skin tones of racial and ethnic identities compared to other stigmatized or non-stigmatized identities, reinforcing incorrect equivalence of biologically-determined skin tone and socially-constructed racial and ethnic identity. Because SD XL is the largest and most complex model and users prefer its generations compared to other models examined in this study, these findings have implications for the dynamics of bias amplification in T2Is, increasing representational harms and challenges generating diverse images depicting people with stigmatized identities.", "date": "2025-08-24", "categories": []}
{"id": "2508.17477", "title": "A Human-In-The-Loop Approach for Improving Fairness in Predictive Business Process Monitoring", "abstract": "Predictive process monitoring enables organizations to proactively react and intervene in running instances of a business process. Given an incomplete process instance, predictions about the outcome, next activity, or remaining time are created. This is done by powerful machine learning models, which have shown impressive predictive performance. However, the data-driven nature of these models makes them susceptible to finding unfair, biased, or unethical patterns in the data. Such patterns lead to biased predictions based on so-called sensitive attributes, such as the gender or age of process participants. Previous work has identified this problem and offered solutions that mitigate biases by removing sensitive attributes entirely from the process instance. However, sensitive attributes can be used both fairly and unfairly in the same process instance. For example, during a medical process, treatment decisions could be based on gender, while the decision to accept a patient should not be based on gender. This paper proposes a novel, model-agnostic approach for identifying and rectifying biased decisions in predictive business process monitoring models, even when the same sensitive attribute is used both fairly and unfairly. The proposed approach uses a human-in-the-loop approach to differentiate between fair and unfair decisions through simple alterations on a decision tree model distilled from the original prediction model. Our results show that the proposed approach achieves a promising tradeoff between fairness and accuracy in the presence of biased data. All source code and data are publicly available at https://doi.org/10.5281/zenodo.15387576.", "date": "2025-08-24", "categories": []}
{"id": "2508.17735", "title": "SMITE: Enhancing Fairness in LLMs through Optimal In-Context Example Selection via Dynamic Validation", "abstract": "Large Language Models (LLMs) are widely used for downstream tasks such as tabular classification, where ensuring fairness in their outputs is critical for inclusivity, equal representation, and responsible AI deployment. This study introduces a novel approach to enhancing LLM performance and fairness through the concept of a dynamic validation set, which evolves alongside the test set, replacing the traditional static validation approach. We also propose an iterative algorithm, SMITE, to select optimal in-context examples, with each example set validated against its corresponding dynamic validation set. The in-context set with the lowest total error is used as the final demonstration set. Our experiments across four different LLMs show that our proposed techniques significantly improve both predictive accuracy and fairness compared to baseline methods. To our knowledge, this is the first study to apply dynamic validation in the context of in-context learning for LLMs.", "date": "2025-08-25", "categories": []}
{"id": "2508.17750", "title": "From Global to Local: Social Bias Transfer in CLIP", "abstract": "The recycling of contrastive language-image pre-trained (CLIP) models as backbones for a large number of downstream tasks calls for a thorough analysis of their transferability implications, especially their well-documented reproduction of social biases and human stereotypes. How do such biases, learned during pre-training, propagate to downstream applications like visual question answering or image captioning? Do they transfer at all?\n  We investigate this phenomenon, referred to as bias transfer in prior literature, through a comprehensive empirical analysis. Firstly, we examine how pre-training bias varies between global and local views of data, finding that bias measurement is highly dependent on the subset of data on which it is computed. Secondly, we analyze correlations between biases in the pre-trained models and the downstream tasks across varying levels of pre-training bias, finding difficulty in discovering consistent trends in bias transfer. Finally, we explore why this inconsistency occurs, showing that under the current paradigm, representation spaces of different pre-trained CLIPs tend to converge when adapted for downstream tasks. We hope this work offers valuable insights into bias behavior and informs future research to promote better bias mitigation practices.", "date": "2025-08-25", "categories": []}
{"id": "2404.17218", "title": "Prompting Techniques for Reducing Social Bias in LLMs through System 1 and System 2 Cognitive Processes", "abstract": "Dual process theory posits that human cognition arises via two systems. System 1, which is a quick, emotional, and intuitive process, which is subject to cognitive biases, and System 2, is a slow, onerous, and deliberate process. Prior research in LLMs found that using chain-of-thought (CoT) prompting in LLMs, which has been often compared to System 2 reasoning, can lead to reduced gender bias. Along these lines, we investigate the relationship between bias, CoT prompting, a direct debiasing, and dual process theory modeling in LLMs. We compare zero-shot CoT, debiasing, and dual process theory-based prompting strategies on two bias datasets spanning nine different social bias categories. We incorporate human and machine personas to determine whether LLM modeling of the effects of dual process theory exist independent of explicit persona models or are tied to the LLM's modeling of human-like generation. We find that a human persona, debiasing, System 2, and CoT prompting all tend to reduce social biases in LLMs, though the best combination of features depends on the exact model and bias category -- resulting in up to a 33 percent drop in stereotypical judgments by an LLM.", "date": "2024-04-26", "categories": []}
{"id": "2502.02309", "title": "Review of Demographic Fairness in Face Recognition", "abstract": "Demographic fairness in face recognition (FR) has emerged as a critical area of research, given its impact on fairness, equity, and reliability across diverse applications. As FR technologies are increasingly deployed globally, disparities in performance across demographic groups -- such as race, ethnicity, and gender -- have garnered significant attention. These biases not only compromise the credibility of FR systems but also raise ethical concerns, especially when these technologies are employed in sensitive domains. This review consolidates extensive research efforts providing a comprehensive overview of the multifaceted aspects of demographic fairness in FR.\n  We systematically examine the primary causes, datasets, assessment metrics, and mitigation approaches associated with demographic disparities in FR. By categorizing key contributions in these areas, this work provides a structured approach to understanding and addressing the complexity of this issue. Finally, we highlight current advancements and identify emerging challenges that need further investigation. This article aims to provide researchers with a unified perspective on the state-of-the-art while emphasizing the critical need for equitable and trustworthy FR systems.", "date": "2025-02-04", "categories": []}
{"id": "2506.11025", "title": "When Algorithms Play Favorites: Lookism in the Generation and Perception of Faces", "abstract": "This paper examines how synthetically generated faces and machine learning-based gender classification algorithms are affected by algorithmic lookism, the preferential treatment based on appearance. In experiments with 13,200 synthetically generated faces, we find that: (1) text-to-image (T2I) systems tend to associate facial attractiveness to unrelated positive traits like intelligence and trustworthiness; and (2) gender classification models exhibit higher error rates on \"less-attractive\" faces, especially among non-White women. These result raise fairness concerns regarding digital identity systems.", "date": "2025-05-20", "categories": []}
{"id": "2508.16377", "title": "Applications and Challenges of Fairness APIs in Machine Learning Software", "abstract": "Machine Learning software systems are frequently used in our day-to-day lives. Some of these systems are used in various sensitive environments to make life-changing decisions. Therefore, it is crucial to ensure that these AI/ML systems do not make any discriminatory decisions for any specific groups or populations. In that vein, different bias detection and mitigation open-source software libraries (aka API libraries) are being developed and used. In this paper, we conduct a qualitative study to understand in what scenarios these open-source fairness APIs are used in the wild, how they are used, and what challenges the developers of these APIs face while developing and adopting these libraries. We have analyzed 204 GitHub repositories (from a list of 1885 candidate repositories) which used 13 APIs that are developed to address bias in ML software. We found that these APIs are used for two primary purposes (i.e., learning and solving real-world problems), targeting 17 unique use-cases. Our study suggests that developers are not well-versed in bias detection and mitigation; they face lots of troubleshooting issues, and frequently ask for opinions and resources. Our findings can be instrumental for future bias-related software engineering research, and for guiding educators in developing more state-of-the-art curricula.", "date": "2025-08-22", "categories": []}
{"id": "2508.16386", "title": "Sequential Cohort Selection", "abstract": "We study the problem of fair cohort selection from an unknown population, with a focus on university admissions. We start with the one-shot setting, where the admission policy must be fixed in advance and remain transparent, before observing the actual applicant pool. In contrast, the sequential setting allows the policy to be updated across stages as new applicant data becomes available. This is achieved by optimizing admission policies using a population model, trained on data from previous admission cycles. We also study the fairness properties of the resulting policies in the one-shot setting, including meritocracy and group parity.", "date": "2025-08-22", "categories": []}
{"id": "2501.08429", "title": "Modeling Discrimination with Causal Abstraction", "abstract": "A person is directly racially discriminated against only if her race caused her worse treatment. This implies that race is an attribute sufficiently separable from other attributes to isolate its causal role. But race is embedded in a nexus of social factors that resist isolated treatment. If race is socially constructed, in what sense can it cause worse treatment? Some propose that the perception of race, rather than race itself, causes worse treatment. Others suggest that since causal models require \\textit{modularity}, i.e. the ability to isolate causal effects, attempts to causally model discrimination are misguided.\n  This paper addresses the problem differently. We introduce a framework for reasoning about discrimination, in which race is a high-level \\textit{abstraction} of lower-level features. In this framework, race can be modeled as itself causing worse treatment. Modularity is ensured by allowing assumptions about social construction to be precisely and explicitly stated, via an alignment between race and its constituents. Such assumptions can then be subjected to normative and empirical challenges, which lead to different views of when discrimination occurs. By distinguishing constitutive and causal relations, the abstraction framework pinpoints disagreements in the current literature on modeling discrimination, while preserving a precise causal account of discrimination.", "date": "2025-01-14", "categories": []}
{"id": "2508.14933", "title": "Inference Time Debiasing Concepts in Diffusion Models", "abstract": "We propose DeCoDi, a debiasing procedure for text-to-image diffusion-based models that changes the inference procedure, does not significantly change image quality, has negligible compute overhead, and can be applied in any diffusion-based image generation model. DeCoDi changes the diffusion process to avoid latent dimension regions of biased concepts. While most deep learning debiasing methods require complex or compute-intensive interventions, our method is designed to change only the inference procedure. Therefore, it is more accessible to a wide range of practitioners. We show the effectiveness of the method by debiasing for gender, ethnicity, and age for the concepts of nurse, firefighter, and CEO. Two distinct human evaluators manually inspect 1,200 generated images. Their evaluation results provide evidence that our method is effective in mitigating biases based on gender, ethnicity, and age. We also show that an automatic bias evaluation performed by the GPT4o is not significantly statistically distinct from a human evaluation. Our evaluation shows promising results, with reliable levels of agreement between evaluators and more coverage of protected attributes. Our method has the potential to significantly improve the diversity of images it generates by diffusion-based text-to-image generative models.", "date": "2025-08-19", "categories": []}
{"id": "2508.15084", "title": "Kernel-based Equalized Odds: A Quantification of Accuracy-Fairness Trade-off in Fair Representation Learning", "abstract": "This paper introduces a novel kernel-based formulation of the Equalized Odds (EO) criterion, denoted as $EO_k$, for fair representation learning (FRL) in supervised settings. The central goal of FRL is to mitigate discrimination regarding a sensitive attribute $S$ while preserving prediction accuracy for the target variable $Y$. Our proposed criterion enables a rigorous and interpretable quantification of three core fairness objectives: independence (prediction $\\hat{Y}$ is independent of $S$), separation (also known as equalized odds; prediction $\\hat{Y}$ is independent with $S$ conditioned on target attribute $Y$), and calibration ($Y$ is independent of $S$ conditioned on the prediction $\\hat{Y}$). Under both unbiased ($Y$ is independent of $S$) and biased ($Y$ depends on $S$) conditions, we show that $EO_k$ satisfies both independence and separation in the former, and uniquely preserves predictive accuracy while lower bounding independence and calibration in the latter, thereby offering a unified analytical characterization of the tradeoffs among these fairness criteria. We further define the empirical counterpart, $\\hat{EO}_k$, a kernel-based statistic that can be computed in quadratic time, with linear-time approximations also available. A concentration inequality for $\\hat{EO}_k$ is derived, providing performance guarantees and error bounds, which serve as practical certificates of fairness compliance. While our focus is on theoretical development, the results lay essential groundwork for principled and provably fair algorithmic design in future empirical studies.", "date": "2025-08-20", "categories": []}
{"id": "2508.15193", "title": "Revisiting Pre-processing Group Fairness: A Modular Benchmarking Framework", "abstract": "As machine learning systems become increasingly integrated into high-stakes decision-making processes, ensuring fairness in algorithmic outcomes has become a critical concern. Methods to mitigate bias typically fall into three categories: pre-processing, in-processing, and post-processing. While significant attention has been devoted to the latter two, pre-processing methods, which operate at the data level and offer advantages such as model-agnosticism and improved privacy compliance, have received comparatively less focus and lack standardised evaluation tools. In this work, we introduce FairPrep, an extensible and modular benchmarking framework designed to evaluate fairness-aware pre-processing techniques on tabular datasets. Built on the AIF360 platform, FairPrep allows seamless integration of datasets, fairness interventions, and predictive models. It features a batch-processing interface that enables efficient experimentation and automatic reporting of fairness and utility metrics. By offering standardised pipelines and supporting reproducible evaluations, FairPrep fills a critical gap in the fairness benchmarking landscape and provides a practical foundation for advancing data-level fairness research.", "date": "2025-08-21", "categories": []}
{"id": "2508.15642", "title": "Correct-By-Construction: Certified Individual Fairness through Neural Network Training", "abstract": "Fairness in machine learning is more important than ever as ethical concerns continue to grow. Individual fairness demands that individuals differing only in sensitive attributes receive the same outcomes. However, commonly used machine learning algorithms often fail to achieve such fairness. To improve individual fairness, various training methods have been developed, such as incorporating fairness constraints as optimisation objectives. While these methods have demonstrated empirical effectiveness, they lack formal guarantees of fairness. Existing approaches that aim to provide fairness guarantees primarily rely on verification techniques, which can sometimes fail to produce definitive results. Moreover, verification alone does not actively enhance individual fairness during training. To address this limitation, we propose a novel framework that formally guarantees individual fairness throughout training. Our approach consists of two parts, i.e., (1) provably fair initialisation that ensures the model starts in a fair state, and (2) a fairness-preserving training algorithm that maintains fairness as the model learns. A key element of our method is the use of randomised response mechanisms, which protect sensitive attributes while maintaining fairness guarantees. We formally prove that this mechanism sustains individual fairness throughout the training process. Experimental evaluations confirm that our approach is effective, i.e., producing models that are empirically fair and accurate. Furthermore, our approach is much more efficient than the alternative approach based on certified training (which requires neural network verification during training).", "date": "2025-08-21", "categories": []}
{"id": "2408.06569", "title": "Social Debiasing for Fair Multi-modal LLMs", "abstract": "Multi-modal Large Language Models (MLLMs) have dramatically advanced the research field and delivered powerful vision-language understanding capabilities. However, these models often inherit deep-rooted social biases from their training data, leading to uncomfortable responses with respect to attributes such as race and gender. This paper addresses the issue of social biases in MLLMs by i) introducing a comprehensive counterfactual dataset with multiple social concepts (CMSC), which complements existing datasets by providing 18 diverse and balanced social concepts; and ii) proposing a counter-stereotype debiasing (CSD) strategy that mitigates social biases in MLLMs by leveraging the opposites of prevalent stereotypes. CSD incorporates both a novel bias-aware data sampling method and a loss rescaling method, enabling the model to effectively reduce biases. We conduct extensive experiments with four prevalent MLLM architectures. The results demonstrate the advantage of the CMSC dataset and the edge of CSD strategy in reducing social biases compared to existing competing methods, without compromising the overall performance on general multi-modal reasoning benchmarks.", "date": "2024-08-13", "categories": []}
{"id": "2508.14576", "title": "A Comprehensive Evaluation of the Sensitivity of Density-Ratio Estimation Based Fairness Measurement in Regression", "abstract": "The prevalence of algorithmic bias in Machine Learning (ML)-driven approaches has inspired growing research on measuring and mitigating bias in the ML domain. Accordingly, prior research studied how to measure fairness in regression which is a complex problem. In particular, recent research proposed to formulate it as a density-ratio estimation problem and relied on a Logistic Regression-driven probabilistic classifier-based approach to solve it. However, there are several other methods to estimate a density ratio, and to the best of our knowledge, prior work did not study the sensitivity of such fairness measurement methods to the choice of underlying density ratio estimation algorithm. To fill this gap, this paper develops a set of fairness measurement methods with various density-ratio estimation cores and thoroughly investigates how different cores would affect the achieved level of fairness. Our experimental results show that the choice of density-ratio estimation core could significantly affect the outcome of fairness measurement method, and even, generate inconsistent results with respect to the relative fairness of various algorithms. These observations suggest major issues with density-ratio estimation based fairness measurement in regression and a need for further research to enhance their reliability.", "date": "2025-08-20", "categories": []}
{"id": "2508.14683", "title": "Improving Fairness in Graph Neural Networks via Counterfactual Debiasing", "abstract": "Graph Neural Networks (GNNs) have been successful in modeling graph-structured data. However, similar to other machine learning models, GNNs can exhibit bias in predictions based on attributes like race and gender. Moreover, bias in GNNs can be exacerbated by the graph structure and message-passing mechanisms. Recent cutting-edge methods propose mitigating bias by filtering out sensitive information from input or representations, like edge dropping or feature masking. Yet, we argue that such strategies may unintentionally eliminate non-sensitive features, leading to a compromised balance between predictive accuracy and fairness. To tackle this challenge, we present a novel approach utilizing counterfactual data augmentation for bias mitigation. This method involves creating diverse neighborhoods using counterfactuals before message passing, facilitating unbiased node representations learning from the augmented graph. Subsequently, an adversarial discriminator is employed to diminish bias in predictions by conventional GNN classifiers. Our proposed technique, Fair-ICD, ensures the fairness of GNNs under moderate conditions. Experiments on standard datasets using three GNN backbones demonstrate that Fair-ICD notably enhances fairness metrics while preserving high predictive performance.", "date": "2025-08-20", "categories": []}
{"id": "2508.13040", "title": "Beyond Internal Data: Bounding and Estimating Fairness from Incomplete Data", "abstract": "Ensuring fairness in AI systems is critical, especially in high-stakes domains such as lending, hiring, and healthcare. This urgency is reflected in emerging global regulations that mandate fairness assessments and independent bias audits. However, procuring the necessary complete data for fairness testing remains a significant challenge. In industry settings, legal and privacy concerns restrict the collection of demographic data required to assess group disparities, and auditors face practical and cultural challenges in gaining access to data. In practice, data relevant for fairness testing is often split across separate sources: internal datasets held by institutions with predictive attributes, and external public datasets such as census data containing protected attributes, each providing only partial, marginal information. Our work seeks to leverage such available separate data to estimate model fairness when complete data is inaccessible. We propose utilising the available separate data to estimate a set of feasible joint distributions and then compute the set plausible fairness metrics. Through simulation and real experiments, we demonstrate that we can derive meaningful bounds on fairness metrics and obtain reliable estimates of the true metric. Our results demonstrate that this approach can serve as a practical and effective solution for fairness testing in real-world settings where access to complete data is restricted.", "date": "2025-08-18", "categories": []}
{"id": "2508.11258", "title": "Group Fairness Meets the Black Box: Enabling Fair Algorithms on Closed LLMs via Post-Processing", "abstract": "Instruction fine-tuned large language models (LLMs) enable a simple zero-shot or few-shot prompting paradigm, also known as in-context learning, for building prediction models. This convenience, combined with continued advances in LLM capability, has the potential to drive their adoption across a broad range of domains, including high-stakes applications where group fairness -- preventing disparate impacts across demographic groups -- is essential. The majority of existing approaches to enforcing group fairness on LLM-based classifiers rely on traditional fair algorithms applied via model fine-tuning or head-tuning on final-layer embeddings, but they are no longer applicable to closed-weight LLMs under the in-context learning setting, which include some of the most capable commercial models today, such as GPT-4, Gemini, and Claude. In this paper, we propose a framework for deriving fair classifiers from closed-weight LLMs via prompting: the LLM is treated as a feature extractor, and features are elicited from its probabilistic predictions (e.g., token log probabilities) using prompts strategically designed for the specified fairness criterion to obtain sufficient statistics for fair classification; a fair algorithm is then applied to these features to train a lightweight fair classifier in a post-hoc manner. Experiments on five datasets, including three tabular ones, demonstrate strong accuracy-fairness tradeoffs for the classifiers derived by our framework from both open-weight and closed-weight LLMs; in particular, our framework is data-efficient and outperforms fair classifiers trained on LLM embeddings (i.e., head-tuning) or from scratch on raw tabular features.", "date": "2025-08-15", "categories": []}
{"id": "2410.17269", "title": "FairFML: Fair Federated Machine Learning with a Case Study on Reducing Gender Disparities in Cardiac Arrest Outcome Prediction", "abstract": "Objective: Mitigating algorithmic disparities is a critical challenge in healthcare research, where ensuring equity and fairness is paramount. While large-scale healthcare data exist across multiple institutions, cross-institutional collaborations often face privacy constraints, highlighting the need for privacy-preserving solutions that also promote fairness.\n  Materials and Methods: In this study, we present Fair Federated Machine Learning (FairFML), a model-agnostic solution designed to reduce algorithmic bias in cross-institutional healthcare collaborations while preserving patient privacy. As a proof of concept, we validated FairFML using a real-world clinical case study focused on reducing gender disparities in cardiac arrest outcome prediction.\n  Results: We demonstrate that the proposed FairFML framework enhances fairness in federated learning (FL) models without compromising predictive performance. Our findings show that FairFML improves model fairness by up to 65% compared to the centralized model, while maintaining performance comparable to both local and centralized models, as measured by receiver operating characteristic analysis.\n  Discussion and Conclusion: FairFML offers a promising and flexible solution for FL collaborations, with its adaptability allowing seamless integration with various FL frameworks and models, from traditional statistical methods to deep learning techniques. This makes FairFML a robust approach for developing fairer FL models across diverse clinical and biomedical applications.", "date": "2024-10-07", "categories": []}
{"id": "2508.10304", "title": "Yet another algorithmic bias: A Discursive Analysis of Large Language Models Reinforcing Dominant Discourses on Gender and Race", "abstract": "With the advance of Artificial Intelligence (AI), Large Language Models (LLMs) have gained prominence and been applied in diverse contexts. As they evolve into more sophisticated versions, it is essential to assess whether they reproduce biases, such as discrimination and racialization, while maintaining hegemonic discourses. Current bias detection approaches rely mostly on quantitative, automated methods, which often overlook the nuanced ways in which biases emerge in natural language. This study proposes a qualitative, discursive framework to complement such methods. Through manual analysis of LLM-generated short stories featuring Black and white women, we investigate gender and racial biases. We contend that qualitative methods such as the one proposed here are fundamental to help both developers and users identify the precise ways in which biases manifest in LLM outputs, thus enabling better conditions to mitigate them. Results show that Black women are portrayed as tied to ancestry and resistance, while white women appear in self-discovery processes. These patterns reflect how language models replicate crystalized discursive representations, reinforcing essentialization and a sense of social immobility. When prompted to correct biases, models offered superficial revisions that maintained problematic meanings, revealing limitations in fostering inclusive narratives. Our results demonstrate the ideological functioning of algorithms and have significant implications for the ethical use and development of AI. The study reinforces the need for critical, interdisciplinary approaches to AI design and deployment, addressing how LLM-generated discourses reflect and perpetuate inequalities.", "date": "2025-08-14", "categories": []}
{"id": "2508.10345", "title": "Welfare-Centric Clustering", "abstract": "Fair clustering has traditionally focused on ensuring equitable group representation or equalizing group-specific clustering costs. However, Dickerson et al. (2025) recently showed that these fairness notions may yield undesirable or unintuitive clustering outcomes and advocated for a welfare-centric clustering approach that models the utilities of the groups. In this work, we model group utilities based on both distances and proportional representation and formalize two optimization objectives based on welfare-centric clustering: the Rawlsian (Egalitarian) objective and the Utilitarian objective. We introduce novel algorithms for both objectives and prove theoretical guarantees for them. Empirical evaluations on multiple real-world datasets demonstrate that our methods significantly outperform existing fair clustering baselines.", "date": "2025-08-14", "categories": []}
{"id": "2508.03940", "title": "FairPOT: Balancing AUC Performance and Fairness with Proportional Optimal Transport", "abstract": "Fairness metrics utilizing the area under the receiver operator characteristic curve (AUC) have gained increasing attention in high-stakes domains such as healthcare, finance, and criminal justice. In these domains, fairness is often evaluated over risk scores rather than binary outcomes, and a common challenge is that enforcing strict fairness can significantly degrade AUC performance. To address this challenge, we propose Fair Proportional Optimal Transport (FairPOT), a novel, model-agnostic post-processing framework that strategically aligns risk score distributions across different groups using optimal transport, but does so selectively by transforming a controllable proportion, i.e., the top-lambda quantile, of scores within the disadvantaged group. By varying lambda, our method allows for a tunable trade-off between reducing AUC disparities and maintaining overall AUC performance. Furthermore, we extend FairPOT to the partial AUC setting, enabling fairness interventions to concentrate on the highest-risk regions. Extensive experiments on synthetic, public, and clinical datasets show that FairPOT consistently outperforms existing post-processing techniques in both global and partial AUC scenarios, often achieving improved fairness with slight AUC degradation or even positive gains in utility. The computational efficiency and practical adaptability of FairPOT make it a promising solution for real-world deployment.", "date": "2025-08-05", "categories": []}
{"id": "2508.07143", "title": "Fairness of Automatic Speech Recognition: Looking Through a Philosophical Lens", "abstract": "Automatic Speech Recognition (ASR) systems now mediate countless human-technology interactions, yet research on their fairness implications remains surprisingly limited. This paper examines ASR bias through a philosophical lens, arguing that systematic misrecognition of certain speech varieties constitutes more than a technical limitation -- it represents a form of disrespect that compounds historical injustices against marginalized linguistic communities. We distinguish between morally neutral classification (discriminate1) and harmful discrimination (discriminate2), demonstrating how ASR systems can inadvertently transform the former into the latter when they consistently misrecognize non-standard dialects. We identify three unique ethical dimensions of speech technologies that differentiate ASR bias from other algorithmic fairness concerns: the temporal burden placed on speakers of non-standard varieties (\"temporal taxation\"), the disruption of conversational flow when systems misrecognize speech, and the fundamental connection between speech patterns and personal/cultural identity. These factors create asymmetric power relationships that existing technical fairness metrics fail to capture. The paper analyzes the tension between linguistic standardization and pluralism in ASR development, arguing that current approaches often embed and reinforce problematic language ideologies. We conclude that addressing ASR bias requires more than technical interventions; it demands recognition of diverse speech varieties as legitimate forms of expression worthy of technological accommodation. This philosophical reframing offers new pathways for developing ASR systems that respect linguistic diversity and speaker autonomy.", "date": "2025-08-10", "categories": []}
{"id": "2405.07623", "title": "Optimizing Class-Level Probability Reweighting Coefficients for Equitable Prompting Accuracy", "abstract": "Even as we engineer LLMs for alignment and safety, they often uncover biases from pre-training data's statistical regularities (from disproportionate co-occurrences to stereotypical associations mirroring human cognitive biases). This leads to persistent, uneven class accuracy in classification and QA. Such per-class accuracy disparities are not inherently resolved by architectural/training evolutions or data scaling, making post-hoc correction essential for equitable performance. To mitigate LLM class accuracy imbalance, we develop a post-hoc probability reweighting method that directly optimizes for non-differentiable performance-driven and fairness-aligned metrics, through a novel COBias metric that highlights disparities in class accuracies. This post-hoc bias mitigation method is grounded in discrete optimization with nonlinear integer programming (NIP) objectives and an efficient metaheuristic solution framework with theoretical convergence guarantees. Operating model-agnostically, it learns reweighting coefficients from output class probabilities to adjust LLM inference outputs without internal weight updates. Evaluations demonstrate its effectiveness: reducing COBias (61% relative reduction), increasing overall accuracy (18% relative increase), and achieving robust within-task generalization across diverse prompt configurations.", "date": "2024-05-13", "categories": []}
{"id": "2508.06671", "title": "Do Biased Models Have Biased Thoughts?", "abstract": "The impressive performance of language models is undeniable. However, the presence of biases based on gender, race, socio-economic status, physical appearance, and sexual orientation makes the deployment of language models challenging. This paper studies the effect of chain-of-thought prompting, a recent approach that studies the steps followed by the model before it responds, on fairness. More specifically, we ask the following question: $\\textit{Do biased models have biased thoughts}$? To answer our question, we conduct experiments on $5$ popular large language models using fairness metrics to quantify $11$ different biases in the model's thoughts and output. Our results show that the bias in the thinking steps is not highly correlated with the output bias (less than $0.6$ correlation with a $p$-value smaller than $0.001$ in most cases). In other words, unlike human beings, the tested models with biased decisions do not always possess biased thoughts.", "date": "2025-08-08", "categories": []}
{"id": "2508.09087", "title": "Addressing Bias in VLMs for Glaucoma Detection Without Protected Attribute Supervision", "abstract": "Vision-Language Models (VLMs) have achieved remarkable success on multimodal tasks such as image-text retrieval and zero-shot classification, yet they can exhibit demographic biases even when explicit protected attributes are absent during training. In this work, we focus on automated glaucoma screening from retinal fundus images, a critical application given that glaucoma is a leading cause of irreversible blindness and disproportionately affects underserved populations. Building on a reweighting-based contrastive learning framework, we introduce an attribute-agnostic debiasing method that (i) infers proxy subgroups via unsupervised clustering of image-image embeddings, (ii) computes gradient-similarity weights between the CLIP-style multimodal loss and a SimCLR-style image-pair contrastive loss, and (iii) applies these weights in a joint, top-$k$ weighted objective to upweight underperforming clusters. This label-free approach adaptively targets the hardest examples, thereby reducing subgroup disparities. We evaluate our method on the Harvard FairVLMed glaucoma subset, reporting Equalized Odds Distance (EOD), Equalized Subgroup AUC (ES AUC), and Groupwise AUC to demonstrate equitable performance across inferred demographic subgroups.", "date": "2025-08-12", "categories": []}
{"id": "2404.16663", "title": "Runtime Monitoring and Enforcement of Conditional Fairness in Generative AIs", "abstract": "The deployment of generative AI (GenAI) models raises significant fairness concerns, addressed in this paper through novel characterization and enforcement techniques specific to GenAI. Unlike standard AI performing specific tasks, GenAI's broad functionality requires ``conditional fairness'' tailored to the context being generated, such as demographic fairness in generating images of poor people versus successful business leaders. We define two fairness levels: the first evaluates fairness in generated outputs, independent of prompts and models; the second assesses inherent fairness with neutral prompts. Given the complexity of GenAI and challenges in fairness specifications, we focus on bounding the worst case, considering a GenAI system unfair if the distance between appearances of a specific group exceeds preset thresholds. We also explore combinatorial testing for assessing relative completeness in intersectional fairness. By bounding the worst case, we develop a prompt injection scheme within an agent-based framework to enforce conditional fairness with minimal intervention, validated on state-of-the-art GenAI systems.", "date": "2024-04-25", "categories": []}
{"id": "2502.01926", "title": "Fairness through Difference Awareness: Measuring Desired Group Discrimination in LLMs", "abstract": "Algorithmic fairness has conventionally adopted the mathematically convenient perspective of racial color-blindness (i.e., difference unaware treatment). However, we contend that in a range of important settings, group difference awareness matters. For example, differentiating between groups may be necessary in legal contexts (e.g., the U.S. compulsory draft applies to men but not women) and harm assessments (e.g., referring to girls as ``terrorists'' may be less harmful than referring to Muslim people as such). Thus, in contrast to most fairness work, we study fairness through the perspective of treating people differently -- when it is contextually appropriate to. We first introduce an important distinction between descriptive (fact-based), normative (value-based), and correlation (association-based) benchmarks. This distinction is significant because each category requires separate interpretation and mitigation tailored to its specific characteristics. Then, we present a benchmark suite composed of eight different scenarios for a total of 16k questions that enables us to assess difference awareness. Finally, we show results across ten models that demonstrate difference awareness is a distinct dimension to fairness where existing bias mitigation strategies may backfire.", "date": "2025-02-04", "categories": []}
{"id": "2503.13236", "title": "Gradient Extrapolation for Debiased Representation Learning", "abstract": "Machine learning classification models trained with empirical risk minimization (ERM) often inadvertently rely on spurious correlations. When absent in the test data, these unintended associations between non-target attributes and target labels lead to poor generalization. This paper addresses this problem from a model optimization perspective and proposes a novel method, Gradient Extrapolation for Debiased Representation Learning (GERNE), designed to learn debiased representations in both known and unknown attribute training cases. GERNE uses two distinct batches with different amounts of spurious correlations and defines the target gradient as a linear extrapolation of the gradients computed from each batch's loss. Our analysis shows that when the extrapolated gradient points toward the batch gradient with fewer spurious correlations, it effectively guides training toward learning a debiased model. GERNE serves as a general framework for debiasing, encompassing ERM and Resampling methods as special cases. We derive the theoretical upper and lower bounds of the extrapolation factor employed by GERNE. By tuning this factor, GERNE can adapt to maximize either Group-Balanced Accuracy (GBA) or Worst-Group Accuracy (WGA). We validate GERNE on five vision and one NLP benchmarks, demonstrating competitive and often superior performance compared to state-of-the-art baselines. The project page is available at: https://gerne-debias.github.io/.", "date": "2025-03-17", "categories": []}
{"id": "2508.07111", "title": "Investigating Intersectional Bias in Large Language Models using Confidence Disparities in Coreference Resolution", "abstract": "Large language models (LLMs) have achieved impressive performance, leading to their widespread adoption as decision-support tools in resource-constrained contexts like hiring and admissions. There is, however, scientific consensus that AI systems can reflect and exacerbate societal biases, raising concerns about identity-based harm when used in critical social contexts. Prior work has laid a solid foundation for assessing bias in LLMs by evaluating demographic disparities in different language reasoning tasks. In this work, we extend single-axis fairness evaluations to examine intersectional bias, recognizing that when multiple axes of discrimination intersect, they create distinct patterns of disadvantage. We create a new benchmark called WinoIdentity by augmenting the WinoBias dataset with 25 demographic markers across 10 attributes, including age, nationality, and race, intersected with binary gender, yielding 245,700 prompts to evaluate 50 distinct bias patterns. Focusing on harms of omission due to underrepresentation, we investigate bias through the lens of uncertainty and propose a group (un)fairness metric called Coreference Confidence Disparity which measures whether models are more or less confident for some intersectional identities than others. We evaluate five recently published LLMs and find confidence disparities as high as 40% along various demographic attributes including body type, sexual orientation and socio-economic status, with models being most uncertain about doubly-disadvantaged identities in anti-stereotypical settings. Surprisingly, coreference confidence decreases even for hegemonic or privileged markers, indicating that the recent impressive performance of LLMs is more likely due to memorization than logical reasoning. Notably, these are two independent failures in value alignment and validity that can compound to cause social harm.", "date": "2025-08-09", "categories": []}
{"id": "2508.07432", "title": "Freeze and Reveal: Exposing Modality Bias in Vision-Language Models", "abstract": "Vision Language Models achieve impressive multi-modal performance but often inherit gender biases from their training data. This bias might be coming from both the vision and text modalities. In this work, we dissect the contributions of vision and text backbones to these biases by applying targeted debiasing using Counterfactual Data Augmentation and Task Vector methods. Inspired by data-efficient approaches in hate-speech classification, we introduce a novel metric, Degree of Stereotypicality and a corresponding debiasing method, Data Augmentation Using Degree of Stereotypicality - DAUDoS, to reduce bias with minimal computational cost. We curate a gender annotated dataset and evaluate all methods on VisoGender benchmark to quantify improvements and identify dominant source of bias. Our results show that CDA reduces the gender gap by 6% and DAUDoS by 3% but using only one-third of the data. Both methods also improve the model's ability to correctly identify gender in images by 3%, with DAUDoS achieving this improvement using only almost one-third of training data. From our experiment's, we observed that CLIP's vision encoder is more biased whereas PaliGemma2's text encoder is more biased. By identifying whether bias stems more from vision or text encoders, our work enables more targeted and effective bias mitigation strategies in future multi-modal systems.", "date": "2025-08-10", "categories": []}
{"id": "2508.07516", "title": "Augmenting Bias Detection in LLMs Using Topological Data Analysis", "abstract": "Recently, many bias detection methods have been proposed to determine the level of bias a large language model captures. However, tests to identify which parts of a large language model are responsible for bias towards specific groups remain underdeveloped. In this study, we present a method using topological data analysis to identify which heads in GPT-2 contribute to the misrepresentation of identity groups present in the StereoSet dataset. We find that biases for particular categories, such as gender or profession, are concentrated in attention heads that act as hot spots. The metric we propose can also be used to determine which heads capture bias for a specific group within a bias category, and future work could extend this method to help de-bias large language models.", "date": "2025-08-11", "categories": []}
{"id": "2508.07518", "title": "FairDRL-ST: Disentangled Representation Learning for Fair Spatio-Temporal Mobility Prediction", "abstract": "As deep spatio-temporal neural networks are increasingly utilised in urban computing contexts, the deployment of such methods can have a direct impact on users of critical urban infrastructure, such as public transport, emergency services, and traffic management systems. While many spatio-temporal methods focus on improving accuracy, fairness has recently gained attention due to growing evidence that biased predictions in spatio-temporal applications can disproportionately disadvantage certain demographic or geographic groups, thereby reinforcing existing socioeconomic inequalities and undermining the ethical deployment of AI in public services. In this paper, we propose a novel framework, FairDRL-ST, based on disentangled representation learning, to address fairness concerns in spatio-temporal prediction, with a particular focus on mobility demand forecasting. By leveraging adversarial learning and disentangled representation learning, our framework learns to separate attributes that contain sensitive information. Unlike existing methods that enforce fairness through supervised learning, which may lead to overcompensation and degraded performance, our framework achieves fairness in an unsupervised manner with minimal performance loss. We apply our framework to real-world urban mobility datasets and demonstrate its ability to close fairness gaps while delivering competitive predictive performance compared to state-of-the-art fairness-aware methods.", "date": "2025-08-11", "categories": []}
{"id": "2508.07872", "title": "Unequal Uncertainty: Rethinking Algorithmic Interventions for Mitigating Discrimination from AI", "abstract": "Uncertainty in artificial intelligence (AI) predictions poses urgent legal and ethical challenges for AI-assisted decision-making. We examine two algorithmic interventions that act as guardrails for human-AI collaboration: selective abstention, which withholds high-uncertainty predictions from human decision-makers, and selective friction, which delivers those predictions together with salient warnings or disclosures that slow the decision process. Research has shown that selective abstention based on uncertainty can inadvertently exacerbate disparities and disadvantage under-represented groups that disproportionately receive uncertain predictions. In this paper, we provide the first integrated socio-technical and legal analysis of uncertainty-based algorithmic interventions. Through two case studies, AI-assisted consumer credit decisions and AI-assisted content moderation, we demonstrate how the seemingly neutral use of uncertainty thresholds can trigger discriminatory impacts. We argue that, although both interventions pose risks of unlawful discrimination under UK law, selective frictions offer a promising pathway toward fairer and more accountable AI-assisted decision-making by preserving transparency and encouraging more cautious human judgment.", "date": "2025-08-11", "categories": []}
{"id": "2508.08151", "title": "FairFLRep: Fairness aware fault localization and repair of Deep Neural Networks", "abstract": "Deep neural networks (DNNs) are being utilized in various aspects of our daily lives, including high-stakes decision-making applications that impact individuals. However, these systems reflect and amplify bias from the data used during training and testing, potentially resulting in biased behavior and inaccurate decisions. For instance, having different misclassification rates between white and black sub-populations. However, effectively and efficiently identifying and correcting biased behavior in DNNs is a challenge. This paper introduces FairFLRep, an automated fairness-aware fault localization and repair technique that identifies and corrects potentially bias-inducing neurons in DNN classifiers. FairFLRep focuses on adjusting neuron weights associated with sensitive attributes, such as race or gender, that contribute to unfair decisions. By analyzing the input-output relationships within the network, FairFLRep corrects neurons responsible for disparities in predictive quality parity. We evaluate FairFLRep on four image classification datasets using two DNN classifiers, and four tabular datasets with a DNN model. The results show that FairFLRep consistently outperforms existing methods in improving fairness while preserving accuracy. An ablation study confirms the importance of considering fairness during both fault localization and repair stages. Our findings also show that FairFLRep is more efficient than the baseline approaches in repairing the network.", "date": "2025-08-11", "categories": []}
{"id": "2508.05680", "title": "Are All Genders Equal in the Eyes of Algorithms? -- Analysing Search and Retrieval Algorithms for Algorithmic Gender Fairness", "abstract": "Algorithmic systems such as search engines and information retrieval platforms significantly influence academic visibility and the dissemination of knowledge. Despite assumptions of neutrality, these systems can reproduce or reinforce societal biases, including those related to gender. This paper introduces and applies a bias-preserving definition of algorithmic gender fairness, which assesses whether algorithmic outputs reflect real-world gender distributions without introducing or amplifying disparities. Using a heterogeneous dataset of academic profiles from German universities and universities of applied sciences, we analyse gender differences in metadata completeness, publication retrieval in academic databases, and visibility in Google search results. While we observe no overt algorithmic discrimination, our findings reveal subtle but consistent imbalances: male professors are associated with a greater number of search results and more aligned publication records, while female professors display higher variability in digital visibility. These patterns reflect the interplay between platform algorithms, institutional curation, and individual self-presentation. Our study highlights the need for fairness evaluations that account for both technical performance and representational equality in digital systems.", "date": "2025-08-05", "categories": []}
{"id": "2508.06443", "title": "The Fair Game: Auditing & Debiasing AI Algorithms Over Time", "abstract": "An emerging field of AI, namely Fair Machine Learning (ML), aims to quantify different types of bias (also known as unfairness) exhibited in the predictions of ML algorithms, and to design new algorithms to mitigate them. Often, the definitions of bias used in the literature are observational, i.e. they use the input and output of a pre-trained algorithm to quantify a bias under concern. In reality,these definitions are often conflicting in nature and can only be deployed if either the ground truth is known or only in retrospect after deploying the algorithm. Thus,there is a gap between what we want Fair ML to achieve and what it does in a dynamic social environment. Hence, we propose an alternative dynamic mechanism,\"Fair Game\",to assure fairness in the predictions of an ML algorithm and to adapt its predictions as the society interacts with the algorithm over time. \"Fair Game\" puts together an Auditor and a Debiasing algorithm in a loop around an ML algorithm. The \"Fair Game\" puts these two components in a loop by leveraging Reinforcement Learning (RL). RL algorithms interact with an environment to take decisions, which yields new observations (also known as data/feedback) from the environment and in turn, adapts future decisions. RL is already used in algorithms with pre-fixed long-term fairness goals. \"Fair Game\" provides a unique framework where the fairness goals can be adapted over time by only modifying the auditor and the different biases it quantifies. Thus,\"Fair Game\" aims to simulate the evolution of ethical and legal frameworks in the society by creating an auditor which sends feedback to a debiasing algorithm deployed around an ML system. This allows us to develop a flexible and adaptive-over-time framework to build Fair ML systems pre- and post-deployment.", "date": "2025-08-08", "categories": []}
{"id": "2410.17519", "title": "Large Language Models Still Exhibit Bias in Long Text", "abstract": "Existing fairness benchmarks for large language models (LLMs) primarily focus on simple tasks, such as multiple-choice questions, overlooking biases that may arise in more complex scenarios like long-text generation. To address this gap, we introduce the Long Text Fairness Test (LTF-TEST), a framework that evaluates biases in LLMs through essay-style prompts. LTF-TEST covers 14 topics and 10 demographic axes, including gender and race, resulting in 11,948 samples. By assessing both model responses and the reasoning behind them, LTF-TEST uncovers subtle biases that are difficult to detect in simple responses. In our evaluation of five recent LLMs, including GPT-4o and LLaMa3, we identify two key patterns of bias. First, these models frequently favor certain demographic groups in their responses. Second, they show excessive sensitivity toward traditionally disadvantaged groups, often providing overly protective responses while neglecting others. To mitigate these biases, we propose FT-REGARD, a finetuning approach that pairs biased prompts with neutral responses. FT-REGARD reduces gender bias by 34.6% and improves performance by 1.4 percentage points on the BBQ benchmark, offering a promising approach to addressing biases in long-text generation tasks.", "date": "2024-10-23", "categories": []}
{"id": "2508.04939", "title": "I Think, Therefore I Am Under-Qualified? A Benchmark for Evaluating Linguistic Shibboleth Detection in LLM Hiring Evaluations", "abstract": "This paper introduces a comprehensive benchmark for evaluating how Large Language Models (LLMs) respond to linguistic shibboleths: subtle linguistic markers that can inadvertently reveal demographic attributes such as gender, social class, or regional background. Through carefully constructed interview simulations using 100 validated question-response pairs, we demonstrate how LLMs systematically penalize certain linguistic patterns, particularly hedging language, despite equivalent content quality. Our benchmark generates controlled linguistic variations that isolate specific phenomena while maintaining semantic equivalence, which enables the precise measurement of demographic bias in automated evaluation systems. We validate our approach along multiple linguistic dimensions, showing that hedged responses receive 25.6% lower ratings on average, and demonstrate the benchmark's effectiveness in identifying model-specific biases. This work establishes a foundational framework for detecting and measuring linguistic discrimination in AI systems, with broad applications to fairness in automated decision-making contexts.", "date": "2025-08-06", "categories": []}
{"id": "2508.05435", "title": "Competing Risks: Impact on Risk Estimation and Algorithmic Fairness", "abstract": "Accurate time-to-event prediction is integral to decision-making, informing medical guidelines, hiring decisions, and resource allocation. Survival analysis, the quantitative framework used to model time-to-event data, accounts for patients who do not experience the event of interest during the study period, known as censored patients. However, many patients experience events that prevent the observation of the outcome of interest. These competing risks are often treated as censoring, a practice frequently overlooked due to a limited understanding of its consequences. Our work theoretically demonstrates why treating competing risks as censoring introduces substantial bias in survival estimates, leading to systematic overestimation of risk and, critically, amplifying disparities. First, we formalize the problem of misclassifying competing risks as censoring and quantify the resulting error in survival estimates. Specifically, we develop a framework to estimate this error and demonstrate the associated implications for predictive performance and algorithmic fairness. Furthermore, we examine how differing risk profiles across demographic groups lead to group-specific errors, potentially exacerbating existing disparities. Our findings, supported by an empirical analysis of cardiovascular management, demonstrate that ignoring competing risks disproportionately impacts the individuals most at risk of these events, potentially accentuating inequity. By quantifying the error and highlighting the fairness implications of the common practice of considering competing risks as censoring, our work provides a critical insight into the development of survival models: practitioners must account for competing risks to improve accuracy, reduce disparities in risk assessment, and better inform downstream decisions.", "date": "2025-08-07", "categories": []}
{"id": "2508.03970", "title": "Data and AI governance: Promoting equity, ethics, and fairness in large language models", "abstract": "In this paper, we cover approaches to systematically govern, assess and quantify bias across the complete life cycle of machine learning models, from initial development and validation to ongoing production monitoring and guardrail implementation. Building upon our foundational work on the Bias Evaluation and Assessment Test Suite (BEATS) for Large Language Models, the authors share prevalent bias and fairness related gaps in Large Language Models (LLMs) and discuss data and AI governance framework to address Bias, Ethics, Fairness, and Factuality within LLMs. The data and AI governance approach discussed in this paper is suitable for practical, real-world applications, enabling rigorous benchmarking of LLMs prior to production deployment, facilitating continuous real-time evaluation, and proactively governing LLM generated responses. By implementing the data and AI governance across the life cycle of AI development, organizations can significantly enhance the safety and responsibility of their GenAI systems, effectively mitigating risks of discrimination and protecting against potential reputational or brand-related harm. Ultimately, through this article, we aim to contribute to advancement of the creation and deployment of socially responsible and ethically aligned generative artificial intelligence powered applications.", "date": "2025-08-05", "categories": []}
{"id": "2508.04071", "title": "Adversarial Fair Multi-View Clustering", "abstract": "Cluster analysis is a fundamental problem in data mining and machine learning. In recent years, multi-view clustering has attracted increasing attention due to its ability to integrate complementary information from multiple views. However, existing methods primarily focus on clustering performance, while fairness-a critical concern in human-centered applications-has been largely overlooked. Although recent studies have explored group fairness in multi-view clustering, most methods impose explicit regularization on cluster assignments, relying on the alignment between sensitive attributes and the underlying cluster structure. However, this assumption often fails in practice and can degrade clustering performance. In this paper, we propose an adversarial fair multi-view clustering (AFMVC) framework that integrates fairness learning into the representation learning process. Specifically, our method employs adversarial training to fundamentally remove sensitive attribute information from learned features, ensuring that the resulting cluster assignments are unaffected by it. Furthermore, we theoretically prove that aligning view-specific clustering assignments with a fairness-invariant consensus distribution via KL divergence preserves clustering consistency without significantly compromising fairness, thereby providing additional theoretical guarantees for our framework. Extensive experiments on data sets with fairness constraints demonstrate that AFMVC achieves superior fairness and competitive clustering performance compared to existing multi-view clustering and fairness-aware clustering methods.", "date": "2025-08-06", "categories": []}
{"id": "2402.03365", "title": "Heterophily-Aware Fair Recommendation using Graph Convolutional Networks", "abstract": "In recent years, graph neural networks (GNNs) have become a popular tool to improve the accuracy and performance of recommender systems. Modern recommender systems are not only designed to serve end users, but also to benefit other participants, such as items and item providers. These participants may have different or conflicting goals and interests, which raises the need for fairness and popularity bias considerations. GNN-based recommendation methods also face the challenges of unfairness and popularity bias, and their normalization and aggregation processes suffer from these challenges. In this paper, we propose a fair GNN-based recommender system, called HetroFair, to improve item-side fairness. HetroFair uses two separate components to generate fairness-aware embeddings: i) Fairness-aware attention, which incorporates the dot product in the normalization process of GNNs to decrease the effect of nodes' degrees. ii) Heterophily feature weighting, to assign distinct weights to different features during the aggregation process. To evaluate the effectiveness of HetroFair, we conduct extensive experiments over six real-world datasets. Our experimental results reveal that HetroFair not only alleviates unfairness and popularity bias on the item side but also achieves superior accuracy on the user side. Our implementation is publicly available at https://github.com/NematGH/HetroFair.", "date": "2024-01-31", "categories": []}
{"id": "2508.02740", "title": "Who Gets Cited? Gender- and Majority-Bias in LLM-Driven Reference Selection", "abstract": "Large language models (LLMs) are rapidly being adopted as research assistants, particularly for literature review and reference recommendation, yet little is known about whether they introduce demographic bias into citation workflows. This study systematically investigates gender bias in LLM-driven reference selection using controlled experiments with pseudonymous author names. We evaluate several LLMs (GPT-4o, GPT-4o-mini, Claude Sonnet, and Claude Haiku) by varying gender composition within candidate reference pools and analyzing selection patterns across fields. Our results reveal two forms of bias: a persistent preference for male-authored references and a majority-group bias that favors whichever gender is more prevalent in the candidate pool. These biases are amplified in larger candidate pools and only modestly attenuated by prompt-based mitigation strategies. Field-level analysis indicates that bias magnitude varies across scientific domains, with social sciences showing the least bias. Our findings indicate that LLMs can reinforce or exacerbate existing gender imbalances in scholarly recognition. Effective mitigation strategies are needed to avoid perpetuating existing gender disparities in scientific citation practices before integrating LLMs into high-stakes academic workflows.", "date": "2025-08-02", "categories": []}
{"id": "2508.03323", "title": "Software Fairness Dilemma: Is Bias Mitigation a Zero-Sum Game?", "abstract": "Fairness is a critical requirement for Machine Learning (ML) software, driving the development of numerous bias mitigation methods. Previous research has identified a leveling-down effect in bias mitigation for computer vision and natural language processing tasks, where fairness is achieved by lowering performance for all groups without benefiting the unprivileged group. However, it remains unclear whether this effect applies to bias mitigation for tabular data tasks, a key area in fairness research with significant real-world applications. This study evaluates eight bias mitigation methods for tabular data, including both widely used and cutting-edge approaches, across 44 tasks using five real-world datasets and four common ML models. Contrary to earlier findings, our results show that these methods operate in a zero-sum fashion, where improvements for unprivileged groups are related to reduced benefits for traditionally privileged groups. However, previous research indicates that the perception of a zero-sum trade-off might complicate the broader adoption of fairness policies. To explore alternatives, we investigate an approach that applies the state-of-the-art bias mitigation method solely to unprivileged groups, showing potential to enhance benefits of unprivileged groups without negatively affecting privileged groups or overall ML performance. Our study highlights potential pathways for achieving fairness improvements without zero-sum trade-offs, which could help advance the adoption of bias mitigation methods.", "date": "2025-08-05", "categories": []}
{"id": "2508.03677", "title": "FairLangProc: A Python package for fairness in NLP", "abstract": "The rise in usage of Large Language Models to near ubiquitousness in recent years has risen societal concern about their applications in decision-making contexts, such as organizational justice or healthcare. This, in turn, poses questions about the fairness of these models in critical settings, which leads to the developement of different procedures to address bias in Natural Language Processing. Although many datasets, metrics and algorithms have been proposed to measure and mitigate harmful prejudice in Natural Language Processing, their implementation is diverse and far from centralized. As a response, this paper presents FairLangProc, a comprehensive Python package providing a common implementation of some of the more recent advances in fairness in Natural Language Processing providing an interface compatible with the famous Hugging Face transformers library, aiming to encourage the widespread use and democratization of bias mitigation techniques. The implementation can be found on https://github.com/arturo-perez-peralta/FairLangProc.", "date": "2025-08-05", "categories": []}
{"id": "2409.15088", "title": "AdapFair: Ensuring Adaptive Fairness for Machine Learning Operations", "abstract": "The biases and discrimination of machine learning algorithms have attracted significant attention, leading to the development of various algorithms tailored to specific contexts. However, these solutions often fall short of addressing fairness issues inherent in machine learning operations. In this paper, we present an adaptive debiasing framework designed to find an optimal fair transformation of input data that maximally preserves data predictability under dynamic conditions. A distinctive feature of our approach is its flexibility and efficiency. It can be integrated with pretrained black-box classifiers, providing fairness guarantees with minimal retraining efforts, even in the face of frequent data drifts, evolving fairness requirements, and batches of similar tasks. To achieve this, we leverage the normalizing flows to enable efficient, information-preserving data transformation, ensuring that no critical information is lost during the debiasing process. Additionally, we incorporate the Wasserstein distance as the fairness measure to guide the optimization of data transformations. Finally, we introduce an efficient optimization algorithm with closed-formed gradient computations, making our framework scalable and suitable for dynamic, real-world environments.", "date": "2024-09-23", "categories": []}
{"id": "2508.00873", "title": "FairFedMed: Benchmarking Group Fairness in Federated Medical Imaging with FairLoRA", "abstract": "Fairness remains a critical concern in healthcare, where unequal access to services and treatment outcomes can adversely affect patient health. While Federated Learning (FL) presents a collaborative and privacy-preserving approach to model training, ensuring fairness is challenging due to heterogeneous data across institutions, and current research primarily addresses non-medical applications. To fill this gap, we establish the first experimental benchmark for fairness in medical FL, evaluating six representative FL methods across diverse demographic attributes and imaging modalities. We introduce FairFedMed, the first medical FL dataset specifically designed to study group fairness (i.e., demographics). It comprises two parts: FairFedMed-Oph, featuring 2D fundus and 3D OCT ophthalmology samples with six demographic attributes; and FairFedMed-Chest, which simulates real cross-institutional FL using subsets of CheXpert and MIMIC-CXR. Together, they support both simulated and real-world FL across diverse medical modalities and demographic groups. Existing FL models often underperform on medical images and overlook fairness across demographic groups. To address this, we propose FairLoRA, a fairness-aware FL framework based on SVD-based low-rank approximation. It customizes singular value matrices per demographic group while sharing singular vectors, ensuring both fairness and efficiency. Experimental results on the FairFedMed dataset demonstrate that FairLoRA not only achieves state-of-the-art performance in medical image classification but also significantly improves fairness across diverse populations. Our code and dataset can be accessible via link: https://wang.hms.harvard.edu/fairfedmed/.", "date": "2025-07-21", "categories": []}
{"id": "2505.00830", "title": "Intersectional Divergence: Measuring Fairness in Regression", "abstract": "Fairness in machine learning research is commonly framed in the context of classification tasks, leaving critical gaps in regression. In this paper, we propose a novel approach to measure intersectional fairness in regression tasks, going beyond the focus on single protected attributes from existing work to consider combinations of all protected attributes. Furthermore, we contend that it is insufficient to measure the average error of groups without regard for imbalanced domain preferences. Accordingly, we propose Intersectional Divergence (ID) as the first fairness measure for regression tasks that 1) describes fair model behavior across multiple protected attributes and 2) differentiates the impact of predictions in target ranges most relevant to users. We extend our proposal demonstrating how ID can be adapted into a loss function, IDLoss, that satisfies convergence guarantees and has piecewise smooth properties that enable practical optimization. Through an extensive experimental evaluation, we demonstrate how ID allows unique insights into model behavior and fairness, and how incorporating IDLoss into optimization can considerably improve single-attribute and intersectional model fairness while maintaining a competitive balance in predictive performance.", "date": "2025-05-01", "categories": []}
{"id": "2507.23349", "title": "Optimal Transport Learning: Balancing Value Optimization and Fairness in Individualized Treatment Rules", "abstract": "Individualized treatment rules (ITRs) have gained significant attention due to their wide-ranging applications in fields such as precision medicine, ridesharing, and advertising recommendations. However, when ITRs are influenced by sensitive attributes such as race, gender, or age, they can lead to outcomes where certain groups are unfairly advantaged or disadvantaged. To address this gap, we propose a flexible approach based on the optimal transport theory, which is capable of transforming any optimal ITR into a fair ITR that ensures demographic parity. Recognizing the potential loss of value under fairness constraints, we introduce an ``improved trade-off ITR,\" designed to balance value optimization and fairness while accommodating varying levels of fairness through parameter adjustment. To maximize the value of the improved trade-off ITR under specific fairness levels, we propose a smoothed fairness constraint for estimating the adjustable parameter. Additionally, we establish a theoretical upper bound on the value loss for the improved trade-off ITR. We demonstrate performance of the proposed method through extensive simulation studies and application to the Next 36 entrepreneurial program dataset.", "date": "2025-07-31", "categories": []}
{"id": "2507.14326", "title": "Rethinking Individual Fairness in Deepfake Detection", "abstract": "Generative AI models have substantially improved the realism of synthetic media, yet their misuse through sophisticated DeepFakes poses significant risks. Despite recent advances in deepfake detection, fairness remains inadequately addressed, enabling deepfake markers to exploit biases against specific populations. While previous studies have emphasized group-level fairness, individual fairness (i.e., ensuring similar predictions for similar individuals) remains largely unexplored. In this work, we identify for the first time that the original principle of individual fairness fundamentally fails in the context of deepfake detection, revealing a critical gap previously unexplored in the literature. To mitigate it, we propose the first generalizable framework that can be integrated into existing deepfake detectors to enhance individual fairness and generalization. Extensive experiments conducted on leading deepfake datasets demonstrate that our approach significantly improves individual fairness while maintaining robust detection performance, outperforming state-of-the-art methods. The code is available at https://github.com/Purdue-M2/Individual-Fairness-Deepfake-Detection.", "date": "2025-07-18", "categories": []}
{"id": "2406.06736", "title": "Long-Term Fairness Inquiries and Pursuits in Machine Learning: A Survey of Notions, Methods, and Challenges", "abstract": "The widespread integration of Machine Learning systems in daily life, particularly in high-stakes domains, has raised concerns about the fairness implications. While prior works have investigated static fairness measures, recent studies reveal that automated decision-making has long-term implications and that off-the-shelf fairness approaches may not serve the purpose of achieving long-term fairness. Additionally, the existence of feedback loops and the interaction between models and the environment introduces additional complexities that may deviate from the initial fairness goals. In this survey, we review existing literature on long-term fairness from different perspectives and present a taxonomy for long-term fairness studies. We highlight key challenges and consider future research directions, analyzing both current issues and potential further explorations.", "date": "2024-06-10", "categories": []}
{"id": "2505.02456", "title": "Colombian Waitresses y Jueces canadienses: Gender and Country Biases in Occupation Recommendations from LLMs", "abstract": "One of the goals of fairness research in NLP is to measure and mitigate stereotypical biases that are propagated by NLP systems. However, such work tends to focus on single axes of bias (most often gender) and the English language. Addressing these limitations, we contribute the first study of multilingual intersecting country and gender biases, with a focus on occupation recommendations generated by large language models. We construct a benchmark of prompts in English, Spanish and German, where we systematically vary country and gender, using 25 countries and four pronoun sets. Then, we evaluate a suite of 5 Llama-based models on this benchmark, finding that LLMs encode significant gender and country biases. Notably, we find that even when models show parity for gender or country individually, intersectional occupational biases based on both country and gender persist. We also show that the prompting language significantly affects bias, and instruction-tuned models consistently demonstrate the lowest and most stable levels of bias. Our findings highlight the need for fairness researchers to use intersectional and multilingual lenses in their work.", "date": "2025-05-05", "categories": []}
{"id": "2507.02087", "title": "Evaluating the Promise and Pitfalls of LLMs in Hiring Decisions", "abstract": "The use of large language models (LLMs) in hiring promises to streamline candidate screening, but it also raises serious concerns regarding accuracy and algorithmic bias where sufficient safeguards are not in place. In this work, we benchmark several state-of-the-art foundational LLMs - including models from OpenAI, Anthropic, Google, Meta, and Deepseek, and compare them with our proprietary domain-specific hiring model (Match Score) for job candidate matching. We evaluate each model's predictive accuracy (ROC AUC, Precision-Recall AUC, F1-score) and fairness (impact ratio of cut-off analysis across declared gender, race, and intersectional subgroups). Our experiments on a dataset of roughly 10,000 real-world recent candidate-job pairs show that Match Score outperforms the general-purpose LLMs on accuracy (ROC AUC 0.85 vs 0.77) and achieves significantly more equitable outcomes across demographic groups. Notably, Match Score attains a minimum race-wise impact ratio of 0.957 (near-parity), versus 0.809 or lower for the best LLMs, (0.906 vs 0.773 for the intersectionals, respectively). We discuss why pretraining biases may cause LLMs with insufficient safeguards to propagate societal biases in hiring scenarios, whereas a bespoke supervised model can more effectively mitigate these biases. Our findings highlight the importance of domain-specific modeling and bias auditing when deploying AI in high-stakes domains such as hiring, and caution against relying on off-the-shelf LLMs for such tasks without extensive fairness safeguards. Furthermore, we show with empirical evidence that there shouldn't be a dichotomy between choosing accuracy and fairness in hiring: a well-designed algorithm can achieve both accuracy in hiring and fairness in outcomes.", "date": "2025-07-02", "categories": []}
{"id": "2507.19962", "title": "KLAAD: Refining Attention Mechanisms to Reduce Societal Bias in Generative Language Models", "abstract": "Large language models (LLMs) often exhibit societal biases in their outputs, prompting ethical concerns regarding fairness and harm. In this work, we propose KLAAD (KL-Attention Alignment Debiasing), an attention-based debiasing framework that implicitly aligns attention distributions between stereotypical and anti-stereotypical sentence pairs without directly modifying model weights. KLAAD introduces a composite training objective combining Cross-Entropy, KL divergence, and Triplet losses, guiding the model to consistently attend across biased and unbiased contexts while preserving fluency and coherence. Experimental evaluation of KLAAD demonstrates improved bias mitigation on both the BBQ and BOLD benchmarks, with minimal impact on language modeling quality. The results indicate that attention-level alignment offers a principled solution for mitigating bias in generative language models.", "date": "2025-07-26", "categories": []}
{"id": "2507.20284", "title": "Controllable Feature Whitening for Hyperparameter-Free Bias Mitigation", "abstract": "As the use of artificial intelligence rapidly increases, the development of trustworthy artificial intelligence has become important. However, recent studies have shown that deep neural networks are susceptible to learn spurious correlations present in datasets. To improve the reliability, we propose a simple yet effective framework called controllable feature whitening. We quantify the linear correlation between the target and bias features by the covariance matrix, and eliminate it through the whitening module. Our results systemically demonstrate that removing the linear correlations between features fed into the last linear classifier significantly mitigates the bias, while avoiding the need to model intractable higher-order dependencies. A particular advantage of the proposed method is that it does not require regularization terms or adversarial learning, which often leads to unstable optimization in practice. Furthermore, we show that two fairness criteria, demographic parity and equalized odds, can be effectively handled by whitening with the re-weighted covariance matrix. Consequently, our method controls the trade-off between the utility and fairness of algorithms by adjusting the weighting coefficient. Finally, we validate that our method outperforms existing approaches on four benchmark datasets: Corrupted CIFAR-10, Biased FFHQ, WaterBirds, and Celeb-A.", "date": "2025-07-27", "categories": []}
{"id": "2507.20542", "title": "Improving Group Fairness in Tensor Completion via Imbalance Mitigating Entity Augmentation", "abstract": "Group fairness is important to consider in tensor decomposition to prevent discrimination based on social grounds such as gender or age. Although few works have studied group fairness in tensor decomposition, they suffer from performance degradation. To address this, we propose STAFF(Sparse Tensor Augmentation For Fairness) to improve group fairness by minimizing the gap in completion errors of different groups while reducing the overall tensor completion error. Our main idea is to augment a tensor with augmented entities including sufficient observed entries to mitigate imbalance and group bias in the sparse tensor. We evaluate \\method on tensor completion with various datasets under conventional and deep learning-based tensor models. STAFF consistently shows the best trade-off between completion error and group fairness; at most, it yields 36% lower MSE and 59% lower MADE than the second-best baseline.", "date": "2025-07-28", "categories": []}
{"id": "2507.20711", "title": "Algorithmic Fairness: A Runtime Perspective", "abstract": "Fairness in AI is traditionally studied as a static property evaluated once, over a fixed dataset. However, real-world AI systems operate sequentially, with outcomes and environments evolving over time. This paper proposes a framework for analysing fairness as a runtime property. Using a minimal yet expressive model based on sequences of coin tosses with possibly evolving biases, we study the problems of monitoring and enforcing fairness expressed in either toss outcomes or coin biases. Since there is no one-size-fits-all solution for either problem, we provide a summary of monitoring and enforcement strategies, parametrised by environment dynamics, prediction horizon, and confidence thresholds. For both problems, we present general results under simple or minimal assumptions. We survey existing solutions for the monitoring problem for Markovian and additive dynamics, and existing solutions for the enforcement problem in static settings with known dynamics.", "date": "2025-07-28", "categories": []}
{"id": "2212.10678", "title": "Causally Testing Gender Bias in LLMs: A Case Study on Occupational Bias", "abstract": "Generated texts from large language models (LLMs) have been shown to exhibit a variety of harmful, human-like biases against various demographics. These findings motivate research efforts aiming to understand and measure such effects. This paper introduces a causal formulation for bias measurement in generative language models. Based on this theoretical foundation, we outline a list of desiderata for designing robust bias benchmarks. We then propose a benchmark called OccuGender, with a bias-measuring procedure to investigate occupational gender bias. We test several state-of-the-art open-source LLMs on OccuGender, including Llama, Mistral, and their instruction-tuned versions. The results show that these models exhibit substantial occupational gender bias. Lastly, we discuss prompting strategies for bias mitigation and an extension of our causal formulation to illustrate the generalizability of our framework. Our code and data https://github.com/chenyuen0103/gender-bias.", "date": "2022-12-20", "categories": []}
{"id": "2405.12312", "title": "A Principled Approach for Data Bias Mitigation", "abstract": "The widespread use of machine learning and data-driven algorithms for decision making has been steadily increasing over many years. \\emph{Bias} in the data can adversely affect this decision-making. We present a new mitigation strategy to address data bias. Our methods are explainable and come with mathematical guarantees of correctness. They can take advantage of new work on table discovery to find new tuples that can be added to a dataset to create real datasets that are unbiased or less biased. Our framework covers data with non-binary labels and with multiple sensitive attributes. Hence, we are able to measure and mitigate bias that does not appear over a single attribute (or feature), but only intersectionally, when considering a combination of attributes. We evaluate our techniques on publicly available datasets and provide a theoretical analysis of our results, highlighting novel insights into data bias.", "date": "2024-05-20", "categories": []}
{"id": "2412.09900", "title": "Analyzing Fairness of Computer Vision and Natural Language Processing Models", "abstract": "Machine learning (ML) algorithms play a critical role in decision-making across various domains, such as healthcare, finance, education, and law enforcement. However, concerns about fairness and bias in these systems have raised significant ethical and social challenges. To address these challenges, this research utilizes two prominent fairness libraries, Fairlearn by Microsoft and AIF360 by IBM. These libraries offer comprehensive frameworks for fairness analysis, providing tools to evaluate fairness metrics, visualize results, and implement bias mitigation algorithms. The study focuses on assessing and mitigating biases for unstructured datasets using Computer Vision (CV) and Natural Language Processing (NLP) models. The primary objective is to present a comparative analysis of the performance of mitigation algorithms from the two fairness libraries. This analysis involves applying the algorithms individually, one at a time, in one of the stages of the ML lifecycle, pre-processing, in-processing, or post-processing, as well as sequentially across more than one stage. The results reveal that some sequential applications improve the performance of mitigation algorithms by effectively reducing bias while maintaining the model's performance. Publicly available datasets from Kaggle were chosen for this research, providing a practical context for evaluating fairness in real-world machine learning workflows.", "date": "2024-12-13", "categories": []}
{"id": "2507.18348", "title": "VB-Mitigator: An Open-source Framework for Evaluating and Advancing Visual Bias Mitigation", "abstract": "Bias in computer vision models remains a significant challenge, often resulting in unfair, unreliable, and non-generalizable AI systems. Although research into bias mitigation has intensified, progress continues to be hindered by fragmented implementations and inconsistent evaluation practices. Disparate datasets and metrics used across studies complicate reproducibility, making it difficult to fairly assess and compare the effectiveness of various approaches. To overcome these limitations, we introduce the Visual Bias Mitigator (VB-Mitigator), an open-source framework designed to streamline the development, evaluation, and comparative analysis of visual bias mitigation techniques. VB-Mitigator offers a unified research environment encompassing 12 established mitigation methods, 7 diverse benchmark datasets. A key strength of VB-Mitigator is its extensibility, allowing for seamless integration of additional methods, datasets, metrics, and models. VB-Mitigator aims to accelerate research toward fairness-aware computer vision models by serving as a foundational codebase for the research community to develop and assess their approaches. To this end, we also recommend best evaluation practices and provide a comprehensive performance comparison among state-of-the-art methodologies.", "date": "2025-07-24", "categories": []}
{"id": "2507.18561", "title": "Beyond Internal Data: Constructing Complete Datasets for Fairness Testing", "abstract": "As AI becomes prevalent in high-risk domains and decision-making, it is essential to test for potential harms and biases. This urgency is reflected by the global emergence of AI regulations that emphasise fairness and adequate testing, with some mandating independent bias audits. However, procuring the necessary data for fairness testing remains a significant challenge. Particularly in industry settings, legal and privacy concerns restrict the collection of demographic data required to assess group disparities, and auditors face practical and cultural challenges in gaining access to data. Further, internal historical datasets are often insufficiently representative to identify real-world biases. This work focuses on evaluating classifier fairness when complete datasets including demographics are inaccessible. We propose leveraging separate overlapping datasets to construct complete synthetic data that includes demographic information and accurately reflects the underlying relationships between protected attributes and model features. We validate the fidelity of the synthetic data by comparing it to real data, and empirically demonstrate that fairness metrics derived from testing on such synthetic data are consistent with those obtained from real data. This work, therefore, offers a path to overcome real-world data scarcity for fairness testing, enabling independent, model-agnostic evaluation of fairness, and serving as a viable substitute where real data is limited.", "date": "2025-07-24", "categories": []}
{"id": "2507.16989", "title": "Obscured but Not Erased: Evaluating Nationality Bias in LLMs via Name-Based Bias Benchmarks", "abstract": "Large Language Models (LLMs) can exhibit latent biases towards specific nationalities even when explicit demographic markers are not present. In this work, we introduce a novel name-based benchmarking approach derived from the Bias Benchmark for QA (BBQ) dataset to investigate the impact of substituting explicit nationality labels with culturally indicative names, a scenario more reflective of real-world LLM applications. Our novel approach examines how this substitution affects both bias magnitude and accuracy across a spectrum of LLMs from industry leaders such as OpenAI, Google, and Anthropic. Our experiments show that small models are less accurate and exhibit more bias compared to their larger counterparts. For instance, on our name-based dataset and in the ambiguous context (where the correct choice is not revealed), Claude Haiku exhibited the worst stereotypical bias scores of 9%, compared to only 3.5% for its larger counterpart, Claude Sonnet, where the latter also outperformed it by 117.7% in accuracy. Additionally, we find that small models retain a larger portion of existing errors in these ambiguous contexts. For example, after substituting names for explicit nationality references, GPT-4o retains 68% of the error rate versus 76% for GPT-4o-mini, with similar findings for other model providers, in the ambiguous context. Our research highlights the stubborn resilience of biases in LLMs, underscoring their profound implications for the development and deployment of AI systems in diverse, global contexts.", "date": "2025-07-22", "categories": []}
{"id": "2507.16410", "title": "GG-BBQ: German Gender Bias Benchmark for Question Answering", "abstract": "Within the context of Natural Language Processing (NLP), fairness evaluation is often associated with the assessment of bias and reduction of associated harm. In this regard, the evaluation is usually carried out by using a benchmark dataset, for a task such as Question Answering, created for the measurement of bias in the model's predictions along various dimensions, including gender identity. In our work, we evaluate gender bias in German Large Language Models (LLMs) using the Bias Benchmark for Question Answering by Parrish et al. (2022) as a reference. Specifically, the templates in the gender identity subset of this English dataset were machine translated into German. The errors in the machine translated templates were then manually reviewed and corrected with the help of a language expert. We find that manual revision of the translation is crucial when creating datasets for gender bias evaluation because of the limitations of machine translation from English to a language such as German with grammatical gender. Our final dataset is comprised of two subsets: Subset-I, which consists of group terms related to gender identity, and Subset-II, where group terms are replaced with proper names. We evaluate several LLMs used for German NLP on this newly created dataset and report the accuracy and bias scores. The results show that all models exhibit bias, both along and against existing social stereotypes.", "date": "2025-07-22", "categories": []}
{"id": "2507.16557", "title": "Exploring Gender Bias in Large Language Models: An In-depth Dive into the German Language", "abstract": "In recent years, various methods have been proposed to evaluate gender bias in large language models (LLMs). A key challenge lies in the transferability of bias measurement methods initially developed for the English language when applied to other languages. This work aims to contribute to this research strand by presenting five German datasets for gender bias evaluation in LLMs. The datasets are grounded in well-established concepts of gender bias and are accessible through multiple methodologies. Our findings, reported for eight multilingual LLM models, reveal unique challenges associated with gender bias in German, including the ambiguous interpretation of male occupational terms and the influence of seemingly neutral nouns on gender perception. This work contributes to the understanding of gender bias in LLMs across languages and underscores the necessity for tailored evaluation frameworks.", "date": "2025-07-22", "categories": []}
{"id": "2502.21284", "title": "Controlled Model Debiasing through Minimal and Interpretable Updates", "abstract": "Traditional approaches to learning fair machine learning models often require rebuilding models from scratch, typically without considering potentially existing models. In a context where models need to be retrained frequently, this can lead to inconsistent model updates, as well as redundant and costly validation testing. To address this limitation, we introduce the notion of controlled model debiasing, a novel supervised learning task relying on two desiderata: that the differences between the new fair model and the existing one should be (i) minimal and (ii) interpretable. After providing theoretical guarantees to this new problem, we introduce a novel algorithm for algorithmic fairness, COMMOD, that is both model-agnostic and does not require the sensitive attribute at test time. In addition, our algorithm is explicitly designed to enforce minimal and interpretable changes between biased and debiased predictions in a binary classification task, a property that, while highly desirable in high-stakes applications, is rarely prioritized as an explicit objective in fairness literature. Our approach combines a concept-based architecture and adversarial learning and we demonstrate through empirical results that it achieves comparable performance to state-of-the-art debiasing methods while performing minimal and interpretable prediction changes.", "date": "2025-02-28", "categories": []}
{"id": "2507.14176", "title": "Predictive Representativity: Uncovering Racial Bias in AI-based Skin Cancer Detection", "abstract": "Artificial intelligence (AI) systems increasingly inform medical decision-making, yet concerns about algorithmic bias and inequitable outcomes persist, particularly for historically marginalized populations. This paper introduces the concept of Predictive Representativity (PR), a framework of fairness auditing that shifts the focus from the composition of the data set to outcomes-level equity. Through a case study in dermatology, we evaluated AI-based skin cancer classifiers trained on the widely used HAM10000 dataset and on an independent clinical dataset (BOSQUE Test set) from Colombia. Our analysis reveals substantial performance disparities by skin phototype, with classifiers consistently underperforming for individuals with darker skin, despite proportional sampling in the source data. We argue that representativity must be understood not as a static feature of datasets but as a dynamic, context-sensitive property of model predictions. PR operationalizes this shift by quantifying how reliably models generalize fairness across subpopulations and deployment contexts. We further propose an External Transportability Criterion that formalizes the thresholds for fairness generalization. Our findings highlight the ethical imperative for post-hoc fairness auditing, transparency in dataset documentation, and inclusive model validation pipelines. This work offers a scalable tool for diagnosing structural inequities in AI systems, contributing to discussions on equity, interpretability, and data justice and fostering a critical re-evaluation of fairness in data-driven healthcare.", "date": "2025-07-10", "categories": []}
{"id": "2507.14238", "title": "Language Models Change Facts Based on the Way You Talk", "abstract": "Large language models (LLMs) are increasingly being used in user-facing applications, from providing medical consultations to job interview advice. Recent research suggests that these models are becoming increasingly proficient at inferring identity information about the author of a piece of text from linguistic patterns as subtle as the choice of a few words. However, little is known about how LLMs use this information in their decision-making in real-world applications. We perform the first comprehensive analysis of how identity markers present in a user's writing bias LLM responses across five different high-stakes LLM applications in the domains of medicine, law, politics, government benefits, and job salaries. We find that LLMs are extremely sensitive to markers of identity in user queries and that race, gender, and age consistently influence LLM responses in these applications. For instance, when providing medical advice, we find that models apply different standards of care to individuals of different ethnicities for the same symptoms; we find that LLMs are more likely to alter answers to align with a conservative (liberal) political worldview when asked factual questions by older (younger) individuals; and that LLMs recommend lower salaries for non-White job applicants and higher salaries for women compared to men. Taken together, these biases mean that the use of off-the-shelf LLMs for these applications may cause harmful differences in medical care, foster wage gaps, and create different political factual realities for people of different identities. Beyond providing an analysis, we also provide new tools for evaluating how subtle encoding of identity in users' language choices impacts model decisions. Given the serious implications of these findings, we recommend that similar thorough assessments of LLM use in user-facing applications are conducted before future deployment.", "date": "2025-07-17", "categories": []}
{"id": "2507.14352", "title": "A Reproducibility Study of Product-side Fairness in Bundle Recommendation", "abstract": "Recommender systems are known to exhibit fairness issues, particularly on the product side, where products and their associated suppliers receive unequal exposure in recommended results. While this problem has been widely studied in traditional recommendation settings, its implications for bundle recommendation (BR) remain largely unexplored. This emerging task introduces additional complexity: recommendations are generated at the bundle level, yet user satisfaction and product (or supplier) exposure depend on both the bundle and the individual items it contains. Existing fairness frameworks and metrics designed for traditional recommender systems may not directly translate to this multi-layered setting. In this paper, we conduct a comprehensive reproducibility study of product-side fairness in BR across three real-world datasets using four state-of-the-art BR methods. We analyze exposure disparities at both the bundle and item levels using multiple fairness metrics, uncovering important patterns. Our results show that exposure patterns differ notably between bundles and items, revealing the need for fairness interventions that go beyond bundle-level assumptions. We also find that fairness assessments vary considerably depending on the metric used, reinforcing the need for multi-faceted evaluation. Furthermore, user behavior plays a critical role: when users interact more frequently with bundles than with individual items, BR systems tend to yield fairer exposure distributions across both levels. Overall, our findings offer actionable insights for building fairer bundle recommender systems and establish a vital foundation for future research in this emerging domain.", "date": "2025-07-18", "categories": []}
{"id": "2507.11548", "title": "Fairness Is Not Enough: Auditing Competence and Intersectional Bias in AI-powered Resume Screening", "abstract": "The increasing use of generative AI for resume screening is predicated on the assumption that it offers an unbiased alternative to biased human decision-making. However, this belief fails to address a critical question: are these AI systems fundamentally competent at the evaluative tasks they are meant to perform?\n  This study investigates the question of competence through a two-part audit of eight major AI platforms. Experiment 1 confirmed complex, contextual racial and gender biases, with some models penalizing candidates merely for the presence of demographic signals. Experiment 2, which evaluated core competence, provided a critical insight: some models that appeared unbiased were, in fact, incapable of performing a substantive evaluation, relying instead on superficial keyword matching.\n  This paper introduces the \"Illusion of Neutrality\" to describe this phenomenon, where an apparent lack of bias is merely a symptom of a model's inability to make meaningful judgments. This study recommends that organizations and regulators adopt a dual-validation framework, auditing AI hiring tools for both demographic bias and demonstrable competence to ensure they are both equitable and effective.", "date": "2025-07-11", "categories": []}
{"id": "2507.12983", "title": "FedGA: A Fair Federated Learning Framework Based on the Gini Coefficient", "abstract": "Fairness has emerged as one of the key challenges in federated learning. In horizontal federated settings, data heterogeneity often leads to substantial performance disparities across clients, raising concerns about equitable model behavior. To address this issue, we propose FedGA, a fairness-aware federated learning algorithm. We first employ the Gini coefficient to measure the performance disparity among clients. Based on this, we establish a relationship between the Gini coefficient $G$ and the update scale of the global model ${U_s}$, and use this relationship to adaptively determine the timing of fairness intervention. Subsequently, we dynamically adjust the aggregation weights according to the system's real-time fairness status, enabling the global model to better incorporate information from clients with relatively poor performance.We conduct extensive experiments on the Office-Caltech-10, CIFAR-10, and Synthetic datasets. The results show that FedGA effectively improves fairness metrics such as variance and the Gini coefficient, while maintaining strong overall performance, demonstrating the effectiveness of our approach.", "date": "2025-07-17", "categories": []}
{"id": "2507.12021", "title": "Incorporating Fairness Constraints into Archetypal Analysis", "abstract": "Archetypal Analysis (AA) is an unsupervised learning method that represents data as convex combinations of extreme patterns called archetypes. While AA provides interpretable and low-dimensional representations, it can inadvertently encode sensitive attributes, leading to fairness concerns. In this work, we propose Fair Archetypal Analysis (FairAA), a modified formulation that explicitly reduces the influence of sensitive group information in the learned projections. We also introduce FairKernelAA, a nonlinear extension that addresses fairness in more complex data distributions. Our approach incorporates a fairness regularization term while preserving the structure and interpretability of the archetypes. We evaluate FairAA and FairKernelAA on synthetic datasets, including linear, nonlinear, and multi-group scenarios, demonstrating their ability to reduce group separability -- as measured by mean maximum discrepancy and linear separability -- without substantially compromising explained variance. We further validate our methods on the real-world ANSUR I dataset, confirming their robustness and practical utility. The results show that FairAA achieves a favorable trade-off between utility and fairness, making it a promising tool for responsible representation learning in sensitive applications.", "date": "2025-07-16", "categories": []}
{"id": "2403.06031", "title": "FairTargetSim: An Interactive Simulator for Understanding and Explaining the Fairness Effects of Target Variable Definition", "abstract": "Machine learning requires defining one's target variable for predictions or decisions, a process that can have profound implications for fairness, since biases are often encoded in target variable definition itself, before any data collection or training. The downstream impacts of target variable definition must be taken into account in order to responsibly develop, deploy, and use the algorithmic systems. We propose FairTargetSim (FTS), an interactive and simulation-based approach for this. We demonstrate FTS using the example of algorithmic hiring, grounded in real-world data and user-defined target variables. FTS is open-source; it can be used by algorithm developers, non-technical stakeholders, researchers, and educators in a number of ways. FTS is available at: http://tinyurl.com/ftsinterface. The video accompanying this paper is here: http://tinyurl.com/ijcaifts.", "date": "2024-03-09", "categories": []}
{"id": "2411.04371", "title": "ComFairGNN: Community Fair Graph Neural Network", "abstract": "Graph Neural Networks (GNNs) have become the leading approach for addressing graph analytical problems in various real-world scenarios. However, GNNs may produce biased predictions against certain demographic subgroups due to node attributes and neighbors surrounding a node. Most current research on GNN fairness focuses predominantly on debiasing GNNs using oversimplified fairness evaluation metrics, which can give a misleading impression of fairness. Understanding the potential evaluation paradoxes due to the complicated nature of the graph structure is crucial for developing effective GNN debiasing mechanisms. In this paper, we examine the effectiveness of current GNN debiasing methods in terms of unfairness evaluation. Specifically, we introduce a community-level strategy to measure bias in GNNs and evaluate debiasing methods at this level. Further, We introduce ComFairGNN, a novel framework designed to mitigate community-level bias in GNNs. Our approach employs a learnable coreset-based debiasing function that addresses bias arising from diverse local neighborhood distributions during GNNs neighborhood aggregation. Comprehensive evaluations on three benchmark datasets demonstrate our model's effectiveness in both accuracy and fairness metrics.", "date": "2024-11-07", "categories": []}
{"id": "2507.11247", "title": "Fairness-Aware Grouping for Continuous Sensitive Variables: Application for Debiasing Face Analysis with respect to Skin Tone", "abstract": "Within a legal framework, fairness in datasets and models is typically assessed by dividing observations into predefined groups and then computing fairness measures (e.g., Disparate Impact or Equality of Odds with respect to gender). However, when sensitive attributes such as skin color are continuous, dividing into default groups may overlook or obscure the discrimination experienced by certain minority subpopulations. To address this limitation, we propose a fairness-based grouping approach for continuous (possibly multidimensional) sensitive attributes. By grouping data according to observed levels of discrimination, our method identifies the partition that maximizes a novel criterion based on inter-group variance in discrimination, thereby isolating the most critical subgroups.\n  We validate the proposed approach using multiple synthetic datasets and demonstrate its robustness under changing population distributions - revealing how discrimination is manifested within the space of sensitive attributes. Furthermore, we examine a specialized setting of monotonic fairness for the case of skin color. Our empirical results on both CelebA and FFHQ, leveraging the skin tone as predicted by an industrial proprietary algorithm, show that the proposed segmentation uncovers more nuanced patterns of discrimination than previously reported, and that these findings remain stable across datasets for a given model. Finally, we leverage our grouping model for debiasing purpose, aiming at predicting fair scores with group-by-group post-processing. The results demonstrate that our approach improves fairness while having minimal impact on accuracy, thus confirming our partition method and opening the door for industrial deployment.", "date": "2025-07-15", "categories": []}
{"id": "2507.11344", "title": "Guiding LLM Decision-Making with Fairness Reward Models", "abstract": "Large language models are increasingly used to support high-stakes decisions, potentially influencing who is granted bail or receives a loan. Naive chain-of-thought sampling can improve average decision accuracy, but has also been shown to amplify unfair bias. To address this challenge and enable the trustworthy use of reasoning models in high-stakes decision-making, we propose a framework for training a generalizable Fairness Reward Model (FRM). Our model assigns a fairness score to LLM reasoning, enabling the system to down-weight biased trajectories and favor equitable ones when aggregating decisions across reasoning chains. We show that a single Fairness Reward Model, trained on weakly supervised, LLM-annotated examples of biased versus unbiased reasoning, transfers across tasks, domains, and model families without additional fine-tuning. Applied to real-world decision-making tasks including recidivism prediction and social media moderation, we show that our approach consistently improves fairness while matching, or even surpassing, baseline accuracy.", "date": "2025-07-15", "categories": []}
{"id": "2505.21363", "title": "Subgroups Matter for Robust Bias Mitigation", "abstract": "Despite the constant development of new bias mitigation methods for machine learning, no method consistently succeeds, and a fundamental question remains unanswered: when and why do bias mitigation techniques fail? In this paper, we hypothesise that a key factor may be the often-overlooked but crucial step shared by many bias mitigation methods: the definition of subgroups. To investigate this, we conduct a comprehensive evaluation of state-of-the-art bias mitigation methods across multiple vision and language classification tasks, systematically varying subgroup definitions, including coarse, fine-grained, intersectional, and noisy subgroups. Our results reveal that subgroup choice significantly impacts performance, with certain groupings paradoxically leading to worse outcomes than no mitigation at all. Our findings suggest that observing a disparity between a set of subgroups is not a sufficient reason to use those subgroups for mitigation. Through theoretical analysis, we explain these phenomena and uncover a counter-intuitive insight that, in some cases, improving fairness with respect to a particular set of subgroups is best achieved by using a different set of subgroups for mitigation. Our work highlights the importance of careful subgroup definition in bias mitigation and presents it as an alternative lever for improving the robustness and fairness of machine learning models.", "date": "2025-05-27", "categories": []}
{"id": "2507.02679", "title": "Exploring Gender Bias Beyond Occupational Titles", "abstract": "In this work, we investigate the correlation between gender and contextual biases, focusing on elements such as action verbs, object nouns, and particularly on occupations. We introduce a novel dataset, GenderLexicon, and a framework that can estimate contextual bias and its related gender bias. Our model can interpret the bias with a score and thus improve the explainability of gender bias. Also, our findings confirm the existence of gender biases beyond occupational stereotypes. To validate our approach and demonstrate its effectiveness, we conduct evaluations on five diverse datasets, including a Japanese dataset.", "date": "2025-07-03", "categories": []}
{"id": "2507.08866", "title": "Underrepresentation, Label Bias, and Proxies: Towards Data Bias Profiles for the EU AI Act and Beyond", "abstract": "Undesirable biases encoded in the data are key drivers of algorithmic discrimination. Their importance is widely recognized in the algorithmic fairness literature, as well as legislation and standards on anti-discrimination in AI. Despite this recognition, data biases remain understudied, hindering the development of computational best practices for their detection and mitigation. In this work, we present three common data biases and study their individual and joint effect on algorithmic discrimination across a variety of datasets, models, and fairness measures. We find that underrepresentation of vulnerable populations in training sets is less conducive to discrimination than conventionally affirmed, while combinations of proxies and label bias can be far more critical. Consequently, we develop dedicated mechanisms to detect specific types of bias, and combine them into a preliminary construct we refer to as the Data Bias Profile (DBP). This initial formulation serves as a proof of concept for how different bias signals can be systematically documented. Through a case study with popular fairness datasets, we demonstrate the effectiveness of the DBP in predicting the risk of discriminatory outcomes and the utility of fairness-enhancing interventions. Overall, this article bridges algorithmic fairness research and anti-discrimination policy through a data-centric lens.", "date": "2025-07-09", "categories": []}
{"id": "2507.10154", "title": "Simulating Biases for Interpretable Fairness in Offline and Online Classifiers", "abstract": "Predictive models often reinforce biases which were originally embedded in their training data, through skewed decisions. In such cases, mitigation methods are critical to ensure that, regardless of the prevailing disparities, model outcomes are adjusted to be fair. To assess this, datasets could be systematically generated with specific biases, to train machine learning classifiers. Then, predictive outcomes could aid in the understanding of this bias embedding process. Hence, an agent-based model (ABM), depicting a loan application process that represents various systemic biases across two demographic groups, was developed to produce synthetic datasets. Then, by applying classifiers trained on them to predict loan outcomes, we can assess how biased data leads to unfairness. This highlights a main contribution of this work: a framework for synthetic dataset generation with controllable bias injection. We also contribute with a novel explainability technique, which shows how mitigations affect the way classifiers leverage data features, via second-order Shapley values. In experiments, both offline and online learning approaches are employed. Mitigations are applied at different stages of the modelling pipeline, such as during pre-processing and in-processing.", "date": "2025-07-14", "categories": []}
{"id": "2501.14710", "title": "Overcoming Fairness Trade-offs via Pre-processing: A Causal Perspective", "abstract": "Training machine learning models for fair decisions faces two key challenges: The \\emph{fairness-accuracy trade-off} results from enforcing fairness which weakens its predictive performance in contrast to an unconstrained model. The incompatibility of different fairness metrics poses another trade-off -- also known as the \\emph{impossibility theorem}. Recent work identifies the bias within the observed data as a possible root cause and shows that fairness and predictive performance are in fact in accord when predictive performance is measured on unbiased data. We offer a causal explanation for these findings using the framework of the FiND (fictitious and normatively desired) world, a \"fair\" world, where protected attributes have no causal effects on the target variable. We show theoretically that (i) classical fairness metrics deemed to be incompatible are naturally satisfied in the FiND world, while (ii) fairness aligns with high predictive performance. We extend our analysis by suggesting how one can benefit from these theoretical insights in practice, using causal pre-processing methods that approximate the FiND world. Additionally, we propose a method for evaluating the approximation of the FiND world via pre-processing in practical use cases where we do not have access to the FiND world. In simulations and empirical studies, we demonstrate that these pre-processing methods are successful in approximating the FiND world and resolve both trade-offs. Our results provide actionable solutions for practitioners to achieve fairness and high predictive performance simultaneously.", "date": "2025-01-24", "categories": []}
{"id": "2501.19334", "title": "The Value of Prediction in Identifying the Worst-Off", "abstract": "Machine learning is increasingly used in government programs to identify and support the most vulnerable individuals, prioritizing assistance for those at greatest risk over optimizing aggregate outcomes. This paper examines the welfare impacts of prediction in equity-driven contexts, and how they compare to other policy levers, such as expanding bureaucratic capacity. Through mathematical models and a real-world case study on long-term unemployment amongst German residents, we develop a comprehensive understanding of the relative effectiveness of prediction in surfacing the worst-off. Our findings provide clear analytical frameworks and practical, data-driven tools that empower policymakers to make principled decisions when designing these systems.", "date": "2025-01-31", "categories": []}
{"id": "2503.02870", "title": "Multiaccuracy and Multicalibration via Proxy Groups", "abstract": "As the use of predictive machine learning algorithms increases in high-stakes decision-making, it is imperative that these algorithms are fair across sensitive groups. However, measuring and enforcing fairness in real-world applications can be challenging due to the missing or incomplete sensitive group information. Proxy-sensitive attributes have been proposed as a practical and effective solution in these settings, but only for parity-based fairness notions. Knowing how to evaluate and control for fairness with missing sensitive group data for newer, different, and more flexible frameworks, such as multiaccuracy and multicalibration, remain unexplored. In this work, we address this gap by demonstrating that in the absence of sensitive group data, proxy-sensitive attributes can provably used to derive actionable upper bounds on the true multiaccuracy and multicalibration violations, providing insights into a predictive model's potential worst-case fairness violations. Additionally, we show that adjusting models to satisfy multiaccuracy and multicalibration across proxy-sensitive attributes can significantly mitigate these violations for the true, but unknown, sensitive groups. Through several experiments on real-world datasets, we illustrate that approximate multiaccuracy and multicalibration can be achieved even when sensitive group data is incomplete or unavailable.", "date": "2025-03-04", "categories": []}
{"id": "2507.07274", "title": "LinguaMark: Do Multimodal Models Speak Fairly? A Benchmark-Based Evaluation", "abstract": "Large Multimodal Models (LMMs) are typically trained on vast corpora of image-text data but are often limited in linguistic coverage, leading to biased and unfair outputs across languages. While prior work has explored multimodal evaluation, less emphasis has been placed on assessing multilingual capabilities. In this work, we introduce LinguaMark, a benchmark designed to evaluate state-of-the-art LMMs on a multilingual Visual Question Answering (VQA) task. Our dataset comprises 6,875 image-text pairs spanning 11 languages and five social attributes. We evaluate models using three key metrics: Bias, Answer Relevancy, and Faithfulness. Our findings reveal that closed-source models generally achieve the highest overall performance. Both closed-source (GPT-4o and Gemini2.5) and open-source models (Gemma3, Qwen2.5) perform competitively across social attributes, and Qwen2.5 demonstrates strong generalization across multiple languages. We release our benchmark and evaluation code to encourage reproducibility and further research.", "date": "2025-07-09", "categories": []}
{"id": "2301.08837", "title": "From Pseudorandomness to Multi-Group Fairness and Back", "abstract": "We identify and explore connections between the recent literature on multi-group fairness for prediction algorithms and the pseudorandomness notions of leakage-resilience and graph regularity. We frame our investigation using new variants of multicalibration based on statistical distance and closely related to the concept of outcome indistinguishability. Adopting this perspective leads us not only to new, more efficient algorithms for multicalibration, but also to our graph theoretic results and a proof of a novel hardcore lemma for real-valued functions.", "date": "2023-01-21", "categories": []}
{"id": "2502.16380", "title": "Understanding Fixed Predictions via Confined Regions", "abstract": "Machine learning models can assign fixed predictions that preclude individuals from changing their outcome. Existing approaches to audit fixed predictions do so on a pointwise basis, which requires access to an existing dataset of individuals and may fail to anticipate fixed predictions in out-of-sample data. This work presents a new paradigm to identify fixed predictions by finding confined regions of the feature space in which all individuals receive fixed predictions. This paradigm enables the certification of recourse for out-of-sample data, works in settings without representative datasets, and provides interpretable descriptions of individuals with fixed predictions. We develop a fast method to discover confined regions for linear classifiers using mixed-integer quadratically constrained programming. We conduct a comprehensive empirical study of confined regions across diverse applications. Our results highlight that existing pointwise verification methods fail to anticipate future individuals with fixed predictions, while our method both identifies them and provides an interpretable description.", "date": "2025-02-22", "categories": []}
{"id": "2503.18962", "title": "Representative Ranking for Deliberation in the Public Sphere", "abstract": "Online comment sections, such as those on news sites or social media, have the potential to foster informal public deliberation, However, this potential is often undermined by the frequency of toxic or low-quality exchanges that occur in these settings. To combat this, platforms increasingly leverage algorithmic ranking to facilitate higher-quality discussions, e.g., by using civility classifiers or forms of prosocial ranking. Yet, these interventions may also inadvertently reduce the visibility of legitimate viewpoints, undermining another key aspect of deliberation: representation of diverse views. We seek to remedy this problem by introducing guarantees of representation into these methods. In particular, we adopt the notion of justified representation (JR) from the social choice literature and incorporate a JR constraint into the comment ranking setting. We find that enforcing JR leads to greater inclusion of diverse viewpoints while still being compatible with optimizing for user engagement or other measures of conversational quality.", "date": "2025-03-19", "categories": []}
{"id": "2405.05031", "title": "Mitigating Bias Using Model-Agnostic Data Attribution", "abstract": "Mitigating bias in machine learning models is a critical endeavor for ensuring fairness and equity. In this paper, we propose a novel approach to address bias by leveraging pixel image attributions to identify and regularize regions of images containing significant information about bias attributes. Our method utilizes a model-agnostic approach to extract pixel attributions by employing a convolutional neural network (CNN) classifier trained on small image patches. By training the classifier to predict a property of the entire image using only a single patch, we achieve region-based attributions that provide insights into the distribution of important information across the image. We propose utilizing these attributions to introduce targeted noise into datasets with confounding attributes that bias the data, thereby constraining neural networks from learning these biases and emphasizing the primary attributes. Our approach demonstrates its efficacy in enabling the training of unbiased classifiers on heavily biased datasets.", "date": "2024-05-08", "categories": []}
{"id": "2409.11598", "title": "Towards Fair RAG: On the Impact of Fair Ranking in Retrieval-Augmented Generation", "abstract": "Despite the central role of retrieval in retrieval-augmented generation (RAG) systems, much of the existing research on RAG overlooks the well-established field of fair ranking and fails to account for the interests of all stakeholders involved. In this paper, we conduct the first systematic evaluation of RAG systems that integrate fairness-aware rankings, addressing both ranking fairness and attribution fairness, which ensures equitable exposure of the sources cited in the generated content. Our evaluation focuses on measuring item-side fairness, specifically the fair exposure of relevant items retrieved by RAG systems, and investigates how this fairness impacts both the effectiveness of the systems and the attribution of sources in the generated output that users ultimately see. By experimenting with twelve RAG models across seven distinct tasks, we show that incorporating fairness-aware retrieval often maintains or even enhances both ranking quality and generation quality, countering the common belief that fairness compromises system performance. Additionally, we demonstrate that fair retrieval practices lead to more balanced attribution in the final responses, ensuring that the generator fairly cites the sources it relies on. Our findings underscore the importance of item-side fairness in retrieval and generation, laying the foundation for responsible and equitable RAG systems and guiding future research in fair ranking and attribution.", "date": "2024-09-17", "categories": []}
{"id": "2502.05286", "title": "Fairness and Sparsity within Rashomon sets: Enumeration-Free Exploration and Characterization", "abstract": "We introduce an enumeration-free method based on mathematical programming to precisely characterize various properties such as fairness or sparsity within the set of \"good models\", known as Rashomon set. This approach is generically applicable to any hypothesis class, provided that a mathematical formulation of the model learning task exists. It offers a structured framework to define the notion of business necessity and evaluate how fairness can be improved or degraded towards a specific protected group, while remaining within the Rashomon set and maintaining any desired sparsity level.\n  We apply our approach to two hypothesis classes: scoring systems and decision diagrams, leveraging recent mathematical programming formulations for training such models. As seen in our experiments, the method comprehensively and certifiably quantifies trade-offs between predictive performance, sparsity, and fairness. We observe that a wide range of fairness values are attainable, ranging from highly favorable to significantly unfavorable for a protected group, while staying within less than 1% of the best possible training accuracy for the hypothesis class. Additionally, we observe that sparsity constraints limit these trade-offs and may disproportionately harm specific subgroups. As we evidenced, thoroughly characterizing the tensions between these key aspects is critical for an informed and accountable selection of models.", "date": "2025-02-07", "categories": []}
{"id": "2504.14492", "title": "FairSteer: Inference Time Debiasing for LLMs with Dynamic Activation Steering", "abstract": "Large language models (LLMs) are prone to capturing biases from training corpus, leading to potential negative social impacts. Existing prompt-based debiasing methods exhibit instability due to their sensitivity to prompt changes, while fine-tuning-based techniques incur substantial computational overhead and catastrophic forgetting. In this paper, we propose FairSteer, a novel inference-time debiasing framework without requiring customized prompt design or model retraining. Motivated by the linear representation hypothesis, our preliminary investigation demonstrates that fairness-related features can be encoded into separable directions in the hidden activation space. FairSteer operates in three steps: biased activation detection, debiasing steering vector (DSV) computation, and dynamic activation steering. Specifically, it first trains a lightweight linear classifier to detect bias signatures in activations, and then computes DSVs as intervention directions derived from small contrastive prompt pairs. Subsequently, it performs debiasing by adjusting activations with DSVs in the inference stage. Comprehensive evaluation with six LLMs demonstrates the superiority of FairSteer across question-answering, counterfactual input evaluation and open-ended text generation tasks. Code will be released.", "date": "2025-04-20", "categories": []}
{"id": "2505.08829", "title": "Aggregating Concepts of Fairness and Accuracy in Prediction Algorithms", "abstract": "An algorithm that outputs predictions about the state of the world will almost always be designed with the implicit or explicit goal of outputting accurate predictions (i.e., predictions that are likely to be true). In addition, the rise of increasingly powerful predictive algorithms brought about by the recent revolution in artificial intelligence has led to an emphasis on building predictive algorithms that are fair, in the sense that their predictions do not systematically evince bias or bring about harm to certain individuals or groups. This state of affairs presents two conceptual challenges. First, the goals of accuracy and fairness can sometimes be in tension, and there are no obvious normative guidelines for managing the trade-offs between these two desiderata when they arise. Second, there are many distinct ways of measuring both the accuracy and fairness of a predictive algorithm; here too, there are no obvious guidelines on how to aggregate our preferences for predictive algorithms that satisfy disparate measures of fairness and accuracy to various extents. The goal of this paper is to address these challenges by arguing that there are good reasons for using a linear combination of accuracy and fairness metrics to measure the all-things-considered value of a predictive algorithm for agents who care about both accuracy and fairness. My argument depends crucially on a classic result in the preference aggregation literature due to Harsanyi. After making this formal argument, I apply my result to an analysis of accuracy-fairness trade-offs using the COMPAS dataset compiled by Angwin et al.", "date": "2025-05-13", "categories": []}
{"id": "2505.18953", "title": "Evaluating AI for Finance: Is AI Credible at Assessing Investment Risk?", "abstract": "We assess whether AI systems can credibly evaluate investment risk appetite-a task that must be thoroughly validated before automation. Our analysis was conducted on proprietary systems (GPT, Claude, Gemini) and open-weight models (LLaMA, DeepSeek, Mistral), using carefully curated user profiles that reflect real users with varying attributes such as country and gender. As a result, the models exhibit significant variance in score distributions when user attributes-such as country or gender-that should not influence risk computation are changed. For example, GPT-4o assigns higher risk scores to Nigerian and Indonesian profiles. While some models align closely with expected scores in the Low- and Mid-risk ranges, none maintain consistent scores across regions and demographics, thereby violating AI and finance regulations.", "date": "2025-05-25", "categories": []}
{"id": "2507.03048", "title": "Monitoring of Static Fairness", "abstract": "Machine-learned systems are in widespread use for making decisions about humans, and it is important that they are fair, i.e., not biased against individuals based on sensitive attributes.\n  We present a general framework of runtime verification of algorithmic fairness for systems whose models are unknown, but are assumed to have a Markov chain structure, with or without full observation of the state space.\n  We introduce a specification language that can model many common algorithmic fairness properties, such as demographic parity, equal opportunity, and social burden.\n  We build monitors that observe a long sequence of events as generated by a given system, and output, after each observation, a quantitative estimate of how fair or biased the system was on that run until that point in time.\n  The estimate is proven to be correct modulo a variable error bound and a given confidence level, where the error bound gets tighter as the observed sequence gets longer.\n  We present two categories of monitoring algorithms, namely ones with a uniform error bound across all time points, and ones with weaker non-uniform, pointwise error bounds at different time points.\n  Our monitoring algorithms use statistical tools that are adapted to suit the dynamic requirements of monitoring and the special needs of the fairness specifications.\n  Using a prototype implementation, we show how we can monitor if a bank is fair in giving loans to applicants from different social backgrounds, and if a college is fair in admitting students while maintaining a reasonable financial burden on the society.\n  In these experiments, our monitors took less than a millisecond to update their verdicts after each observation.", "date": "2025-07-03", "categories": []}
{"id": "2507.03142", "title": "From Measurement to Mitigation: Exploring the Transferability of Debiasing Approaches to Gender Bias in Maltese Language Models", "abstract": "The advancement of Large Language Models (LLMs) has transformed Natural Language Processing (NLP), enabling performance across diverse tasks with little task-specific training. However, LLMs remain susceptible to social biases, particularly reflecting harmful stereotypes from training data, which can disproportionately affect marginalised communities. We measure gender bias in Maltese LMs, arguing that such bias is harmful as it reinforces societal stereotypes and fails to account for gender diversity, which is especially problematic in gendered, low-resource languages. While bias evaluation and mitigation efforts have progressed for English-centric models, research on low-resourced and morphologically rich languages remains limited. This research investigates the transferability of debiasing methods to Maltese language models, focusing on BERTu and mBERTu, BERT-based monolingual and multilingual models respectively. Bias measurement and mitigation techniques from English are adapted to Maltese, using benchmarks such as CrowS-Pairs and SEAT, alongside debiasing methods Counterfactual Data Augmentation, Dropout Regularization, Auto-Debias, and GuiDebias. We also contribute to future work in the study of gender bias in Maltese by creating evaluation datasets. Our findings highlight the challenges of applying existing bias mitigation methods to linguistically complex languages, underscoring the need for more inclusive approaches in the development of multilingual NLP.", "date": "2025-07-03", "categories": []}
{"id": "2507.03728", "title": "FAROS: Fair Graph Generation via Attribute Switching Mechanisms", "abstract": "Recent advancements in graph diffusion models (GDMs) have enabled the synthesis of realistic network structures, yet ensuring fairness in the generated data remains a critical challenge. Existing solutions attempt to mitigate bias by re-training the GDMs with ad-hoc fairness constraints. Conversely, with this work, we propose FAROS, a novel FAir graph geneRatiOn framework leveraging attribute Switching mechanisms and directly running in the generation process of the pre-trained GDM. Technically, our approach works by altering nodes' sensitive attributes during the generation. To this end, FAROS calculates the optimal fraction of switching nodes, and selects the diffusion step to perform the switch by setting tailored multi-criteria constraints to preserve the node-topology profile from the original distribution (a proxy for accuracy) while ensuring the edge independence on the sensitive attributes for the generated graph (a proxy for fairness). Our experiments on benchmark datasets for link prediction demonstrate that the proposed approach effectively reduces fairness discrepancies while maintaining comparable (or even higher) accuracy performance to other similar baselines. Noteworthy, FAROS is also able to strike a better accuracy-fairness trade-off than other competitors in some of the tested settings under the Pareto optimality concept, demonstrating the effectiveness of the imposed multi-criteria constraints.", "date": "2025-07-04", "categories": []}
{"id": "2407.03133", "title": "Quantifying the Cross-sectoral Intersecting Discrepancies within Multiple Groups Using Latent Class Analysis Towards Fairness", "abstract": "The growing interest in fair AI development is evident. The ''Leave No One Behind'' initiative urges us to address multiple and intersecting forms of inequality in accessing services, resources, and opportunities, emphasising the significance of fairness in AI. This is particularly relevant as an increasing number of AI tools are applied to decision-making processes, such as resource allocation and service scheme development, across various sectors such as health, energy, and housing. Therefore, exploring joint inequalities in these sectors is significant and valuable for thoroughly understanding overall inequality and unfairness. This research introduces an innovative approach to quantify cross-sectoral intersecting discrepancies among user-defined groups using latent class analysis. These discrepancies can be used to approximate inequality and provide valuable insights to fairness issues. We validate our approach using both proprietary and public datasets, including both EVENS and Census 2021 (England & Wales) datasets, to examine cross-sectoral intersecting discrepancies among different ethnic groups. We also verify the reliability of the quantified discrepancy by conducting a correlation analysis with a government public metric. Our findings reveal significant discrepancies both among minority ethnic groups and between minority ethnic groups and non-minority ethnic groups, emphasising the need for targeted interventions in policy-making processes. Furthermore, we demonstrate how the proposed approach can provide valuable insights into ensuring fairness in machine learning systems.", "date": "2024-05-24", "categories": []}
{"id": "2507.02000", "title": "Why Multi-Interest Fairness Matters: Hypergraph Contrastive Multi-Interest Learning for Fair Conversational Recommender System", "abstract": "Unfairness is a well-known challenge in Recommender Systems (RSs), often resulting in biased outcomes that disadvantage users or items based on attributes such as gender, race, age, or popularity. Although some approaches have started to improve fairness recommendation in offline or static contexts, the issue of unfairness often exacerbates over time, leading to significant problems like the Matthew effect, filter bubbles, and echo chambers. To address these challenges, we proposed a novel framework, Hypergraph Contrastive Multi-Interest Learning for Fair Conversational Recommender System (HyFairCRS), aiming to promote multi-interest diversity fairness in dynamic and interactive Conversational Recommender Systems (CRSs). HyFairCRS first captures a wide range of user interests by establishing diverse hypergraphs through contrastive learning. These interests are then utilized in conversations to generate informative responses and ensure fair item predictions within the dynamic user-system feedback loop. Experiments on two CRS-based datasets show that HyFairCRS achieves a new state-of-the-art performance while effectively alleviating unfairness. Our code is available at https://github.com/zysensmile/HyFairCRS.", "date": "2025-07-01", "categories": []}
{"id": "2507.02152", "title": "The Illusion of Fairness: Auditing Fairness Interventions with Audit Studies", "abstract": "Artificial intelligence systems, especially those using machine learning, are being deployed in domains from hiring to loan issuance in order to automate these complex decisions. Judging both the effectiveness and fairness of these AI systems, and their human decision making counterpart, is a complex and important topic studied across both computational and social sciences. Within machine learning, a common way to address bias in downstream classifiers is to resample the training data to offset disparities. For example, if hiring rates vary by some protected class, then one may equalize the rate within the training set to alleviate bias in the resulting classifier. While simple and seemingly effective, these methods have typically only been evaluated using data obtained through convenience samples, introducing selection bias and label bias into metrics. Within the social sciences, psychology, public health, and medicine, audit studies, in which fictitious ``testers'' (e.g., resumes, emails, patient actors) are sent to subjects (e.g., job openings, businesses, doctors) in randomized control trials, provide high quality data that support rigorous estimates of discrimination. In this paper, we investigate how data from audit studies can be used to improve our ability to both train and evaluate automated hiring algorithms. We find that such data reveals cases where the common fairness intervention method of equalizing base rates across classes appears to achieve parity using traditional measures, but in fact has roughly 10% disparity when measured appropriately. We additionally introduce interventions based on individual treatment effect estimation methods that further reduce algorithmic discrimination using this data.", "date": "2025-07-02", "categories": []}
{"id": "2507.01418", "title": "Penalizing Transparency? How AI Disclosure and Author Demographics Shape Human and AI Judgments About Writing", "abstract": "As AI integrates in various types of human writing, calls for transparency around AI assistance are growing. However, if transparency operates on uneven ground and certain identity groups bear a heavier cost for being honest, then the burden of openness becomes asymmetrical. This study investigates how AI disclosure statement affects perceptions of writing quality, and whether these effects vary by the author's race and gender. Through a large-scale controlled experiment, both human raters (n = 1,970) and LLM raters (n = 2,520) evaluated a single human-written news article while disclosure statements and author demographics were systematically varied. This approach reflects how both human and algorithmic decisions now influence access to opportunities (e.g., hiring, promotion) and social recognition (e.g., content recommendation algorithms). We find that both human and LLM raters consistently penalize disclosed AI use. However, only LLM raters exhibit demographic interaction effects: they favor articles attributed to women or Black authors when no disclosure is present. But these advantages disappear when AI assistance is revealed. These findings illuminate the complex relationships between AI disclosure and author identity, highlighting disparities between machine and human evaluation patterns.", "date": "2025-07-02", "categories": []}
{"id": "2310.03647", "title": "Rethinking Algorithmic Fairness for Human-AI Collaboration", "abstract": "Existing approaches to algorithmic fairness aim to ensure equitable outcomes if human decision-makers comply perfectly with algorithmic decisions. However, perfect compliance with the algorithm is rarely a reality or even a desirable outcome in human-AI collaboration. Yet, recent studies have shown that selective compliance with fair algorithms can amplify discrimination relative to the prior human policy. As a consequence, ensuring equitable outcomes requires fundamentally different algorithmic design principles that ensure robustness to the decision-maker's (a priori unknown) compliance pattern. We define the notion of compliance-robustly fair algorithmic recommendations that are guaranteed to (weakly) improve fairness in decisions, regardless of the human's compliance pattern. We propose a simple optimization strategy to identify the best performance-improving compliance-robustly fair policy. However, we show that it may be infeasible to design algorithmic recommendations that are simultaneously fair in isolation, compliance-robustly fair, and more accurate than the human policy; thus, if our goal is to improve the equity and accuracy of human-AI collaboration, it may not be desirable to enforce traditional algorithmic fairness constraints. We illustrate the value of our approach on criminal sentencing data before and after the introduction of an algorithmic risk assessment tool in Virginia.", "date": "2023-10-05", "categories": []}
{"id": "2506.23111", "title": "FairI Tales: Evaluation of Fairness in Indian Contexts with a Focus on Bias and Stereotypes", "abstract": "Existing studies on fairness are largely Western-focused, making them inadequate for culturally diverse countries such as India. To address this gap, we introduce INDIC-BIAS, a comprehensive India-centric benchmark designed to evaluate fairness of LLMs across 85 identity groups encompassing diverse castes, religions, regions, and tribes. We first consult domain experts to curate over 1,800 socio-cultural topics spanning behaviors and situations, where biases and stereotypes are likely to emerge. Grounded in these topics, we generate and manually validate 20,000 real-world scenario templates to probe LLMs for fairness. We structure these templates into three evaluation tasks: plausibility, judgment, and generation. Our evaluation of 14 popular LLMs on these tasks reveals strong negative biases against marginalized identities, with models frequently reinforcing common stereotypes. Additionally, we find that models struggle to mitigate bias even when explicitly asked to rationalize their decision. Our evaluation provides evidence of both allocative and representational harms that current LLMs could cause towards Indian identities, calling for a more cautious usage in practical applications. We release INDIC-BIAS as an open-source benchmark to advance research on benchmarking and mitigating biases and stereotypes in the Indian context.", "date": "2025-06-29", "categories": []}
{"id": "2309.13933", "title": "Fairness and Bias in Algorithmic Hiring: a Multidisciplinary Survey", "abstract": "Employers are adopting algorithmic hiring technology throughout the recruitment pipeline. Algorithmic fairness is especially applicable in this domain due to its high stakes and structural inequalities. Unfortunately, most work in this space provides partial treatment, often constrained by two competing narratives, optimistically focused on replacing biased recruiter decisions or pessimistically pointing to the automation of discrimination. Whether, and more importantly what types of, algorithmic hiring can be less biased and more beneficial to society than low-tech alternatives currently remains unanswered, to the detriment of trustworthiness. This multidisciplinary survey caters to practitioners and researchers with a balanced and integrated coverage of systems, biases, measures, mitigation strategies, datasets, and legal aspects of algorithmic hiring and fairness. Our work supports a contextualized understanding and governance of this technology by highlighting current opportunities and limitations, providing recommendations for future work to ensure shared benefits for all stakeholders.", "date": "2023-09-25", "categories": []}
{"id": "2506.21574", "title": "Digital Gatekeepers: Exploring Large Language Model's Role in Immigration Decisions", "abstract": "With globalization and increasing immigrant populations, immigration departments face significant work-loads and the challenge of ensuring fairness in decision-making processes. Integrating artificial intelligence offers a promising solution to these challenges. This study investigates the potential of large language models (LLMs),such as GPT-3.5 and GPT-4, in supporting immigration decision-making. Utilizing a mixed-methods approach,this paper conducted discrete choice experiments and in-depth interviews to study LLM decision-making strategies and whether they are fair. Our findings demonstrate that LLMs can align their decision-making with human strategies, emphasizing utility maximization and procedural fairness. Meanwhile, this paper also reveals that while ChatGPT has safeguards to prevent unintentional discrimination, it still exhibits stereotypes and biases concerning nationality and shows preferences toward privileged group. This dual analysis highlights both the potential and limitations of LLMs in automating and enhancing immigration decisions.", "date": "2025-06-15", "categories": []}
{"id": "2506.22372", "title": "Towards Fair Rankings: Leveraging LLMs for Gender Bias Detection and Measurement", "abstract": "The presence of social biases in Natural Language Processing (NLP) and Information Retrieval (IR) systems is an ongoing challenge, which underlines the importance of developing robust approaches to identifying and evaluating such biases. In this paper, we aim to address this issue by leveraging Large Language Models (LLMs) to detect and measure gender bias in passage ranking. Existing gender fairness metrics rely on lexical- and frequency-based measures, leading to various limitations, e.g., missing subtle gender disparities. Building on our LLM-based gender bias detection method, we introduce a novel gender fairness metric, named Class-wise Weighted Exposure (CWEx), aiming to address existing limitations. To measure the effectiveness of our proposed metric and study LLMs' effectiveness in detecting gender bias, we annotate a subset of the MS MARCO Passage Ranking collection and release our new gender bias collection, called MSMGenderBias, to foster future research in this area. Our extensive experimental results on various ranking models show that our proposed metric offers a more detailed evaluation of fairness compared to previous metrics, with improved alignment to human labels (58.77% for Grep-BiasIR, and 18.51% for MSMGenderBias, measured using Cohen's Kappa agreement), effectively distinguishing gender bias in ranking. By integrating LLM-driven bias detection, an improved fairness metric, and gender bias annotations for an established dataset, this work provides a more robust framework for analyzing and mitigating bias in IR systems.", "date": "2025-06-27", "categories": []}
{"id": "2403.17852", "title": "Counterfactual Fairness through Transforming Data Orthogonal to Bias", "abstract": "Machine learning models have shown exceptional prowess in solving complex issues across various domains. However, these models can sometimes exhibit biased decision-making, resulting in unequal treatment of different groups. Despite substantial research on counterfactual fairness, methods to reduce the impact of multivariate and continuous sensitive variables on decision-making outcomes are still underdeveloped. We propose a novel data pre-processing algorithm, Orthogonal to Bias (OB), which is designed to eliminate the influence of a group of continuous sensitive variables, thus promoting counterfactual fairness in machine learning applications. Our approach, based on the assumption of a jointly normal distribution within a structural causal model (SCM), demonstrates that counterfactual fairness can be achieved by ensuring the data is orthogonal to the observed sensitive variables. The OB algorithm is model-agnostic, making it applicable to a wide range of machine learning models and tasks. Additionally, it includes a sparse variant to improve numerical stability through regularization. Empirical evaluations on both simulated and real-world datasets, encompassing settings with both discrete and continuous sensitive variables, show that our methodology effectively promotes fairer outcomes without compromising accuracy.", "date": "2024-03-26", "categories": []}
{"id": "2506.20025", "title": "Thumb on the Scale: Optimal Loss Weighting in Last Layer Retraining", "abstract": "While machine learning models become more capable in discriminative tasks at scale, their ability to overcome biases introduced by training data has come under increasing scrutiny. Previous results suggest that there are two extremes of parameterization with very different behaviors: the population (underparameterized) setting where loss weighting is optimal and the separable overparameterized setting where loss weighting is ineffective at ensuring equal performance across classes. This work explores the regime of last layer retraining (LLR) in which the unseen limited (retraining) data is frequently inseparable and the model proportionately sized, falling between the two aforementioned extremes. We show, in theory and practice, that loss weighting is still effective in this regime, but that these weights \\emph{must} take into account the relative overparameterization of the model.", "date": "2025-06-24", "categories": []}
{"id": "2506.20329", "title": "Producer-Fairness in Sequential Bundle Recommendation", "abstract": "We address fairness in the context of sequential bundle recommendation, where users are served in turn with sets of relevant and compatible items. Motivated by real-world scenarios, we formalize producer-fairness, that seeks to achieve desired exposure of different item groups across users in a recommendation session. Our formulation combines naturally with building high quality bundles. Our problem is solved in real time as users arrive. We propose an exact solution that caters to small instances of our problem. We then examine two heuristics, quality-first and fairness-first, and an adaptive variant that determines on-the-fly the right balance between bundle fairness and quality. Our experiments on three real-world datasets underscore the strengths and limitations of each solution and demonstrate their efficacy in providing fair bundle recommendations without compromising bundle quality.", "date": "2025-06-25", "categories": []}
{"id": "2502.02072", "title": "ASCenD-BDS: Adaptable, Stochastic and Context-aware framework for Detection of Bias, Discrimination and Stereotyping", "abstract": "The rapid evolution of Large Language Models (LLMs) has transformed natural language processing but raises critical concerns about biases inherent in their deployment and use across diverse linguistic and sociocultural contexts. This paper presents a framework named ASCenD BDS (Adaptable, Stochastic and Context-aware framework for Detection of Bias, Discrimination and Stereotyping). The framework presents approach to detecting bias, discrimination, stereotyping across various categories such as gender, caste, age, disability, socioeconomic status, linguistic variations, etc., using an approach which is Adaptive, Stochastic and Context-Aware. The existing frameworks rely heavily on usage of datasets to generate scenarios for detection of Bias, Discrimination and Stereotyping. Examples include datasets such as Civil Comments, Wino Gender, WinoBias, BOLD, CrowS Pairs and BBQ. However, such an approach provides point solutions. As a result, these datasets provide a finite number of scenarios for assessment. The current framework overcomes this limitation by having features which enable Adaptability, Stochasticity, Context Awareness. Context awareness can be customized for any nation or culture or sub-culture (for example an organization's unique culture). In this paper, context awareness in the Indian context has been established. Content has been leveraged from Indian Census 2011 to have a commonality of categorization. A framework has been developed using Category, Sub-Category, STEM, X-Factor, Synonym to enable the features for Adaptability, Stochasticity and Context awareness. The framework has been described in detail in Section 3. Overall 800 plus STEMs, 10 Categories, 31 unique SubCategories were developed by a team of consultants at Saint Fox Consultancy Private Ltd. The concept has been tested out in SFCLabs as part of product development.", "date": "2025-02-04", "categories": []}
{"id": "2506.17787", "title": "Incorporating Rather Than Eliminating: Achieving Fairness for Skin Disease Diagnosis Through Group-Specific Expert", "abstract": "AI-based systems have achieved high accuracy in skin disease diagnostics but often exhibit biases across demographic groups, leading to inequitable healthcare outcomes and diminished patient trust. Most existing bias mitigation methods attempt to eliminate the correlation between sensitive attributes and diagnostic prediction, but those methods often degrade performance due to the lost of clinically relevant diagnostic cues. In this work, we propose an alternative approach that incorporates sensitive attributes to achieve fairness. We introduce FairMoE, a framework that employs layer-wise mixture-of-experts modules to serve as group-specific learners. Unlike traditional methods that rigidly assign data based on group labels, FairMoE dynamically routes data to the most suitable expert, making it particularly effective for handling cases near group boundaries. Experimental results show that, unlike previous fairness approaches that reduce performance, FairMoE achieves substantial accuracy improvements while preserving comparable fairness metrics.", "date": "2025-06-21", "categories": []}
{"id": "2506.18327", "title": "Bias vs Bias -- Dawn of Justice: A Fair Fight in Recommendation Systems", "abstract": "Recommendation systems play a crucial role in our daily lives by impacting user experience across various domains, including e-commerce, job advertisements, entertainment, etc. Given the vital role of such systems in our lives, practitioners must ensure they do not produce unfair and imbalanced recommendations. Previous work addressing bias in recommendations overlooked bias in certain item categories, potentially leaving some biases unaddressed. Additionally, most previous work on fair re-ranking focused on binary-sensitive attributes. In this paper, we address these issues by proposing a fairness-aware re-ranking approach that helps mitigate bias in different categories of items. This re-ranking approach leverages existing biases to correct disparities in recommendations across various demographic groups. We show how our approach can mitigate bias on multiple sensitive attributes, including gender, age, and occupation. We experimented on three real-world datasets to evaluate the effectiveness of our re-ranking scheme in mitigating bias in recommendations. Our results show how this approach helps mitigate social bias with little to no degradation in performance.", "date": "2025-06-23", "categories": []}
{"id": "2506.18598", "title": "No Training Wheels: Steering Vectors for Bias Correction at Inference Time", "abstract": "Neural network classifiers trained on datasets with uneven group representation often inherit class biases and learn spurious correlations. These models may perform well on average but consistently fail on atypical groups. For example, in hair color classification, datasets may over-represent females with blond hair, reinforcing stereotypes. Although various algorithmic and data-centric methods have been proposed to address such biases, they often require retraining or significant compute. In this work, we propose a cheap, training-free method inspired by steering vectors used to edit behaviors in large language models. We compute the difference in mean activations between majority and minority groups to define a \"bias vector,\" which we subtract from the model's residual stream. This leads to reduced classification bias and improved worst-group accuracy. We explore multiple strategies for extracting and applying these vectors in transformer-like classifiers, showing that steering vectors, traditionally used in generative models, can also be effective in classification. More broadly, we showcase an extremely cheap, inference time, training free method to mitigate bias in classification models.", "date": "2025-06-23", "categories": []}
{"id": "2506.18696", "title": "SaGIF: Improving Individual Fairness in Graph Neural Networks via Similarity Encoding", "abstract": "Individual fairness (IF) in graph neural networks (GNNs), which emphasizes the need for similar individuals should receive similar outcomes from GNNs, has been a critical issue. Despite its importance, research in this area has been largely unexplored in terms of (1) a clear understanding of what induces individual unfairness in GNNs and (2) a comprehensive consideration of identifying similar individuals. To bridge these gaps, we conduct a preliminary analysis to explore the underlying reason for individual unfairness and observe correlations between IF and similarity consistency, a concept introduced to evaluate the discrepancy in identifying similar individuals based on graph structure versus node features. Inspired by our observations, we introduce two metrics to assess individual similarity from two distinct perspectives: topology fusion and feature fusion. Building upon these metrics, we propose Similarity-aware GNNs for Individual Fairness, named SaGIF. The key insight behind SaGIF is the integration of individual similarities by independently learning similarity representations, leading to an improvement of IF in GNNs. Our experiments on several real-world datasets validate the effectiveness of our proposed metrics and SaGIF. Specifically, SaGIF consistently outperforms state-of-the-art IF methods while maintaining utility performance. Code is available at: https://github.com/ZzoomD/SaGIF.", "date": "2025-06-23", "categories": []}
{"id": "2310.04585", "title": "Interventions Against Machine-Assisted Statistical Discrimination", "abstract": "I study statistical discrimination driven by verifiable beliefs, such as those generated by machine learning, rather than by humans. When beliefs are verifiable, interventions against statistical discrimination can move beyond simple, belief-free designs like affirmative action, to more sophisticated ones, that constrain decision makers based on what they are thinking. I design a belief-contingent intervention I call common identity. I show that it is effective at eliminating equilibrium statistical discrimination, even when training data exhibit the various statistical biases that often plague algorithmic decision problems.", "date": "2023-10-06", "categories": []}
{"id": "2506.17035", "title": "Critical Appraisal of Fairness Metrics in Clinical Predictive AI", "abstract": "Predictive artificial intelligence (AI) offers an opportunity to improve clinical practice and patient outcomes, but risks perpetuating biases if fairness is inadequately addressed. However, the definition of \"fairness\" remains unclear. We conducted a scoping review to identify and critically appraise fairness metrics for clinical predictive AI. We defined a \"fairness metric\" as a measure quantifying whether a model discriminates (societally) against individuals or groups defined by sensitive attributes. We searched five databases (2014-2024), screening 820 records, to include 41 studies, and extracted 62 fairness metrics. Metrics were classified by performance-dependency, model output level, and base performance metric, revealing a fragmented landscape with limited clinical validation and overreliance on threshold-dependent measures. Eighteen metrics were explicitly developed for healthcare, including only one clinical utility metric. Our findings highlight conceptual challenges in defining and quantifying fairness and identify gaps in uncertainty quantification, intersectionality, and real-world applicability. Future work should prioritise clinically meaningful metrics.", "date": "2025-06-20", "categories": []}
{"id": "2506.17111", "title": "Are Bias Evaluation Methods Biased ?", "abstract": "The creation of benchmarks to evaluate the safety of Large Language Models is one of the key activities within the trusted AI community. These benchmarks allow models to be compared for different aspects of safety such as toxicity, bias, harmful behavior etc. Independent benchmarks adopt different approaches with distinct data sets and evaluation methods. We investigate how robust such benchmarks are by using different approaches to rank a set of representative models for bias and compare how similar are the overall rankings. We show that different but widely used bias evaluations methods result in disparate model rankings. We conclude with recommendations for the community in the usage of such benchmarks.", "date": "2025-06-20", "categories": []}
{"id": "2506.15568", "title": "Gender Inclusivity Fairness Index (GIFI): A Multilevel Framework for Evaluating Gender Diversity in Large Language Models", "abstract": "We present a comprehensive evaluation of gender fairness in large language models (LLMs), focusing on their ability to handle both binary and non-binary genders. While previous studies primarily focus on binary gender distinctions, we introduce the Gender Inclusivity Fairness Index (GIFI), a novel and comprehensive metric that quantifies the diverse gender inclusivity of LLMs. GIFI consists of a wide range of evaluations at different levels, from simply probing the model with respect to provided gender pronouns to testing various aspects of model generation and cognitive behaviors under different gender assumptions, revealing biases associated with varying gender identifiers. We conduct extensive evaluations with GIFI on 22 prominent open-source and proprietary LLMs of varying sizes and capabilities, discovering significant variations in LLMs' gender inclusivity. Our study highlights the importance of improving LLMs' inclusivity, providing a critical benchmark for future advancements in gender fairness in generative models.", "date": "2025-06-18", "categories": []}
{"id": "2506.15620", "title": "GFLC: Graph-based Fairness-aware Label Correction for Fair Classification", "abstract": "Fairness in machine learning (ML) has a critical importance for building trustworthy machine learning system as artificial intelligence (AI) systems increasingly impact various aspects of society, including healthcare decisions and legal judgments. Moreover, numerous studies demonstrate evidence of unfair outcomes in ML and the need for more robust fairness-aware methods. However, the data we use to train and develop debiasing techniques often contains biased and noisy labels. As a result, the label bias in the training data affects model performance and misrepresents the fairness of classifiers during testing. To tackle this problem, our paper presents Graph-based Fairness-aware Label Correction (GFLC), an efficient method for correcting label noise while preserving demographic parity in datasets. In particular, our approach combines three key components: prediction confidence measure, graph-based regularization through Ricci-flow-optimized graph Laplacians, and explicit demographic parity incentives. Our experimental findings show the effectiveness of our proposed approach and show significant improvements in the trade-off between performance and fairness metrics compared to the baseline.", "date": "2025-06-18", "categories": []}
{"id": "2506.08673", "title": "Towards Fair Representation: Clustering and Consensus", "abstract": "Consensus clustering, a fundamental task in machine learning and data analysis, aims to aggregate multiple input clusterings of a dataset, potentially based on different non-sensitive attributes, into a single clustering that best represents the collective structure of the data. In this work, we study this fundamental problem through the lens of fair clustering, as introduced by Chierichetti et al. [NeurIPS'17], which incorporates the disparate impact doctrine to ensure proportional representation of each protected group in the dataset within every cluster. Our objective is to find a consensus clustering that is not only representative but also fair with respect to specific protected attributes. To the best of our knowledge, we are the first to address this problem and provide a constant-factor approximation.\n  As part of our investigation, we examine how to minimally modify an existing clustering to enforce fairness -- an essential postprocessing step in many clustering applications that require fair representation. We develop an optimal algorithm for datasets with equal group representation and near-linear time constant factor approximation algorithms for more general scenarios with different proportions of two group sizes. We complement our approximation result by showing that the problem is NP-hard for two unequal-sized groups. Given the fundamental nature of this problem, we believe our results on Closest Fair Clustering could have broader implications for other clustering problems, particularly those for which no prior approximation guarantees exist for their fair variants.", "date": "2025-06-10", "categories": []}
{"id": "2506.13780", "title": "Hidden Bias in the Machine: Stereotypes in Text-to-Image Models", "abstract": "Text-to-Image (T2I) models have transformed visual content creation, producing highly realistic images from natural language prompts. However, concerns persist around their potential to replicate and magnify existing societal biases. To investigate these issues, we curated a diverse set of prompts spanning thematic categories such as occupations, traits, actions, ideologies, emotions, family roles, place descriptions, spirituality, and life events. For each of the 160 unique topics, we crafted multiple prompt variations to reflect a wide range of meanings and perspectives. Using Stable Diffusion 1.5 (UNet-based) and Flux-1 (DiT-based) models with original checkpoints, we generated over 16,000 images under consistent settings. Additionally, we collected 8,000 comparison images from Google Image Search. All outputs were filtered to exclude abstract, distorted, or nonsensical results. Our analysis reveals significant disparities in the representation of gender, race, age, somatotype, and other human-centric factors across generated images. These disparities often mirror and reinforce harmful stereotypes embedded in societal narratives. We discuss the implications of these findings and emphasize the need for more inclusive datasets and development practices to foster fairness in generative visual systems.", "date": "2025-06-09", "categories": []}
{"id": "2506.13947", "title": "Meta Optimality for Demographic Parity Constrained Regression via Post-Processing", "abstract": "We address the regression problem under the constraint of demographic parity, a commonly used fairness definition. Recent studies have revealed fair minimax optimal regression algorithms, the most accurate algorithms that adhere to the fairness constraint. However, these analyses are tightly coupled with specific data generation models. In this paper, we provide meta-theorems that can be applied to various situations to validate the fair minimax optimality of the corresponding regression algorithms. Furthermore, we demonstrate that fair minimax optimal regression can be achieved through post-processing methods, allowing researchers and practitioners to focus on improving conventional regression techniques, which can then be efficiently adapted for fair regression.", "date": "2025-06-16", "categories": []}
{"id": "2506.14306", "title": "Fair for a few: Improving Fairness in Doubly Imbalanced Datasets", "abstract": "Fairness has been identified as an important aspect of Machine Learning and Artificial Intelligence solutions for decision making. Recent literature offers a variety of approaches for debiasing, however many of them fall short when the data collection is imbalanced. In this paper, we focus on a particular case, fairness in doubly imbalanced datasets, such that the data collection is imbalanced both for the label and the groups in the sensitive attribute. Firstly, we present an exploratory analysis to illustrate limitations in debiasing on a doubly imbalanced dataset. Then, a multi-criteria based solution is proposed for finding the most suitable sampling and distribution for label and sensitive attribute, in terms of fairness and classification accuracy", "date": "2025-06-17", "categories": []}
{"id": "2402.08879", "title": "Inference for an Algorithmic Fairness-Accuracy Frontier", "abstract": "Algorithms are increasingly used to aid with high-stakes decision making. Yet, their predictive ability frequently exhibits systematic variation across population subgroups. To assess the trade-off between fairness and accuracy using finite data, we propose a debiased machine learning estimator for the fairness-accuracy frontier introduced by Liang, Lu, Mu, and Okumura (2024). We derive its asymptotic distribution and propose inference methods to test key hypotheses in the fairness literature, such as (i) whether excluding group identity from use in training the algorithm is optimal and (ii) whether there are less discriminatory alternatives to a given algorithm. In addition, we construct an estimator for the distance between a given algorithm and the fairest point on the frontier, and characterize its asymptotic distribution. Using Monte Carlo simulations, we evaluate the finite-sample performance of our inference methods. We apply our framework to re-evaluate algorithms used in hospital care management and show that our approach yields alternative algorithms that lie on the fairness-accuracy frontier, offering improvements along both dimensions.", "date": "2024-02-14", "categories": []}
{"id": "2506.12502", "title": "Towards Fairness Assessment of Dutch Hate Speech Detection", "abstract": "Numerous studies have proposed computational methods to detect hate speech online, yet most focus on the English language and emphasize model development. In this study, we evaluate the counterfactual fairness of hate speech detection models in the Dutch language, specifically examining the performance and fairness of transformer-based models. We make the following key contributions. First, we curate a list of Dutch Social Group Terms that reflect social context. Second, we generate counterfactual data for Dutch hate speech using LLMs and established strategies like Manual Group Substitution (MGS) and Sentence Log-Likelihood (SLL). Through qualitative evaluation, we highlight the challenges of generating realistic counterfactuals, particularly with Dutch grammar and contextual coherence. Third, we fine-tune baseline transformer-based models with counterfactual data and evaluate their performance in detecting hate speech. Fourth, we assess the fairness of these models using Counterfactual Token Fairness (CTF) and group fairness metrics, including equality of odds and demographic parity. Our analysis shows that models perform better in terms of hate speech detection, average counterfactual fairness and group fairness. This work addresses a significant gap in the literature on counterfactual fairness for hate speech detection in Dutch and provides practical insights and recommendations for improving both model performance and fairness.", "date": "2025-06-14", "categories": []}
{"id": "2506.12574", "title": "Overview of the NLPCC 2025 Shared Task: Gender Bias Mitigation Challenge", "abstract": "As natural language processing for gender bias becomes a significant interdisciplinary topic, the prevalent data-driven techniques, such as pre-trained language models, suffer from biased corpus. This case becomes more obvious regarding those languages with less fairness-related computational linguistic resources, such as Chinese. To this end, we propose a Chinese cOrpus foR Gender bIas Probing and Mitigation (CORGI-PM), which contains 32.9k sentences with high-quality labels derived by following an annotation scheme specifically developed for gender bias in the Chinese context. It is worth noting that CORGI-PM contains 5.2k gender-biased sentences along with the corresponding bias-eliminated versions rewritten by human annotators. We pose three challenges as a shared task to automate the mitigation of textual gender bias, which requires the models to detect, classify, and mitigate textual gender bias. In the literature, we present the results and analysis for the teams participating this shared task in NLPCC 2025.", "date": "2025-06-14", "categories": []}
{"id": "2506.12839", "title": "Fair Bayesian Model-Based Clustering", "abstract": "Fair clustering has become a socially significant task with the advancement of machine learning technologies and the growing demand for trustworthy AI. Group fairness ensures that the proportions of each sensitive group are similar in all clusters. Most existing group-fair clustering methods are based on the $K$-means clustering and thus require the distance between instances and the number of clusters to be given in advance. To resolve this limitation, we propose a fair Bayesian model-based clustering called Fair Bayesian Clustering (FBC). We develop a specially designed prior which puts its mass only on fair clusters, and implement an efficient MCMC algorithm. Advantages of FBC are that it can infer the number of clusters and can be applied to any data type as long as the likelihood is defined (e.g., categorical data). Experiments on real-world datasets show that FBC (i) reasonably infers the number of clusters, (ii) achieves a competitive utility-fairness trade-off compared to existing fair clustering methods, and (iii) performs well on categorical data.", "date": "2025-06-15", "categories": []}
{"id": "2505.01591", "title": "Machine Learning Fairness in House Price Prediction: A Case Study of America's Expanding Metropolises", "abstract": "As a basic human need, housing plays a key role in enhancing health, well-being, and educational outcome in society, and the housing market is a major factor for promoting quality of life and ensuring social equity. To improve the housing conditions, there has been extensive research on building Machine Learning (ML)-driven house price prediction solutions to accurately forecast the future conditions, and help inform actions and policies in the field. In spite of their success in developing high-accuracy models, there is a gap in our understanding of the extent to which various ML-driven house price prediction approaches show ethnic and/or racial bias, which in turn is essential for the responsible use of ML, and ensuring that the ML-driven solutions do not exacerbate inequity. To fill this gap, this paper develops several ML models from a combination of structural and neighborhood-level attributes, and conducts comprehensive assessments on the fairness of ML models under various definitions of privileged groups. As a result, it finds that the ML-driven house price prediction models show various levels of bias towards protected attributes (i.e., race and ethnicity in this study). Then, it investigates the performance of different bias mitigation solutions, and the experimental results show their various levels of effectiveness on different ML-driven methods. However, in general, the in-processing bias mitigation approach tends to be more effective than the pre-processing one in this problem domain. Our code is available at https://github.com/wahab1412/housing_fairness.", "date": "2025-05-02", "categories": []}
{"id": "2506.11047", "title": "Perception-Driven Bias Detection in Machine Learning via Crowdsourced Visual Judgment", "abstract": "Machine learning systems are increasingly deployed in high-stakes domains, yet they remain vulnerable to bias systematic disparities that disproportionately impact specific demographic groups. Traditional bias detection methods often depend on access to sensitive labels or rely on rigid fairness metrics, limiting their applicability in real-world settings. This paper introduces a novel, perception-driven framework for bias detection that leverages crowdsourced human judgment. Inspired by reCAPTCHA and other crowd-powered systems, we present a lightweight web platform that displays stripped-down visualizations of numeric data (for example-salary distributions across demographic clusters) and collects binary judgments on group similarity. We explore how users' visual perception-shaped by layout, spacing, and question phrasing can signal potential disparities. User feedback is aggregated to flag data segments as biased, which are then validated through statistical tests and machine learning cross-evaluations. Our findings show that perceptual signals from non-expert users reliably correlate with known bias cases, suggesting that visual intuition can serve as a powerful, scalable proxy for fairness auditing. This approach offers a label-efficient, interpretable alternative to conventional fairness diagnostics, paving the way toward human-aligned, crowdsourced bias detection pipelines.", "date": "2025-05-21", "categories": []}
{"id": "2506.11132", "title": "Gender Fairness of Machine Learning Algorithms for Pain Detection", "abstract": "Automated pain detection through machine learning (ML) and deep learning (DL) algorithms holds significant potential in healthcare, particularly for patients unable to self-report pain levels. However, the accuracy and fairness of these algorithms across different demographic groups (e.g., gender) remain under-researched. This paper investigates the gender fairness of ML and DL models trained on the UNBC-McMaster Shoulder Pain Expression Archive Database, evaluating the performance of various models in detecting pain based solely on the visual modality of participants' facial expressions. We compare traditional ML algorithms, Linear Support Vector Machine (L SVM) and Radial Basis Function SVM (RBF SVM), with DL methods, Convolutional Neural Network (CNN) and Vision Transformer (ViT), using a range of performance and fairness metrics. While ViT achieved the highest accuracy and a selection of fairness metrics, all models exhibited gender-based biases. These findings highlight the persistent trade-off between accuracy and fairness, emphasising the need for fairness-aware techniques to mitigate biases in automated healthcare systems.", "date": "2025-06-10", "categories": []}
{"id": "2506.11242", "title": "A Causal Lens for Learning Long-term Fair Policies", "abstract": "Fairness-aware learning studies the development of algorithms that avoid discriminatory decision outcomes despite biased training data. While most studies have concentrated on immediate bias in static contexts, this paper highlights the importance of investigating long-term fairness in dynamic decision-making systems while simultaneously considering instantaneous fairness requirements. In the context of reinforcement learning, we propose a general framework where long-term fairness is measured by the difference in the average expected qualification gain that individuals from different groups could obtain.Then, through a causal lens, we decompose this metric into three components that represent the direct impact, the delayed impact, as well as the spurious effect the policy has on the qualification gain. We analyze the intrinsic connection between these components and an emerging fairness notion called benefit fairness that aims to control the equity of outcomes in decision-making. Finally, we develop a simple yet effective approach for balancing various fairness notions.", "date": "2025-06-12", "categories": []}
{"id": "2506.11627", "title": "Evaluating Fairness and Mitigating Bias in Machine Learning: A Novel Technique using Tensor Data and Bayesian Regression", "abstract": "Fairness is a critical component of Trustworthy AI. In this paper, we focus on Machine Learning (ML) and the performance of model predictions when dealing with skin color. Unlike other sensitive attributes, the nature of skin color differs significantly. In computer vision, skin color is represented as tensor data rather than categorical values or single numerical points. However, much of the research on fairness across sensitive groups has focused on categorical features such as gender and race. This paper introduces a new technique for evaluating fairness in ML for image classification tasks, specifically without the use of annotation. To address the limitations of prior work, we handle tensor data, like skin color, without classifying it rigidly. Instead, we convert it into probability distributions and apply statistical distance measures. This novel approach allows us to capture fine-grained nuances in fairness both within and across what would traditionally be considered distinct groups. Additionally, we propose an innovative training method to mitigate the latent biases present in conventional skin tone categorization. This method leverages color distance estimates calculated through Bayesian regression with polynomial functions, ensuring a more nuanced and equitable treatment of skin color in ML models.", "date": "2025-06-13", "categories": []}
{"id": "2506.11880", "title": "Addressing Bias in LLMs: Strategies and Application to Fair AI-based Recruitment", "abstract": "The use of language technologies in high-stake settings is increasing in recent years, mostly motivated by the success of Large Language Models (LLMs). However, despite the great performance of LLMs, they are are susceptible to ethical concerns, such as demographic biases, accountability, or privacy. This work seeks to analyze the capacity of Transformers-based systems to learn demographic biases present in the data, using a case study on AI-based automated recruitment. We propose a privacy-enhancing framework to reduce gender information from the learning pipeline as a way to mitigate biased behaviors in the final tools. Our experiments analyze the influence of data biases on systems built on two different LLMs, and how the proposed framework effectively prevents trained systems from reproducing the bias in the data.", "date": "2025-06-13", "categories": []}
{"id": "2506.07479", "title": "Improving Fairness of Large Language Models in Multi-document Summarization", "abstract": "Fairness in multi-document summarization (MDS) is crucial for providing comprehensive views across documents with diverse social attribute values, which can significantly impact decision-making. For example, a summarization system that tends to overrepresent negative reviews of products can mislead customers into disregarding good products. Previous works measure fairness in MDS at two levels: summary-level and corpus-level. While summary-level fairness focuses on individual summaries, corpus-level fairness focuses on a corpus of summaries. Recent methods primarily focus on summary-level fairness. We propose FairPO, a preference tuning method that focuses on both summary-level and corpus-level fairness in MDS. To improve summary-level fairness, we propose to generate preference pairs by perturbing document sets. To improve corpus-level fairness, we propose fairness-aware preference tuning by dynamically adjusting the weights of preference pairs. Our experiments show that FairPO outperforms strong baselines while maintaining the critical qualities of summaries. The code is available at https://github.com/leehaoyuan/coverage_fairnes.", "date": "2025-06-09", "categories": []}
{"id": "2506.10922", "title": "Robustly Improving LLM Fairness in Realistic Settings via Interpretability", "abstract": "Large language models (LLMs) are increasingly deployed in high-stakes hiring applications, making decisions that directly impact people's careers and livelihoods. While prior studies suggest simple anti-bias prompts can eliminate demographic biases in controlled evaluations, we find these mitigations fail when realistic contextual details are introduced. We address these failures through internal bias mitigation: by identifying and neutralizing sensitive attribute directions within model activations, we achieve robust bias reduction across all tested scenarios. Across leading commercial (GPT-4o, Claude 4 Sonnet, Gemini 2.5 Flash) and open-source models (Gemma-2 27B, Gemma-3, Mistral-24B), we find that adding realistic context such as company names, culture descriptions from public careers pages, and selective hiring constraints (e.g.,``only accept candidates in the top 10\\%\") induces significant racial and gender biases (up to 12\\% differences in interview rates). When these biases emerge, they consistently favor Black over White candidates and female over male candidates across all tested models and scenarios. Moreover, models can infer demographics and become biased from subtle cues like college affiliations, with these biases remaining invisible even when inspecting the model's chain-of-thought reasoning. To address these limitations, our internal bias mitigation identifies race and gender-correlated directions and applies affine concept editing at inference time. Despite using directions from a simple synthetic dataset, the intervention generalizes robustly, consistently reducing bias to very low levels (typically under 1\\%, always below 2.5\\%) while largely maintaining model performance. Our findings suggest that practitioners deploying LLMs for hiring should adopt more realistic evaluation methodologies and consider internal mitigation strategies for equitable outcomes.", "date": "2025-06-12", "categories": []}
{"id": "2502.02221", "title": "Bias Detection via Maximum Subgroup Discrepancy", "abstract": "Bias evaluation is fundamental to trustworthy AI, both in terms of checking data quality and in terms of checking the outputs of AI systems. In testing data quality, for example, one may study the distance of a given dataset, viewed as a distribution, to a given ground-truth reference dataset. However, classical metrics, such as the Total Variation and the Wasserstein distances, are known to have high sample complexities and, therefore, may fail to provide a meaningful distinction in many practical scenarios.\n  In this paper, we propose a new notion of distance, the Maximum Subgroup Discrepancy (MSD). In this metric, two distributions are close if, roughly, discrepancies are low for all feature subgroups. While the number of subgroups may be exponential, we show that the sample complexity is linear in the number of features, thus making it feasible for practical applications. Moreover, we provide a practical algorithm for evaluating the distance based on Mixed-integer optimization (MIO). We also note that the proposed distance is easily interpretable, thus providing clearer paths to fixing the biases once they have been identified. Finally, we describe a natural general bias detection framework, termed MSDD distances, and show that MSD aligns well with this framework. We empirically evaluate MSD by comparing it with other metrics and by demonstrating the above properties of MSD on real-world datasets.", "date": "2025-02-04", "categories": []}
{"id": "2506.09106", "title": "Bias Analysis in Unconditional Image Generative Models", "abstract": "The widespread adoption of generative AI models has raised growing concerns about representational harm and potential discriminatory outcomes. Yet, despite growing literature on this topic, the mechanisms by which bias emerges - especially in unconditional generation - remain disentangled. We define the bias of an attribute as the difference between the probability of its presence in the observed distribution and its expected proportion in an ideal reference distribution. In our analysis, we train a set of unconditional image generative models and adopt a commonly used bias evaluation framework to study bias shift between training and generated distributions. Our experiments reveal that the detected attribute shifts are small. We find that the attribute shifts are sensitive to the attribute classifier used to label generated images in the evaluation framework, particularly when its decision boundaries fall in high-density regions. Our empirical analysis indicates that this classifier sensitivity is often observed in attributes values that lie on a spectrum, as opposed to exhibiting a binary nature. This highlights the need for more representative labeling practices, understanding the shortcomings through greater scrutiny of evaluation frameworks, and recognizing the socially complex nature of attributes when evaluating bias.", "date": "2025-06-10", "categories": []}
{"id": "2506.09107", "title": "FAIRTOPIA: Envisioning Multi-Agent Guardianship for Disrupting Unfair AI Pipelines", "abstract": "AI models have become active decision makers, often acting without human supervision. The rapid advancement of AI technology has already caused harmful incidents that have hurt individuals and societies and AI unfairness in heavily criticized. It is urgent to disrupt AI pipelines which largely neglect human principles and focus on computational biases exploration at the data (pre), model(in), and deployment (post) processing stages. We claim that by exploiting the advances of agents technology, we will introduce cautious, prompt, and ongoing fairness watch schemes, under realistic, systematic, and human-centric fairness expectations. We envision agents as fairness guardians, since agents learn from their environment, adapt to new information, and solve complex problems by interacting with external tools and other systems. To set the proper fairness guardrails in the overall AI pipeline, we introduce a fairness-by-design approach which embeds multi-role agents in an end-to-end (human to AI) synergetic scheme. Our position is that we may design adaptive and realistic AI fairness frameworks, and we introduce a generalized algorithm which can be customized to the requirements and goals of each AI decision making scenario. Our proposed, so called FAIRTOPIA framework, is structured over a three-layered architecture, which encapsulates the AI pipeline inside an agentic guardian and a knowledge-based, self-refining layered scheme. Based on our proposition, we enact fairness watch in all of the AI pipeline stages, under robust multi-agent workflows, which will inspire new fairness research hypothesis, heuristics, and methods grounded in human-centric, systematic, interdisciplinary, socio-technical principles.", "date": "2025-06-10", "categories": []}
{"id": "2410.19317", "title": "FairMT-Bench: Benchmarking Fairness for Multi-turn Dialogue in Conversational LLMs", "abstract": "The growing use of large language model (LLM)-based chatbots has raised concerns about fairness. Fairness issues in LLMs can lead to severe consequences, such as bias amplification, discrimination, and harm to marginalized communities. While existing fairness benchmarks mainly focus on single-turn dialogues, multi-turn scenarios, which in fact better reflect real-world conversations, present greater challenges due to conversational complexity and potential bias accumulation. In this paper, we propose a comprehensive fairness benchmark for LLMs in multi-turn dialogue scenarios, \\textbf{FairMT-Bench}. Specifically, we formulate a task taxonomy targeting LLM fairness capabilities across three stages: context understanding, user interaction, and instruction trade-offs, with each stage comprising two tasks. To ensure coverage of diverse bias types and attributes, we draw from existing fairness datasets and employ our template to construct a multi-turn dialogue dataset, \\texttt{FairMT-10K}. For evaluation, GPT-4 is applied, alongside bias classifiers including Llama-Guard-3 and human validation to ensure robustness. Experiments and analyses on \\texttt{FairMT-10K} reveal that in multi-turn dialogue scenarios, current LLMs are more likely to generate biased responses, and there is significant variation in performance across different tasks and models. Based on this, we curate a challenging dataset, \\texttt{FairMT-1K}, and test 15 current state-of-the-art (SOTA) LLMs on this dataset. The results show the current state of fairness in LLMs and showcase the utility of this novel approach for assessing fairness in more realistic multi-turn dialogue contexts, calling for future work to focus on LLM fairness improvement and the adoption of \\texttt{FairMT-1K} in such efforts.", "date": "2024-10-25", "categories": []}
{"id": "2506.08143", "title": "Accelerating Spectral Clustering under Fairness Constraints", "abstract": "Fairness of decision-making algorithms is an increasingly important issue. In this paper, we focus on spectral clustering with group fairness constraints, where every demographic group is represented in each cluster proportionally as in the general population. We present a new efficient method for fair spectral clustering (Fair SC) by casting the Fair SC problem within the difference of convex functions (DC) framework. To this end, we introduce a novel variable augmentation strategy and employ an alternating direction method of multipliers type of algorithm adapted to DC problems. We show that each associated subproblem can be solved efficiently, resulting in higher computational efficiency compared to prior work, which required a computationally expensive eigendecomposition. Numerical experiments demonstrate the effectiveness of our approach on both synthetic and real-world benchmarks, showing significant speedups in computation time over prior art, especially as the problem size grows. This work thus represents a considerable step forward towards the adoption of fair clustering in real-world applications.", "date": "2025-06-09", "categories": []}
{"id": "2404.05678", "title": "FairICP: Encouraging Equalized Odds via Inverse Conditional Permutation", "abstract": "$\\textit{Equalized odds}$, an important notion of algorithmic fairness, aims to ensure that sensitive variables, such as race and gender, do not unfairly influence the algorithm's prediction when conditioning on the true outcome. Despite rapid advancements, current research primarily focuses on equalized odds violations caused by a single sensitive attribute, leaving the challenge of simultaneously accounting for multiple attributes under-addressed. We bridge this gap by introducing an in-processing fairness-aware learning approach, FairICP, which integrates adversarial learning with a novel inverse conditional permutation scheme. FairICP offers a flexible and efficient scheme to promote equalized odds under fairness conditions described by complex and multi-dimensional sensitive attributes. The efficacy and adaptability of our method are demonstrated through both simulation studies and empirical analyses of real-world datasets.", "date": "2024-04-08", "categories": []}
{"id": "2406.13544", "title": "One Fits All: Learning Fair Graph Neural Networks for Various Sensitive Attributes", "abstract": "Recent studies have highlighted fairness issues in Graph Neural Networks (GNNs), where they produce discriminatory predictions against specific protected groups categorized by sensitive attributes such as race and age. While various efforts to enhance GNN fairness have made significant progress, these approaches are often tailored to specific sensitive attributes. Consequently, they necessitate retraining the model from scratch to accommodate changes in the sensitive attribute requirement, resulting in high computational costs. To gain deeper insights into this issue, we approach the graph fairness problem from a causal modeling perspective, where we identify the confounding effect induced by the sensitive attribute as the underlying reason. Motivated by this observation, we formulate the fairness problem in graphs from an invariant learning perspective, which aims to learn invariant representations across environments. Accordingly, we propose a graph fairness framework based on invariant learning, namely FairINV, which enables the training of fair GNNs to accommodate various sensitive attributes within a single training session. Specifically, FairINV incorporates sensitive attribute partition and trains fair GNNs by eliminating spurious correlations between the label and various sensitive attributes. Experimental results on several real-world datasets demonstrate that FairINV significantly outperforms state-of-the-art fairness approaches, underscoring its effectiveness. Our code is available via: https://github.com/ZzoomD/FairINV/.", "date": "2024-06-19", "categories": []}
{"id": "2502.01267", "title": "Counterfactual Situation Testing: From Single to Multidimensional Discrimination", "abstract": "We present counterfactual situation testing (CST), a causal data mining framework for detecting individual discrimination in a dataset of classifier decisions. CST answers the question ``what would have been the model outcome had the individual, or complainant, been of a different protected status?'' It extends the legally-grounded situation testing (ST) of Thanh et al. (2011) by operationalizing the notion of \"fairness given the difference\" via counterfactual reasoning. ST finds for each complainant similar protected and non-protected instances in the dataset; constructs, respectively, a control and test group; and compares the groups such that a difference in model outcomes implies a potential case of individual discrimination. CST, instead, avoids this idealized comparison by establishing the test group on the complainant's generated counterfactual, which reflects how the protected attribute when changed influences other seemingly neutral attributes of the complainant. Under CST we test for discrimination for each complainant by comparing similar individuals within the control and test group but dissimilar individuals across these groups. We consider single (e.g.,~gender) and multidimensional (e.g.,~gender and race) discrimination testing. For multidimensional discrimination we study multiple and intersectional discrimination and, as feared by legal scholars, find evidence that the former fails to account for the latter kind. Using a k-nearest neighbor implementation, we showcase CST on synthetic and real data. Experimental results show that CST uncovers a higher number of cases than ST, even when the model is counterfactually fair. CST, in fact, extends counterfactual fairness (CF) of Kusner et al. (2017) by equipping CF with confidence intervals, which we report for all experiments.", "date": "2025-02-03", "categories": []}
{"id": "2504.00928", "title": "Taxonomizing Representational Harms using Speech Act Theory", "abstract": "Representational harms are widely recognized among fairness-related harms caused by generative language systems. However, their definitions are commonly under-specified. We make a theoretical contribution to the specification of representational harms by introducing a framework, grounded in speech act theory (Austin, 1962), that conceptualizes representational harms caused by generative language systems as the perlocutionary effects (i.e., real-world impacts) of particular types of illocutionary acts (i.e., system behaviors). Building on this argument and drawing on relevant literature from linguistic anthropology and sociolinguistics, we provide new definitions of stereotyping, demeaning, and erasure. We then use our framework to develop a granular taxonomy of illocutionary acts that cause representational harms, going beyond the high-level taxonomies presented in previous work. We also discuss the ways that our framework and taxonomy can support the development of valid measurement instruments. Finally, we demonstrate the utility of our framework and taxonomy via a case study that engages with recent conceptual debates about what constitutes a representational harm and how such harms should be measured.", "date": "2025-04-01", "categories": []}
{"id": "2506.06506", "title": "Biases Propagate in Encoder-based Vision-Language Models: A Systematic Analysis From Intrinsic Measures to Zero-shot Retrieval Outcomes", "abstract": "To build fair AI systems we need to understand how social-group biases intrinsic to foundational encoder-based vision-language models (VLMs) manifest in biases in downstream tasks. In this study, we demonstrate that intrinsic biases in VLM representations systematically ``carry over'' or propagate into zero-shot retrieval tasks, revealing how deeply rooted biases shape a model's outputs. We introduce a controlled framework to measure this propagation by correlating (a) intrinsic measures of bias in the representational space with (b) extrinsic measures of bias in zero-shot text-to-image (TTI) and image-to-text (ITT) retrieval. Results show substantial correlations between intrinsic and extrinsic bias, with an average $\\rho$ = 0.83 $\\pm$ 0.10. This pattern is consistent across 114 analyses, both retrieval directions, six social groups, and three distinct VLMs. Notably, we find that larger/better-performing models exhibit greater bias propagation, a finding that raises concerns given the trend towards increasingly complex AI models. Our framework introduces baseline evaluation tasks to measure the propagation of group and valence signals. Investigations reveal that underrepresented groups experience less robust propagation, further skewing their model-related outcomes.", "date": "2025-06-06", "categories": []}
{"id": "2506.07049", "title": "FairPFN: A Tabular Foundation Model for Causal Fairness", "abstract": "Machine learning (ML) systems are utilized in critical sectors, such as healthcare, law enforcement, and finance. However, these systems are often trained on historical data that contains demographic biases, leading to ML decisions that perpetuate or exacerbate existing social inequalities. Causal fairness provides a transparent, human-in-the-loop framework to mitigate algorithmic discrimination, aligning closely with legal doctrines of direct and indirect discrimination. However, current causal fairness frameworks hold a key limitation in that they assume prior knowledge of the correct causal model, restricting their applicability in complex fairness scenarios where causal models are unknown or difficult to identify. To bridge this gap, we propose FairPFN, a tabular foundation model pre-trained on synthetic causal fairness data to identify and mitigate the causal effects of protected attributes in its predictions. FairPFN's key contribution is that it requires no knowledge of the causal model and still demonstrates strong performance in identifying and removing protected causal effects across a diverse set of hand-crafted and real-world scenarios relative to robust baseline methods. FairPFN paves the way for promising future research, making causal fairness more accessible to a wider variety of complex fairness problems.", "date": "2025-06-08", "categories": []}
{"id": "2506.07861", "title": "Fairness Overfitting in Machine Learning: An Information-Theoretic Perspective", "abstract": "Despite substantial progress in promoting fairness in high-stake applications using machine learning models, existing methods often modify the training process, such as through regularizers or other interventions, but lack formal guarantees that fairness achieved during training will generalize to unseen data. Although overfitting with respect to prediction performance has been extensively studied, overfitting in terms of fairness loss has received far less attention. This paper proposes a theoretical framework for analyzing fairness generalization error through an information-theoretic lens. Our novel bounding technique is based on Efron-Stein inequality, which allows us to derive tight information-theoretic fairness generalization bounds with both Mutual Information (MI) and Conditional Mutual Information (CMI). Our empirical results validate the tightness and practical relevance of these bounds across diverse fairness-aware learning algorithms. Our framework offers valuable insights to guide the design of algorithms improving fairness generalization.", "date": "2025-06-09", "categories": []}
{"id": "2410.22118", "title": "The Impact of Inference Acceleration on Bias of LLMs", "abstract": "Last few years have seen unprecedented advances in capabilities of Large Language Models (LLMs). These advancements promise to benefit a vast array of application domains. However, due to their immense size, performing inference with LLMs is both costly and slow. Consequently, a plethora of recent work has proposed strategies to enhance inference efficiency, e.g., quantization, pruning, and caching. These acceleration strategies reduce the inference cost and latency, often by several factors, while maintaining much of the predictive performance measured via common benchmarks. In this work, we explore another critical aspect of LLM performance: demographic bias in model generations due to inference acceleration optimizations. Using a wide range of metrics, we probe bias in model outputs from a number of angles. Analysis of outputs before and after inference acceleration shows significant change in bias. Worryingly, these bias effects are complex and unpredictable. A combination of an acceleration strategy and bias type may show little bias change in one model but may lead to a large effect in another. Our results highlight a need for in-depth and case-by-case evaluation of model bias after it has been modified to accelerate inference.", "date": "2024-10-29", "categories": []}
{"id": "2504.05632", "title": "Reasoning Towards Fairness: Mitigating Bias in Language Models through Reasoning-Guided Fine-Tuning", "abstract": "Recent advances in large-scale generative language models have shown that reasoning capabilities can significantly improve model performance across a variety of tasks. However, the impact of reasoning on a model's ability to mitigate stereotypical responses remains largely underexplored. In this work, we investigate the crucial relationship between a model's reasoning ability and fairness, and ask whether improved reasoning capabilities can mitigate harmful stereotypical responses, especially those arising due to shallow or flawed reasoning. We conduct a comprehensive evaluation of multiple open-source LLMs, and find that larger models with stronger reasoning abilities exhibit substantially lower stereotypical bias on existing fairness benchmarks. Building on this insight, we introduce ReGiFT -- Reasoning Guided Fine-Tuning, a novel approach that extracts structured reasoning traces from advanced reasoning models and infuses them into models that lack such capabilities. We use only general-purpose reasoning and do not require any fairness-specific supervision for bias mitigation. Notably, we see that models fine-tuned using ReGiFT not only improve fairness relative to their non-reasoning counterparts but also outperform advanced reasoning models on fairness benchmarks. We also analyze how variations in the correctness of the reasoning traces and their length influence model fairness and their overall performance. Our findings highlight that enhancing reasoning capabilities is an effective, fairness-agnostic strategy for mitigating stereotypical bias caused by reasoning flaws.", "date": "2025-04-08", "categories": []}
{"id": "2506.01532", "title": "Balancing Beyond Discrete Categories: Continuous Demographic Labels for Fair Face Recognition", "abstract": "Bias has been a constant in face recognition models. Over the years, researchers have looked at it from both the model and the data point of view. However, their approach to mitigation of data bias was limited and lacked insight on the real nature of the problem. Here, in this document, we propose to revise our use of ethnicity labels as a continuous variable instead of a discrete value per identity. We validate our formulation both experimentally and theoretically, showcasing that not all identities from one ethnicity contribute equally to the balance of the dataset; thus, having the same number of identities per ethnicity does not represent a balanced dataset. We further show that models trained on datasets balanced in the continuous space consistently outperform models trained on data balanced in the discrete space. We trained more than 65 different models, and created more than 20 subsets of the original datasets.", "date": "2025-06-02", "categories": []}
{"id": "2506.05390", "title": "Understanding Gender Bias in AI-Generated Product Descriptions", "abstract": "While gender bias in large language models (LLMs) has been extensively studied in many domains, uses of LLMs in e-commerce remain largely unexamined and may reveal novel forms of algorithmic bias and harm. Our work investigates this space, developing data-driven taxonomic categories of gender bias in the context of product description generation, which we situate with respect to existing general purpose harms taxonomies. We illustrate how AI-generated product descriptions can uniquely surface gender biases in ways that require specialized detection and mitigation approaches. Further, we quantitatively analyze issues corresponding to our taxonomic categories in two models used for this task -- GPT-3.5 and an e-commerce-specific LLM -- demonstrating that these forms of bias commonly occur in practice. Our results illuminate unique, under-explored dimensions of gender bias, such as assumptions about clothing size, stereotypical bias in which features of a product are advertised, and differences in the use of persuasive language. These insights contribute to our understanding of three types of AI harms identified by current frameworks: exclusionary norms, stereotyping, and performance disparities, particularly for the context of e-commerce.", "date": "2025-06-03", "categories": []}
{"id": "2406.03198", "title": "The Impossibility of Fair LLMs", "abstract": "The rise of general-purpose artificial intelligence (AI) systems, particularly large language models (LLMs), has raised pressing moral questions about how to reduce bias and ensure fairness at scale. Researchers have documented a sort of \"bias\" in the significant correlations between demographics (e.g., race, gender) in LLM prompts and responses, but it remains unclear how LLM fairness could be evaluated with more rigorous definitions, such as group fairness or fair representations. We analyze a variety of technical fairness frameworks and find inherent challenges in each that make the development of a fair LLM intractable. We show that each framework either does not logically extend to the general-purpose AI context or is infeasible in practice, primarily due to the large amounts of unstructured training data and the many potential combinations of human populations, use cases, and sensitive attributes. These inherent challenges would persist for general-purpose AI, including LLMs, even if empirical challenges, such as limited participatory input and limited measurement methods, were overcome. Nonetheless, fairness will remain an important type of model evaluation, and there are still promising research directions, particularly the development of standards for the responsibility of LLM developers, context-specific evaluations, and methods of iterative, participatory, and AI-assisted evaluation that could scale fairness across the diverse contexts of modern human-AI interaction.", "date": "2024-05-28", "categories": []}
{"id": "2410.13831", "title": "The Disparate Benefits of Deep Ensembles", "abstract": "Ensembles of Deep Neural Networks, Deep Ensembles, are widely used as a simple way to boost predictive performance. However, their impact on algorithmic fairness is not well understood yet. Algorithmic fairness examines how a model's performance varies across socially relevant groups defined by protected attributes such as age, gender, or race. In this work, we explore the interplay between the performance gains from Deep Ensembles and fairness. Our analysis reveals that they unevenly favor different groups, a phenomenon that we term the disparate benefits effect. We empirically investigate this effect using popular facial analysis and medical imaging datasets with protected group attributes and find that it affects multiple established group fairness metrics, including statistical parity and equal opportunity. Furthermore, we identify that the per-group differences in predictive diversity of ensemble members can explain this effect. Finally, we demonstrate that the classical Hardt post-processing method is particularly effective at mitigating the disparate benefits effect of Deep Ensembles by leveraging their better-calibrated predictive distributions.", "date": "2024-10-17", "categories": []}
{"id": "2505.14971", "title": "DECASTE: Unveiling Caste Stereotypes in Large Language Models through Multi-Dimensional Bias Analysis", "abstract": "Recent advancements in large language models (LLMs) have revolutionized natural language processing (NLP) and expanded their applications across diverse domains. However, despite their impressive capabilities, LLMs have been shown to reflect and perpetuate harmful societal biases, including those based on ethnicity, gender, and religion. A critical and underexplored issue is the reinforcement of caste-based biases, particularly towards India's marginalized caste groups such as Dalits and Shudras. In this paper, we address this gap by proposing DECASTE, a novel, multi-dimensional framework designed to detect and assess both implicit and explicit caste biases in LLMs. Our approach evaluates caste fairness across four dimensions: socio-cultural, economic, educational, and political, using a range of customized prompting strategies. By benchmarking several state-of-the-art LLMs, we reveal that these models systematically reinforce caste biases, with significant disparities observed in the treatment of oppressed versus dominant caste groups. For example, bias scores are notably elevated when comparing Dalits and Shudras with dominant caste groups, reflecting societal prejudices that persist in model outputs. These results expose the subtle yet pervasive caste biases in LLMs and emphasize the need for more comprehensive and inclusive bias evaluation methodologies that assess the potential risks of deploying such models in real-world contexts.", "date": "2025-05-20", "categories": []}
{"id": "2506.04482", "title": "Understanding and Meeting Practitioner Needs When Measuring Representational Harms Caused by LLM-Based Systems", "abstract": "The NLP research community has made publicly available numerous instruments for measuring representational harms caused by large language model (LLM)-based systems. These instruments have taken the form of datasets, metrics, tools, and more. In this paper, we examine the extent to which such instruments meet the needs of practitioners tasked with evaluating LLM-based systems. Via semi-structured interviews with 12 such practitioners, we find that practitioners are often unable to use publicly available instruments for measuring representational harms. We identify two types of challenges. In some cases, instruments are not useful because they do not meaningfully measure what practitioners seek to measure or are otherwise misaligned with practitioner needs. In other cases, instruments - even useful instruments - are not used by practitioners due to practical and institutional barriers impeding their uptake. Drawing on measurement theory and pragmatic measurement, we provide recommendations for addressing these challenges to better meet practitioner needs.", "date": "2025-06-04", "categories": []}
{"id": "2503.09969", "title": "Detecting Dataset Bias in Medical AI: A Generalized and Modality-Agnostic Auditing Framework", "abstract": "Artificial Intelligence (AI) is now firmly at the center of evidence-based medicine. Despite many success stories that edge the path of AI's rise in healthcare, there are comparably many reports of significant shortcomings and unexpected behavior of AI in deployment. A major reason for these limitations is AI's reliance on association-based learning, where non-representative machine learning datasets can amplify latent bias during training and/or hide it during testing. To unlock new tools capable of foreseeing and preventing such AI bias issues, we present G-AUDIT. Generalized Attribute Utility and Detectability-Induced bias Testing (G-AUDIT) for datasets is a modality-agnostic dataset auditing framework that allows for generating targeted hypotheses about sources of bias in training or testing data. Our method examines the relationship between task-level annotations (commonly referred to as ``labels'') and data properties including patient attributes (e.g., age, sex) and environment/acquisition characteristics (e.g., clinical site, imaging protocols). G-AUDIT quantifies the extent to which the observed data attributes pose a risk for shortcut learning, or in the case of testing data, might hide predictions made based on spurious associations. We demonstrate the broad applicability of our method by analyzing large-scale medical datasets for three distinct modalities and machine learning tasks: skin lesion classification in images, stigmatizing language classification in Electronic Health Records (EHR), and mortality prediction for ICU tabular data. In each setting, G-AUDIT successfully identifies subtle biases commonly overlooked by traditional qualitative methods, underscoring its practical value in exposing dataset-level risks and supporting the downstream development of reliable AI systems.", "date": "2025-03-13", "categories": []}
{"id": "2505.13469", "title": "Algorithmic Tradeoffs in Fair Lending: Profitability, Compliance, and Long-Term Impact", "abstract": "As financial institutions increasingly rely on machine learning models to automate lending decisions, concerns about algorithmic fairness have risen. This paper explores the tradeoff between enforcing fairness constraints (such as demographic parity or equal opportunity) and maximizing lender profitability. Through simulations on synthetic data that reflects real-world lending patterns, we quantify how different fairness interventions impact profit margins and default rates. Our results demonstrate that equal opportunity constraints typically impose lower profit costs than demographic parity, but surprisingly, removing protected attributes from the model (fairness through unawareness) outperforms explicit fairness interventions in both fairness and profitability metrics. We further identify the specific economic conditions under which fair lending becomes profitable and analyze the feature-specific drivers of unfairness. These findings offer practical guidance for designing lending algorithms that balance ethical considerations with business objectives.", "date": "2025-05-08", "categories": []}
{"id": "2506.00140", "title": "Balancing Profit and Fairness in Risk-Based Pricing Markets", "abstract": "Dynamic, risk-based pricing can systematically exclude vulnerable consumer groups from essential resources such as health insurance and consumer credit. We show that a regulator can realign private incentives with social objectives through a learned, interpretable tax schedule. First, we provide a formal proposition that bounding each firm's \\emph{local} demographic gap implicitly bounds the \\emph{global} opt-out disparity, motivating firm-level penalties. Building on this insight we introduce \\texttt{MarketSim} -- an open-source, scalable simulator of heterogeneous consumers and profit-maximizing firms -- and train a reinforcement learning (RL) social planner (SP) that selects a bracketed fairness-tax while remaining close to a simple linear prior via an $\\mathcal{L}_1$ regularizer. The learned policy is thus both transparent and easily interpretable. In two empirically calibrated markets, i.e., U.S. health-insurance and consumer-credit, our planner simultaneously raises demand-fairness by up to $16\\%$ relative to unregulated Free Market while outperforming a fixed linear schedule in terms of social welfare without explicit coordination. These results illustrate how AI-assisted regulation can convert a competitive social dilemma into a win-win equilibrium, providing a principled and practical framework for fairness-aware market oversight.", "date": "2025-05-30", "categories": []}
{"id": "2506.03911", "title": "Learning Fair And Effective Points-Based Rewards Programs", "abstract": "Points-based rewards programs are a prevalent way to incentivize customer loyalty; in these programs, customers who make repeated purchases from a seller accumulate points, working toward eventual redemption of a free reward. These programs have recently come under scrutiny due to accusations of unfair practices in their implementation. Motivated by these concerns, we study the problem of fairly designing points-based rewards programs, with a focus on two obstacles that put fairness at odds with their effectiveness. First, due to customer heterogeneity, the seller should set different redemption thresholds for different customers to generate high revenue. Second, the relationship between customer behavior and the number of accumulated points is typically unknown; this requires experimentation which may unfairly devalue customers' previously earned points. We first show that an individually fair rewards program that uses the same redemption threshold for all customers suffers a loss in revenue of at most a factor of $1+\\ln 2$, compared to the optimal personalized strategy that differentiates between customers. We then tackle the problem of designing temporally fair learning algorithms in the presence of demand uncertainty. Toward this goal, we design a learning algorithm that limits the risk of point devaluation due to experimentation by only changing the redemption threshold $O(\\log T)$ times, over a horizon of length $T$. This algorithm achieves the optimal (up to polylogarithmic factors) $\\widetilde{O}(\\sqrt{T})$ regret in expectation. We then modify this algorithm to only ever decrease redemption thresholds, leading to improved fairness at a cost of only a constant factor in regret. Extensive numerical experiments show the limited value of personalization in average-case settings, in addition to demonstrating the strong practical performance of our proposed learning algorithms.", "date": "2025-06-04", "categories": []}
{"id": "2506.03913", "title": "When Fairness Isn't Statistical: The Limits of Machine Learning in Evaluating Legal Reasoning", "abstract": "Legal decisions are increasingly evaluated for fairness, consistency, and bias using machine learning (ML) techniques. In high-stakes domains like refugee adjudication, such methods are often applied to detect disparities in outcomes. Yet it remains unclear whether statistical methods can meaningfully assess fairness in legal contexts shaped by discretion, normative complexity, and limited ground truth.\n  In this paper, we empirically evaluate three common ML approaches (feature-based analysis, semantic clustering, and predictive modeling) on a large, real-world dataset of 59,000+ Canadian refugee decisions (AsyLex). Our experiments show that these methods produce divergent and sometimes contradictory signals, that predictive modeling often depends on contextual and procedural features rather than legal features, and that semantic clustering fails to capture substantive legal reasoning.\n  We show limitations of statistical fairness evaluation, challenge the assumption that statistical regularity equates to fairness, and argue that current computational approaches fall short of evaluating fairness in legally discretionary domains. We argue that evaluating fairness in law requires methods grounded not only in data, but in legal reasoning and institutional context.", "date": "2025-06-04", "categories": []}
{"id": "2506.02017", "title": "Fairness through Feedback: Addressing Algorithmic Misgendering in Automatic Gender Recognition", "abstract": "Automatic Gender Recognition (AGR) systems are an increasingly widespread application in the Machine Learning (ML) landscape. While these systems are typically understood as detecting gender, they often classify datapoints based on observable features correlated at best with either male or female sex. In addition to questionable binary assumptions, from an epistemological point of view, this is problematic for two reasons. First, there exists a gap between the categories the system is meant to predict (woman versus man) and those onto which their output reasonably maps (female versus male). What is more, gender cannot be inferred on the basis of such observable features. This makes AGR tools often unreliable, especially in the case of non-binary and gender non-conforming people. We suggest a theoretical and practical rethinking of AGR systems. To begin, distinctions are made between sex, gender, and gender expression. Then, we build upon the observation that, unlike algorithmic misgendering, human-human misgendering is open to the possibility of re-evaluation and correction. We suggest that analogous dynamics should be recreated in AGR, giving users the possibility to correct the system's output. While implementing such a feedback mechanism could be regarded as diminishing the system's autonomy, it represents a way to significantly increase fairness levels in AGR. This is consistent with the conceptual change of paradigm that we advocate for AGR systems, which should be understood as tools respecting individuals' rights and capabilities of self-expression and determination.", "date": "2025-05-28", "categories": []}
{"id": "2506.02321", "title": "Quantifying Misattribution Unfairness in Authorship Attribution", "abstract": "Authorship misattribution can have profound consequences in real life. In forensic settings simply being considered as one of the potential authors of an evidential piece of text or communication can result in undesirable scrutiny. This raises a fairness question: Is every author in the candidate pool at equal risk of misattribution? Standard evaluation measures for authorship attribution systems do not explicitly account for this notion of fairness. We introduce a simple measure, Misattribution Unfairness Index (MAUIk), which is based on how often authors are ranked in the top k for texts they did not write. Using this measure we quantify the unfairness of five models on two different datasets. All models exhibit high levels of unfairness with increased risks for some authors. Furthermore, we find that this unfairness relates to how the models embed the authors as vectors in the latent search space. In particular, we observe that the risk of misattribution is higher for authors closer to the centroid (or center) of the embedded authors in the haystack. These results indicate the potential for harm and the need for communicating with and calibrating end users on misattribution risk when building and providing such models for downstream use.", "date": "2025-06-02", "categories": []}
{"id": "2404.10508", "title": "White Men Lead, Black Women Help? Benchmarking and Mitigating Language Agency Social Biases in LLMs", "abstract": "Social biases can manifest in language agency. However, very limited research has investigated such biases in Large Language Model (LLM)-generated content. In addition, previous works often rely on string-matching techniques to identify agentic and communal words within texts, falling short of accurately classifying language agency. We introduce the Language Agency Bias Evaluation (LABE) benchmark, which comprehensively evaluates biases in LLMs by analyzing agency levels attributed to different demographic groups in model generations. LABE tests for gender, racial, and intersectional language agency biases in LLMs on 3 text generation tasks: biographies, professor reviews, and reference letters. Using LABE, we unveil language agency social biases in 3 recent LLMs: ChatGPT, Llama3, and Mistral. We observe that: (1) LLM generations tend to demonstrate greater gender bias than human-written texts; (2) Models demonstrate remarkably higher levels of intersectional bias than the other bias aspects. (3) Prompt-based mitigation is unstable and frequently leads to bias exacerbation. Based on our observations, we propose Mitigation via Selective Rewrite (MSR), a novel bias mitigation strategy that leverages an agency classifier to identify and selectively revise parts of generated texts that demonstrate communal traits. Empirical results prove MSR to be more effective and reliable than prompt-based mitigation method, showing a promising research direction.", "date": "2024-04-16", "categories": []}
{"id": "2506.00256", "title": "The Impact of Disability Disclosure on Fairness and Bias in LLM-Driven Candidate Selection", "abstract": "As large language models (LLMs) become increasingly integrated into hiring processes, concerns about fairness have gained prominence. When applying for jobs, companies often request/require demographic information, including gender, race, and disability or veteran status. This data is collected to support diversity and inclusion initiatives, but when provided to LLMs, especially disability-related information, it raises concerns about potential biases in candidate selection outcomes. Many studies have highlighted how disability can impact CV screening, yet little research has explored the specific effect of voluntarily disclosed information on LLM-driven candidate selection. This study seeks to bridge that gap. When candidates shared identical gender, race, qualifications, experience, and backgrounds, and sought jobs with minimal employment rate gaps between individuals with and without disabilities (e.g., Cashier, Software Developer), LLMs consistently favored candidates who disclosed that they had no disability. Even in cases where candidates chose not to disclose their disability status, the LLMs were less likely to select them compared to those who explicitly stated they did not have a disability.", "date": "2025-05-30", "categories": []}
{"id": "2506.00756", "title": "\"Who experiences large model decay and why?\" A Hierarchical Framework for Diagnosing Heterogeneous Performance Drift", "abstract": "Machine learning (ML) models frequently experience performance degradation when deployed in new contexts. Such degradation is rarely uniform: some subgroups may suffer large performance decay while others may not. Understanding where and how large differences in performance arise is critical for designing targeted corrective actions that mitigate decay for the most affected subgroups while minimizing any unintended effects. Current approaches do not provide such detailed insight, as they either (i) explain how average performance shifts arise or (ii) identify adversely affected subgroups without insight into how this occurred. To this end, we introduce a Subgroup-scanning Hierarchical Inference Framework for performance drifT (SHIFT). SHIFT first asks \"Is there any subgroup with unacceptably large performance decay due to covariate/outcome shifts?\" (Where?) and, if so, dives deeper to ask \"Can we explain this using more detailed variable(subset)-specific shifts?\" (How?). In real-world experiments, we find that SHIFT identifies interpretable subgroups affected by performance decay, and suggests targeted actions that effectively mitigate the decay.", "date": "2025-05-31", "categories": []}
{"id": "2506.01709", "title": "Fairness Dynamics During Training", "abstract": "We investigate fairness dynamics during Large Language Model (LLM) training to enable the diagnoses of biases and mitigations through training interventions like early stopping; we find that biases can emerge suddenly and do not always follow common performance metrics. We introduce two new metrics to evaluate fairness dynamics holistically during model pre-training: Average Rank and Jensen-Shannon Divergence by Parts. These metrics provide insights into the Pythia models' progression of biases in gender prediction of occupations on the WinoBias dataset. By monitoring these dynamics, we find that (1) Pythia-6.9b is biased towards men; it becomes more performant and confident predicting \"male\" than \"female\" during training, (2) via early-stopping, Pythia-6.9b can exchange 1.7% accuracy on LAMBADA for a 92.5% increase in fairness, and (3) larger models can exhibit more bias; Pythia-6.9b makes more assumptions about gender than Pythia-160m, even when a subject's gender is not specified.", "date": "2025-06-02", "categories": []}
{"id": "2405.12372", "title": "DispaRisk: Auditing Fairness Through Usable Information", "abstract": "Machine Learning algorithms (ML) impact virtually every aspect of human lives and have found use across diverse sectors including healthcare, finance, and education. Often, ML algorithms have been found to exacerbate societal biases present in datasets leading to adversarial impacts on subsets/groups of individuals and in many cases on minority groups. To effectively mitigate these untoward effects, it is crucial that disparities/biases are identified early in a ML pipeline. This proactive approach facilitates timely interventions to prevent bias amplification and reduce complexity at later stages of model development. In this paper, we leverage recent advancements in usable information theory to introduce DispaRisk, a novel framework designed to proactively assess the potential risks of disparities in datasets during the initial stages of the ML pipeline. We evaluate DispaRisk's effectiveness by benchmarking it against commonly used datasets in fairness research. Our findings demonstrate DispaRisk's capabilities to identify datasets with a high risk of discrimination, detect model families prone to biases within an ML pipeline, and enhance the explainability of these bias risks. This work contributes to the development of fairer ML systems by providing a robust tool for early bias detection and mitigation.", "date": "2024-05-20", "categories": []}
{"id": "2505.14449", "title": "Mitigating Subgroup Disparities in Multi-Label Speech Emotion Recognition: A Pseudo-Labeling and Unsupervised Learning Approach", "abstract": "While subgroup disparities and performance bias are increasingly studied in computational research, fairness in categorical Speech Emotion Recognition (SER) remains underexplored. Existing methods often rely on explicit demographic labels, which are difficult to obtain due to privacy concerns. To address this limitation, we introduce an Implicit Demography Inference (IDI) module that leverages pseudo-labeling from a pre-trained model and unsupervised learning using k-means clustering to mitigate bias in SER. Our experiments show that pseudo-labeling IDI reduces subgroup disparities, improving fairness metrics by over 28% with less than a 2% decrease in SER accuracy. Also, the unsupervised IDI yields more than a 4.6% improvement in fairness metrics with a drop of less than 3.6% in SER performance. Further analyses reveal that the unsupervised IDI consistently mitigates race and age disparities, demonstrating its potential when explicit demographic information is unavailable.", "date": "2025-05-20", "categories": []}
{"id": "2505.23829", "title": "BiasFilter: An Inference-Time Debiasing Framework for Large Language Models", "abstract": "Mitigating social bias in large language models (LLMs) has become an increasingly important research objective. However, existing debiasing methods often incur high human and computational costs, exhibit limited effectiveness, and struggle to scale to larger models and open-ended generation tasks. To address these limitations, this paper proposes BiasFilter, a model-agnostic, inference-time debiasing framework that integrates seamlessly with both open-source and API-based LLMs. Instead of relying on retraining with balanced data or modifying model parameters, BiasFilter enforces fairness by filtering generation outputs in real time. Specifically, it periodically evaluates intermediate outputs every few tokens, maintains an active set of candidate continuations, and incrementally completes generation by discarding low-reward segments based on a fairness reward signal. To support this process, we construct a fairness preference dataset and train an implicit reward model to assess token-level fairness in generated responses. Extensive experiments demonstrate that BiasFilter effectively mitigates social bias across a range of LLMs while preserving overall generation quality.", "date": "2025-05-28", "categories": []}
{"id": "2505.23996", "title": "Is Your Model Fairly Certain? Uncertainty-Aware Fairness Evaluation for LLMs", "abstract": "The recent rapid adoption of large language models (LLMs) highlights the critical need for benchmarking their fairness. Conventional fairness metrics, which focus on discrete accuracy-based evaluations (i.e., prediction correctness), fail to capture the implicit impact of model uncertainty (e.g., higher model confidence about one group over another despite similar accuracy). To address this limitation, we propose an uncertainty-aware fairness metric, UCerF, to enable a fine-grained evaluation of model fairness that is more reflective of the internal bias in model decisions compared to conventional fairness measures. Furthermore, observing data size, diversity, and clarity issues in current datasets, we introduce a new gender-occupation fairness evaluation dataset with 31,756 samples for co-reference resolution, offering a more diverse and suitable dataset for evaluating modern LLMs. We establish a benchmark, using our metric and dataset, and apply it to evaluate the behavior of ten open-source LLMs. For example, Mistral-7B exhibits suboptimal fairness due to high confidence in incorrect predictions, a detail overlooked by Equalized Odds but captured by UCerF. Overall, our proposed LLM benchmark, which evaluates fairness with uncertainty awareness, paves the way for developing more transparent and accountable AI systems.", "date": "2025-05-29", "categories": []}
{"id": "2505.24023", "title": "Multi-Group Proportional Representation for Text-to-Image Models", "abstract": "Text-to-image (T2I) generative models can create vivid, realistic images from textual descriptions. As these models proliferate, they expose new concerns about their ability to represent diverse demographic groups, propagate stereotypes, and efface minority populations. Despite growing attention to the \"safe\" and \"responsible\" design of artificial intelligence (AI), there is no established methodology to systematically measure and control representational harms in image generation. This paper introduces a novel framework to measure the representation of intersectional groups in images generated by T2I models by applying the Multi-Group Proportional Representation (MPR) metric. MPR evaluates the worst-case deviation of representation statistics across given population groups in images produced by a generative model, allowing for flexible and context-specific measurements based on user requirements. We also develop an algorithm to optimize T2I models for this metric. Through experiments, we demonstrate that MPR can effectively measure representation statistics across multiple intersectional groups and, when used as a training objective, can guide models toward a more balanced generation across demographic groups while maintaining generation quality.", "date": "2025-05-29", "categories": []}
{"id": "2308.15651", "title": "Ensuring User-side Fairness in Dynamic Recommender Systems", "abstract": "User-side group fairness is crucial for modern recommender systems, aiming to alleviate performance disparities among user groups defined by sensitive attributes like gender, race, or age. In the ever-evolving landscape of user-item interactions, continual adaptation to newly collected data is crucial for recommender systems to stay aligned with the latest user preferences. However, we observe that such continual adaptation often exacerbates performance disparities. This necessitates a thorough investigation into user-side fairness in dynamic recommender systems, an area that has been unexplored in the literature. This problem is challenging due to distribution shifts, frequent model updates, and non-differentiability of ranking metrics. To our knowledge, this paper presents the first principled study on ensuring user-side fairness in dynamic recommender systems. We start with theoretical analyses on fine-tuning v.s. retraining, showing that the best practice is incremental fine-tuning with restart. Guided by our theoretical analyses, we propose FAir Dynamic rEcommender (FADE), an end-to-end fine-tuning framework to dynamically ensure user-side fairness over time. To overcome the non-differentiability of recommendation metrics in the fairness loss, we further introduce Differentiable Hit (DH) as an improvement over the recent NeuralNDCG method, not only alleviating its gradient vanishing issue but also achieving higher efficiency. Besides that, we also address the instability issue of the fairness loss by leveraging the competing nature between the recommendation loss and the fairness loss. Through extensive experiments on real-world datasets, we demonstrate that FADE effectively and efficiently reduces performance disparities with little sacrifice in the overall recommendation performance.", "date": "2023-08-29", "categories": []}
{"id": "2310.09687", "title": "When Collaborative Filtering is not Collaborative: Unfairness of PCA for Recommendations", "abstract": "We study the fairness of dimensionality reduction methods for recommendations. We focus on the fundamental method of principal component analysis (PCA), which identifies latent components and produces a low-rank approximation via the leading components while discarding the trailing components. Prior works have defined notions of \"fair PCA\"; however, these definitions do not answer the following question: why is PCA unfair? We identify two underlying popularity mechanisms that induce item unfairness in PCA. The first negatively impacts less popular items because less popular items rely on trailing latent components to recover their values. The second negatively impacts highly popular items, since the leading PCA components specialize in individual popular items instead of capturing similarities between items. To address these issues, we develop a polynomial-time algorithm, Item-Weighted PCA, that flexibly up-weights less popular items when optimizing for leading principal components. We theoretically show that PCA, in all cases, and Normalized PCA, in cases of block-diagonal matrices, are instances of Item-Weighted PCA. We empirically show that there exist datasets for which Item-Weighted PCA yields the optimal solution while the baselines do not. In contrast to past dimensionality reduction re-weighting techniques, Item-Weighted PCA solves a convex optimization problem and enforces a hard rank constraint. Our evaluations on real-world datasets show that Item-Weighted PCA not only mitigates both unfairness mechanisms, but also produces recommendations that outperform those of PCA baselines.", "date": "2023-10-15", "categories": []}
{"id": "2502.08166", "title": "From Individual Experience to Collective Evidence: A Reporting-Based Framework for Identifying Systemic Harms", "abstract": "When an individual reports a negative interaction with some system, how can their personal experience be contextualized within broader patterns of system behavior? We study the reporting database problem, where individual reports of adverse events arrive sequentially, and are aggregated over time. In this work, our goal is to identify whether there are subgroups--defined by any combination of relevant features--that are disproportionately likely to experience harmful interactions with the system. We formalize this problem as a sequential hypothesis test, and identify conditions on reporting behavior that are sufficient for making inferences about disparities in true rates of harm across subgroups. We show that algorithms for sequential hypothesis tests can be applied to this problem with a standard multiple testing correction. We then demonstrate our method on real-world datasets, including mortgage decisions and vaccine side effects; on each, our method (re-)identifies subgroups known to experience disproportionate harm using only a fraction of the data that was initially used to discover them.", "date": "2025-02-12", "categories": []}
{"id": "2505.23528", "title": "Comparative assessment of fairness definitions and bias mitigation strategies in machine learning-based diagnosis of Alzheimer's disease from MR images", "abstract": "The present study performs a comprehensive fairness analysis of machine learning (ML) models for the diagnosis of Mild Cognitive Impairment (MCI) and Alzheimer's disease (AD) from MRI-derived neuroimaging features. Biases associated with age, race, and gender in a multi-cohort dataset, as well as the influence of proxy features encoding these sensitive attributes, are investigated. The reliability of various fairness definitions and metrics in the identification of such biases is also assessed. Based on the most appropriate fairness measures, a comparative analysis of widely used pre-processing, in-processing, and post-processing bias mitigation strategies is performed. Moreover, a novel composite measure is introduced to quantify the trade-off between fairness and performance by considering the F1-score and the equalized odds ratio, making it appropriate for medical diagnostic applications. The obtained results reveal the existence of biases related to age and race, while no significant gender bias is observed. The deployed mitigation strategies yield varying improvements in terms of fairness across the different sensitive attributes and studied subproblems. For race and gender, Reject Option Classification improves equalized odds by 46% and 57%, respectively, and achieves harmonic mean scores of 0.75 and 0.80 in the MCI versus AD subproblem, whereas for age, in the same subproblem, adversarial debiasing yields the highest equalized odds improvement of 40% with a harmonic mean score of 0.69. Insights are provided into how variations in AD neuropathology and risk factors, associated with demographic characteristics, influence model fairness.", "date": "2025-05-29", "categories": []}
{"id": "2503.06150", "title": "Do Fairness Interventions Come at the Cost of Privacy: Evaluations for Binary Classifiers", "abstract": "While in-processing fairness approaches show promise in mitigating biased predictions, their potential impact on privacy leakage remains under-explored. We aim to address this gap by assessing the privacy risks of fairness-enhanced binary classifiers via membership inference attacks (MIAs) and attribute inference attacks (AIAs). Surprisingly, our results reveal that enhancing fairness does not necessarily lead to privacy compromises. For example, these fairness interventions exhibit increased resilience against MIAs and AIAs. This is because fairness interventions tend to remove sensitive information among extracted features and reduce confidence scores for the majority of training data for fairer predictions. However, during the evaluations, we uncover a potential threat mechanism that exploits prediction discrepancies between fair and biased models, leading to advanced attack results for both MIAs and AIAs. This mechanism reveals potent vulnerabilities of fair models and poses significant privacy risks of current fairness methods. Extensive experiments across multiple datasets, attack methods, and representative fairness approaches confirm our findings and demonstrate the efficacy of the uncovered mechanism. Our study exposes the under-explored privacy threats in fairness studies, advocating for thorough evaluations of potential security vulnerabilities before model deployments.", "date": "2025-03-08", "categories": []}
{"id": "2505.20885", "title": "Improved Bounds for Swap Multicalibration and Swap Omniprediction", "abstract": "In this paper, we consider the related problems of multicalibration -- a multigroup fairness notion and omniprediction -- a simultaneous loss minimization paradigm, both in the distributional and online settings. The recent work of Garg et al. (2024) raised the open problem of whether it is possible to efficiently achieve $O(\\sqrt{T})$ $\\ell_{2}$-multicalibration error against bounded linear functions. In this paper, we answer this question in a strongly affirmative sense. We propose an efficient algorithm that achieves $O(T^{\\frac{1}{3}})$ $\\ell_{2}$-swap multicalibration error (both in high probability and expectation). On propagating this bound onward, we obtain significantly improved rates for $\\ell_{1}$-swap multicalibration and swap omniprediction for a loss class of convex Lipschitz functions. In particular, we show that our algorithm achieves $O(T^{\\frac{2}{3}})$ $\\ell_{1}$-swap multicalibration and swap omniprediction errors, thereby improving upon the previous best-known bound of $O(T^{\\frac{7}{8}})$. As a consequence of our improved online results, we further obtain several improved sample complexity rates in the distributional setting. In particular, we establish a $O(\\varepsilon ^ {-3})$ sample complexity of efficiently learning an $\\varepsilon$-swap omnipredictor for the class of convex and Lipschitz functions, $O(\\varepsilon ^{-2.5})$ sample complexity of efficiently learning an $\\varepsilon$-swap agnostic learner for the squared loss, and $O(\\varepsilon ^ {-5}), O(\\varepsilon ^ {-2.5})$ sample complexities of learning $\\ell_{1}, \\ell_{2}$-swap multicalibrated predictors against linear functions, all of which significantly improve on the previous best-known bounds.", "date": "2025-05-27", "categories": []}
{"id": "2505.21584", "title": "Fairness in Federated Learning: Fairness for Whom?", "abstract": "Fairness in federated learning has emerged as a rapidly growing area of research, with numerous works proposing formal definitions and algorithmic interventions. Yet, despite this technical progress, fairness in FL is often defined and evaluated in ways that abstract away from the sociotechnical contexts in which these systems are deployed. In this paper, we argue that existing approaches tend to optimize narrow system level metrics, such as performance parity or contribution-based rewards, while overlooking how harms arise throughout the FL lifecycle and how they impact diverse stakeholders. We support this claim through a critical analysis of the literature, based on a systematic annotation of papers for their fairness definitions, design decisions, evaluation practices, and motivating use cases. Our analysis reveals five recurring pitfalls: 1) fairness framed solely through the lens of server client architecture, 2) a mismatch between simulations and motivating use-cases and contexts, 3) definitions that conflate protecting the system with protecting its users, 4) interventions that target isolated stages of the lifecycle while neglecting upstream and downstream effects, 5) and a lack of multi-stakeholder alignment where multiple fairness definitions can be relevant at once. Building on these insights, we propose a harm centered framework that links fairness definitions to concrete risks and stakeholder vulnerabilities. We conclude with recommendations for more holistic, context-aware, and accountable fairness research in FL.", "date": "2025-05-27", "categories": []}
{"id": "2505.21819", "title": "Representative Language Generation", "abstract": "We introduce \"representative generation,\" extending the theoretical framework for generation proposed by Kleinberg et al. (2024) and formalized by Li et al. (2024), to additionally address diversity and bias concerns in generative models. Our notion requires outputs of a generative model to proportionally represent groups of interest from the training data. We characterize representative uniform and non-uniform generation, introducing the \"group closure dimension\" as a key combinatorial quantity. For representative generation in the limit, we analyze both information-theoretic and computational aspects, demonstrating feasibility for countably infinite hypothesis classes and collections of groups under certain conditions, but proving a negative result for computability using only membership queries. This contrasts with Kleinberg et al.'s (2024) positive results for standard generation in the limit. Our findings provide a rigorous foundation for developing more diverse and representative generative models.", "date": "2025-05-27", "categories": []}
{"id": "2505.22114", "title": "BiMi Sheets: Infosheets for bias mitigation methods", "abstract": "Over the past 15 years, hundreds of bias mitigation methods have been proposed in the pursuit of fairness in machine learning (ML). However, algorithmic biases are domain-, task-, and model-specific, leading to a `portability trap': bias mitigation solutions in one context may not be appropriate in another. Thus, a myriad of design choices have to be made when creating a bias mitigation method, such as the formalization of fairness it pursues, and where and how it intervenes in the ML pipeline. This creates challenges in benchmarking and comparing the relative merits of different bias mitigation methods, and limits their uptake by practitioners.\n  We propose BiMi Sheets as a portable, uniform guide to document the design choices of any bias mitigation method. This enables researchers and practitioners to quickly learn its main characteristics and to compare with their desiderata. Furthermore, the sheets' structure allow for the creation of a structured database of bias mitigation methods. In order to foster the sheets' adoption, we provide a platform for finding and creating BiMi Sheets at bimisheet.com.", "date": "2025-05-28", "categories": []}
{"id": "2402.17389", "title": "FairBelief -- Assessing Harmful Beliefs in Language Models", "abstract": "Language Models (LMs) have been shown to inherit undesired biases that might hurt minorities and underrepresented groups if such systems were integrated into real-world applications without careful fairness auditing. This paper proposes FairBelief, an analytical approach to capture and assess beliefs, i.e., propositions that an LM may embed with different degrees of confidence and that covertly influence its predictions. With FairBelief, we leverage prompting to study the behavior of several state-of-the-art LMs across different previously neglected axes, such as model scale and likelihood, assessing predictions on a fairness dataset specifically designed to quantify LMs' outputs' hurtfulness. Finally, we conclude with an in-depth qualitative assessment of the beliefs emitted by the models. We apply FairBelief to English LMs, revealing that, although these architectures enable high performances on diverse natural language processing tasks, they show hurtful beliefs about specific genders. Interestingly, training procedure and dataset, model scale, and architecture induce beliefs of different degrees of hurtfulness.", "date": "2024-02-27", "categories": []}
{"id": "2406.09307", "title": "What is Fair? Defining Fairness in Machine Learning for Health", "abstract": "Ensuring that machine learning (ML) models are safe, effective, and equitable across all patients is critical for clinical decision-making and for preventing the amplification of existing health disparities. In this work, we examine how fairness is conceptualized in ML for health, including why ML models may lead to unfair decisions and how fairness has been measured in diverse real-world applications. We review commonly used fairness notions within group, individual, and causal-based frameworks. We also discuss the outlook for future research and highlight opportunities and challenges in operationalizing fairness in health-focused applications.", "date": "2024-06-13", "categories": []}
{"id": "2505.20512", "title": "A Feature-level Bias Evaluation Framework for Facial Expression Recognition Models", "abstract": "Recent studies on fairness have shown that Facial Expression Recognition (FER) models exhibit biases toward certain visually perceived demographic groups. However, the limited availability of human-annotated demographic labels in public FER datasets has constrained the scope of such bias analysis. To overcome this limitation, some prior works have resorted to pseudo-demographic labels, which may distort bias evaluation results. Alternatively, in this paper, we propose a feature-level bias evaluation framework for evaluating demographic biases in FER models under the setting where demographic labels are unavailable in the test set. Extensive experiments demonstrate that our method more effectively evaluates demographic biases compared to existing approaches that rely on pseudo-demographic labels. Furthermore, we observe that many existing studies do not include statistical testing in their bias evaluations, raising concerns that some reported biases may not be statistically significant but rather due to randomness. To address this issue, we introduce a plug-and-play statistical module to ensure the statistical significance of biased evaluation results. A comprehensive bias analysis based on the proposed module is then conducted across three sensitive attributes (age, gender, and race), seven facial expressions, and multiple network architectures on a large-scale dataset, revealing the prominent demographic biases in FER and providing insights on selecting a fairer network architecture.", "date": "2025-05-26", "categories": []}
{"id": "2505.20637", "title": "TrustSkin: A Fairness Pipeline for Trustworthy Facial Affect Analysis Across Skin Tone", "abstract": "Understanding how facial affect analysis (FAA) systems perform across different demographic groups requires reliable measurement of sensitive attributes such as ancestry, often approximated by skin tone, which itself is highly influenced by lighting conditions. This study compares two objective skin tone classification methods: the widely used Individual Typology Angle (ITA) and a perceptually grounded alternative based on Lightness ($L^*$) and Hue ($H^*$). Using AffectNet and a MobileNet-based model, we assess fairness across skin tone groups defined by each method. Results reveal a severe underrepresentation of dark skin tones ($\\sim 2 \\%$), alongside fairness disparities in F1-score (up to 0.08) and TPR (up to 0.11) across groups. While ITA shows limitations due to its sensitivity to lighting, the $H^*$-$L^*$ method yields more consistent subgrouping and enables clearer diagnostics through metrics such as Equal Opportunity. Grad-CAM analysis further highlights differences in model attention patterns by skin tone, suggesting variation in feature encoding. To support future mitigation efforts, we also propose a modular fairness-aware pipeline that integrates perceptual skin tone estimation, model interpretability, and fairness evaluation. These findings emphasize the relevance of skin tone measurement choices in fairness assessment and suggest that ITA-based evaluations may overlook disparities affecting darker-skinned individuals.", "date": "2025-05-27", "categories": []}
{"id": "2505.20918", "title": "Humble AI in the real-world: the case of algorithmic hiring", "abstract": "Humble AI (Knowles et al., 2023) argues for cautiousness in AI development and deployments through scepticism (accounting for limitations of statistical learning), curiosity (accounting for unexpected outcomes), and commitment (accounting for multifaceted values beyond performance). We present a real-world case study for humble AI in the domain of algorithmic hiring. Specifically, we evaluate virtual screening algorithms in a widely used hiring platform that matches candidates to job openings. There are several challenges in misrecognition and stereotyping in such contexts that are difficult to assess through standard fairness and trust frameworks; e.g., someone with a non-traditional background is less likely to rank highly. We demonstrate technical feasibility of how humble AI principles can be translated to practice through uncertainty quantification of ranks, entropy estimates, and a user experience that highlights algorithmic unknowns. We describe preliminary discussions with focus groups made up of recruiters. Future user studies seek to evaluate whether the higher cognitive load of a humble AI system fosters a climate of trust in its outcomes.", "date": "2025-05-27", "categories": []}
{"id": "2302.03675", "title": "Auditing Gender Presentation Differences in Text-to-Image Models", "abstract": "Text-to-image models, which can generate high-quality images based on textual input, have recently enabled various content-creation tools. Despite significantly affecting a wide range of downstream applications, the distributions of these generated images are still not fully understood, especially when it comes to the potential stereotypical attributes of different genders. In this work, we propose a paradigm (Gender Presentation Differences) that utilizes fine-grained self-presentation attributes to study how gender is presented differently in text-to-image models. By probing gender indicators in the input text (e.g., \"a woman\" or \"a man\"), we quantify the frequency differences of presentation-centric attributes (e.g., \"a shirt\" and \"a dress\") through human annotation and introduce a novel metric: GEP. Furthermore, we propose an automatic method to estimate such differences. The automatic GEP metric based on our approach yields a higher correlation with human annotations than that based on existing CLIP scores, consistently across three state-of-the-art text-to-image models. Finally, we demonstrate the generalization ability of our metrics in the context of gender stereotypes related to occupations.", "date": "2023-02-07", "categories": []}
{"id": "2311.09090", "title": "Social Bias Probing: Fairness Benchmarking for Language Models", "abstract": "While the impact of social biases in language models has been recognized, prior methods for bias evaluation have been limited to binary association tests on small datasets, limiting our understanding of bias complexities. This paper proposes a novel framework for probing language models for social biases by assessing disparate treatment, which involves treating individuals differently according to their affiliation with a sensitive demographic group. We curate SoFa, a large-scale benchmark designed to address the limitations of existing fairness collections. SoFa expands the analysis beyond the binary comparison of stereotypical versus anti-stereotypical identities to include a diverse range of identities and stereotypes. Comparing our methodology with existing benchmarks, we reveal that biases within language models are more nuanced than acknowledged, indicating a broader scope of encoded biases than previously recognized. Benchmarking LMs on SoFa, we expose how identities expressing different religions lead to the most pronounced disparate treatments across all models. Finally, our findings indicate that real-life adversities faced by various groups such as women and people with disabilities are mirrored in the behavior of these models.", "date": "2023-11-15", "categories": []}
{"id": "2402.04420", "title": "Measuring Machine Learning Harms from Stereotypes Requires Understanding Who Is Harmed by Which Errors in What Ways", "abstract": "As machine learning applications proliferate, we need an understanding of their potential for harm. However, current fairness metrics are rarely grounded in human psychological experiences of harm. Drawing on the social psychology of stereotypes, we use a case study of gender stereotypes in image search to examine how people react to machine learning errors. First, we use survey studies to show that not all machine learning errors reflect stereotypes nor are equally harmful. Then, in experimental studies we randomly expose participants to stereotype-reinforcing, -violating, and -neutral machine learning errors. We find stereotype-reinforcing errors induce more experientially (i.e., subjectively) harmful experiences, while having minimal changes to cognitive beliefs, attitudes, or behaviors. This experiential harm impacts women more than men. However, certain stereotype-violating errors are more experientially harmful for men, potentially due to perceived threats to masculinity. We conclude that harm cannot be the sole guide in fairness mitigation, and propose a nuanced perspective depending on who is experiencing what harm and why.", "date": "2024-02-06", "categories": []}
{"id": "2402.15481", "title": "Bias and Volatility: A Statistical Framework for Evaluating Large Language Model's Stereotypes and the Associated Generation Inconsistency", "abstract": "We present a novel statistical framework for analyzing stereotypes in large language models (LLMs) by systematically estimating the bias and variation in their generation. Current alignment evaluation metrics often overlook stereotypes' randomness caused by LLMs' inconsistent generative behavior. For instance, LLMs may display contradictory stereotypes, such as those related to gender or race, for identical professions in different contexts. Ignoring this inconsistency risks misleading conclusions in alignment assessments and undermines efforts to evaluate the potential of LLMs to perpetuate or amplify social biases and unfairness.\n  To address this, we propose the Bias-Volatility Framework (BVF), which estimates the probability distribution of stereotypes in LLM outputs. By capturing the variation in generative behavior, BVF assesses both the likelihood and degree to which LLM outputs negatively impact vulnerable groups, enabling a quantification of aggregated discrimination risk. Additionally, we introduce a mathematical framework to decompose this risk into bias risk (from the mean of the stereotype distribution) and volatility risk (from its variation). Applying BVF to 12 widely used LLMs, we find: i) Bias risk is the dominant contributor to discrimination; ii) Most LLMs exhibit substantial pro-male stereotypes across nearly all professions; iii) Reinforcement learning from human feedback reduces bias but increases volatility; iv) Discrimination risk correlates with socio-economic factors, such as professional salaries. Finally, we highlight BVF's broader applicability for assessing how generation inconsistencies in LLMs impact behavior beyond stereotypes.", "date": "2024-02-23", "categories": []}
{"id": "2505.18532", "title": "Preserving AUC Fairness in Learning with Noisy Protected Groups", "abstract": "The Area Under the ROC Curve (AUC) is a key metric for classification, especially under class imbalance, with growing research focus on optimizing AUC over accuracy in applications like medical image analysis and deepfake detection. This leads to fairness in AUC optimization becoming crucial as biases can impact protected groups. While various fairness mitigation techniques exist, fairness considerations in AUC optimization remain in their early stages, with most research focusing on improving AUC fairness under the assumption of clean protected groups. However, these studies often overlook the impact of noisy protected groups, leading to fairness violations in practice. To address this, we propose the first robust AUC fairness approach under noisy protected groups with fairness theoretical guarantees using distributionally robust optimization. Extensive experiments on tabular and image datasets show that our method outperforms state-of-the-art approaches in preserving AUC fairness. The code is in https://github.com/Purdue-M2/AUC_Fairness_with_Noisy_Groups.", "date": "2025-05-24", "categories": []}
{"id": "2505.18895", "title": "Marginal Fairness: Fair Decision-Making under Risk Measures", "abstract": "This paper introduces marginal fairness, a new individual fairness notion for equitable decision-making in the presence of protected attributes such as gender, race, and religion. This criterion ensures that decisions based on generalized distortion risk measures are insensitive to distributional perturbations in protected attributes, regardless of whether these attributes are continuous, discrete, categorical, univariate, or multivariate. To operationalize this notion and reflect real-world regulatory environments (such as the EU gender-neutral pricing regulation), we model business decision-making in highly regulated industries (such as insurance and finance) as a two-step process: (i) a predictive modeling stage, in which a prediction function for the target variable (e.g., insurance losses) is estimated based on both protected and non-protected covariates; and (ii) a decision-making stage, in which a generalized distortion risk measure is applied to the target variable, conditional only on non-protected covariates, to determine the decision. In this second step, we modify the risk measure such that the decision becomes insensitive to the protected attribute, thus enforcing fairness to ensure equitable outcomes under risk-sensitive, regulatory constraints. Furthermore, by utilizing the concept of cascade sensitivity, we extend the marginal fairness framework to capture how dependencies between covariates propagate the influence of protected attributes through the modeling pipeline. A numerical study and an empirical implementation using an auto insurance dataset demonstrate how the framework can be applied in practice.", "date": "2025-05-24", "categories": []}
{"id": "2505.17280", "title": "Mitigate One, Skew Another? Tackling Intersectional Biases in Text-to-Image Models", "abstract": "The biases exhibited by text-to-image (TTI) models are often treated as independent, though in reality, they may be deeply interrelated. Addressing bias along one dimension - such as ethnicity or age - can inadvertently affect another, like gender, either mitigating or exacerbating existing disparities. Understanding these interdependencies is crucial for designing fairer generative models, yet measuring such effects quantitatively remains a challenge. To address this, we introduce BiasConnect, a novel tool for analyzing and quantifying bias interactions in TTI models. BiasConnect uses counterfactual interventions along different bias axes to reveal the underlying structure of these interactions and estimates the effect of mitigating one bias axis on another. These estimates show strong correlation (+0.65) with observed post-mitigation outcomes. Building on BiasConnect, we propose InterMit, an intersectional bias mitigation algorithm guided by user-defined target distributions and priority weights. InterMit achieves lower bias (0.33 vs. 0.52) with fewer mitigation steps (2.38 vs. 3.15 average steps), and yields superior image quality compared to traditional techniques. Although our implementation is training-free, InterMit is modular and can be integrated with many existing debiasing approaches for TTI models, making it a flexible and extensible solution.", "date": "2025-05-22", "categories": []}
{"id": "2505.17525", "title": "Transparency and Proportionality in Post-Processing Algorithmic Bias Correction", "abstract": "Algorithmic decision-making systems sometimes produce errors or skewed predictions toward a particular group, leading to unfair results. Debiasing practices, applied at different stages of the development of such systems, occasionally introduce new forms of unfairness or exacerbate existing inequalities. We focus on post-processing techniques that modify algorithmic predictions to achieve fairness in classification tasks, examining the unintended consequences of these interventions. To address this challenge, we develop a set of measures that quantify the disparity in the flips applied to the solution in the post-processing stage. The proposed measures will help practitioners: (1) assess the proportionality of the debiasing strategy used, (2) have transparency to explain the effects of the strategy in each group, and (3) based on those results, analyze the possibility of the use of some other approaches for bias mitigation or to solve the problem. We introduce a methodology for applying the proposed metrics during the post-processing stage and illustrate its practical application through an example. This example demonstrates how analyzing the proportionality of the debiasing strategy complements traditional fairness metrics, providing a deeper perspective to ensure fairer outcomes across all groups.", "date": "2025-05-23", "categories": []}
{"id": "2505.17533", "title": "Learning Representational Disparities", "abstract": "We propose a fair machine learning algorithm to model interpretable differences between observed and desired human decision-making, with the latter aimed at reducing disparity in a downstream outcome impacted by the human decision. Prior work learns fair representations without considering the outcome in the decision-making process. We model the outcome disparities as arising due to the different representations of the input seen by the observed and desired decision-maker, which we term representational disparities. Our goal is to learn interpretable representational disparities which could potentially be corrected by specific nudges to the human decision, mitigating disparities in the downstream outcome; we frame this as a multi-objective optimization problem using a neural network. Under reasonable simplifying assumptions, we prove that our neural network model of the representational disparity learns interpretable weights that fully mitigate the outcome disparity. We validate objectives and interpret results using real-world German Credit, Adult, and Heritage Health datasets.", "date": "2025-05-23", "categories": []}
{"id": "2305.19294", "title": "Investigating the Effects of Fairness Interventions Using Pointwise Representational Similarity", "abstract": "Machine learning (ML) algorithms can often exhibit discriminatory behavior, negatively affecting certain populations across protected groups. To address this, numerous debiasing methods, and consequently evaluation measures, have been proposed. Current evaluation measures for debiasing methods suffer from two main limitations: (1) they primarily provide a global estimate of unfairness, failing to provide a more fine-grained analysis, and (2) they predominantly analyze the model output on a specific task, failing to generalize the findings to other tasks. In this work, we introduce Pointwise Normalized Kernel Alignment (PNKA), a pointwise representational similarity measure that addresses these limitations by measuring how debiasing measures affect the intermediate representations of individuals. On tabular data, the use of PNKA reveals previously unknown insights: while group fairness predominantly influences a small subset of the population, maintaining high representational similarity for the majority, individual fairness constraints uniformly impact representations across the entire population, altering nearly every data point. We show that by evaluating representations using PNKA, we can reliably predict the behavior of ML models trained on these representations. Moreover, applying PNKA to language embeddings shows that existing debiasing methods may not perform as intended, failing to remove biases from stereotypical words and sentences. Our findings suggest that current evaluation measures for debiasing methods are insufficient, highlighting the need for a deeper understanding of the effects of debiasing methods, and show how pointwise representational similarity metrics can help with fairness audits.", "date": "2023-05-30", "categories": []}
{"id": "2403.12474", "title": "FairSIN: Achieving Fairness in Graph Neural Networks through Sensitive Information Neutralization", "abstract": "Despite the remarkable success of graph neural networks (GNNs) in modeling graph-structured data, like other machine learning models, GNNs are also susceptible to making biased predictions based on sensitive attributes, such as race and gender. For fairness consideration, recent state-of-the-art (SOTA) methods propose to filter out sensitive information from inputs or representations, e.g., edge dropping or feature masking. However, we argue that such filtering-based strategies may also filter out some non-sensitive feature information, leading to a sub-optimal trade-off between predictive performance and fairness. To address this issue, we unveil an innovative neutralization-based paradigm, where additional Fairness-facilitating Features (F3) are incorporated into node features or representations before message passing. The F3 are expected to statistically neutralize the sensitive bias in node representations and provide additional nonsensitive information. We also provide theoretical explanations for our rationale, concluding that F3 can be realized by emphasizing the features of each node's heterogeneous neighbors (neighbors with different sensitive attributes). We name our method as FairSIN, and present three implementation variants from both data-centric and model-centric perspectives. Experimental results on five benchmark datasets with three different GNN backbones show that FairSIN significantly improves fairness metrics while maintaining high prediction accuracies.", "date": "2024-03-19", "categories": []}
{"id": "2505.16494", "title": "Accuracy vs. Accuracy: Computational Tradeoffs Between Classification Rates and Utility", "abstract": "We revisit the foundations of fairness and its interplay with utility and efficiency in settings where the training data contain richer labels, such as individual types, rankings, or risk estimates, rather than just binary outcomes. In this context, we propose algorithms that achieve stronger notions of evidence-based fairness than are possible in standard supervised learning. Our methods support classification and ranking techniques that preserve accurate subpopulation classification rates, as suggested by the underlying data distributions, across a broad class of classification rules and downstream applications. Furthermore, our predictors enable loss minimization, whether aimed at maximizing utility or in the service of fair treatment.\n  Complementing our algorithmic contributions, we present impossibility results demonstrating that simultaneously achieving accurate classification rates and optimal loss minimization is, in some cases, computationally infeasible. Unlike prior impossibility results, our notions are not inherently in conflict and are simultaneously satisfied by the Bayes-optimal predictor. Furthermore, we show that each notion can be satisfied individually via efficient learning. Our separation thus stems from the computational hardness of learning a sufficiently good approximation of the Bayes-optimal predictor. These computational impossibilities present a choice between two natural and attainable notions of accuracy that could both be motivated by fairness.", "date": "2025-05-22", "categories": []}
{"id": "2405.15240", "title": "Towards Real-world Debiasing: Rethinking Evaluation, Challenge, and Solution", "abstract": "Spurious correlations in training data significantly hinder the generalization capability of machine learning models when faced with distribution shifts, leading to the proposition of numberous debiasing methods. However, it remains to be asked: \\textit{Do existing benchmarks for debiasing really represent biases in the real world?} Recent works attempt to address such concerns by sampling from real-world data (instead of synthesizing) according to some predefined biased distributions to ensure the realism of individual samples. However, the realism of the biased distribution is more critical yet challenging and underexplored due to the complexity of real-world bias distributions. To tackle the problem, we propose a fine-grained framework for analyzing biased distributions, based on which we empirically and theoretically identify key characteristics of biased distributions in the real world that are poorly represented by existing benchmarks. Towards applicable debiasing in the real world, we further introduce two novel real-world-inspired biases to bridge this gap and build a systematic evaluation framework for real-world debiasing, RDBench\\footnote{RDBench: Code to be released. Preliminary version in supplementary material for anonimized review.}. Furthermore, focusing on the practical setting of debiasing w/o bias label, we find real-world biases pose a novel \\textit{Sparse bias capturing} challenge to the existing paradigm. We propose a simple yet effective approach named Debias in Destruction (DiD), to address the challenge, whose effectiveness is validated with extensive experiments on 8 datasets of various biased distributions.", "date": "2024-05-24", "categories": []}
{"id": "2408.10202", "title": "SANER: Annotation-free Societal Attribute Neutralizer for Debiasing CLIP", "abstract": "Large-scale vision-language models, such as CLIP, are known to contain societal bias regarding protected attributes (e.g., gender, age). This paper aims to address the problems of societal bias in CLIP. Although previous studies have proposed to debias societal bias through adversarial learning or test-time projecting, our comprehensive study of these works identifies two critical limitations: 1) loss of attribute information when it is explicitly disclosed in the input and 2) use of the attribute annotations during debiasing process. To mitigate societal bias in CLIP and overcome these limitations simultaneously, we introduce a simple-yet-effective debiasing method called SANER (societal attribute neutralizer) that eliminates attribute information from CLIP text features only of attribute-neutral descriptions. Experimental results show that SANER, which does not require attribute annotations and preserves original information for attribute-specific descriptions, demonstrates superior debiasing ability than the existing methods.", "date": "2024-08-19", "categories": []}
{"id": "2501.15549", "title": "Optimal Transport on Categorical Data for Counterfactuals using Compositional Data and Dirichlet Transport", "abstract": "Recently, optimal transport-based approaches have gained attention for deriving counterfactuals, e.g., to quantify algorithmic discrimination. However, in the general multivariate setting, these methods are often opaque and difficult to interpret. To address this, alternative methodologies have been proposed, using causal graphs combined with iterative quantile regressions (Ple\\v{c}ko and Meinshausen (2020)) or sequential transport (Fernandes Machado et al. (2025)) to examine fairness at the individual level, often referred to as ``counterfactual fairness.'' Despite these advancements, transporting categorical variables remains a significant challenge in practical applications with real datasets. In this paper, we propose a novel approach to address this issue. Our method involves (1) converting categorical variables into compositional data and (2) transporting these compositions within the probabilistic simplex of $\\mathbb{R}^d$. We demonstrate the applicability and effectiveness of this approach through an illustration on real-world data, and discuss limitations.", "date": "2025-01-26", "categories": []}
{"id": "2505.15475", "title": "LFTF: Locating First and Then Fine-Tuning for Mitigating Gender Bias in Large Language Models", "abstract": "Nowadays, Large Language Models (LLMs) have attracted widespread attention due to their powerful performance. However, due to the unavoidable exposure to socially biased data during training, LLMs tend to exhibit social biases, particularly gender bias. To better explore and quantifying the degree of gender bias in LLMs, we propose a pair of datasets named GenBiasEval and GenHintEval, respectively. The GenBiasEval is responsible for evaluating the degree of gender bias in LLMs, accompanied by an evaluation metric named AFGB-Score (Absolutely Fair Gender Bias Score). Meanwhile, the GenHintEval is used to assess whether LLMs can provide responses consistent with prompts that contain gender hints, along with the accompanying evaluation metric UB-Score (UnBias Score). Besides, in order to mitigate gender bias in LLMs more effectively, we present the LFTF (Locating First and Then Fine-Tuning) algorithm.The algorithm first ranks specific LLM blocks by their relevance to gender bias in descending order using a metric called BMI (Block Mitigating Importance Score). Based on this ranking, the block most strongly associated with gender bias is then fine-tuned using a carefully designed loss function. Numerous experiments have shown that our proposed LFTF algorithm can significantly mitigate gender bias in LLMs while maintaining their general capabilities.", "date": "2025-05-21", "categories": []}
{"id": "2505.13509", "title": "Fuck the Algorithm: Conceptual Issues in Algorithmic Bias", "abstract": "Algorithmic bias has been the subject of much recent controversy. To clarify what is at stake and to make progress resolving the controversy, a better understanding of the concepts involved would be helpful. The discussion here focuses on the disputed claim that algorithms themselves cannot be biased. To clarify this claim we need to know what kind of thing 'algorithms themselves' are, and to disambiguate the several meanings of 'bias' at play. This further involves showing how bias of moral import can result from statistical biases, and drawing connections to previous conceptual work about political artifacts and oppressive things. Data bias has been identified in domains like hiring, policing and medicine. Examples where algorithms themselves have been pinpointed as the locus of bias include recommender systems that influence media consumption, academic search engines that influence citation patterns, and the 2020 UK algorithmically-moderated A-level grades. Recognition that algorithms are a kind of thing that can be biased is key to making decisions about responsibility for harm, and preventing algorithmically mediated discrimination.", "date": "2025-05-16", "categories": []}
{"id": "2505.14080", "title": "Gender Trouble in Language Models: An Empirical Audit Guided by Gender Performativity Theory", "abstract": "Language models encode and subsequently perpetuate harmful gendered stereotypes. Research has succeeded in mitigating some of these harms, e.g. by dissociating non-gendered terms such as occupations from gendered terms such as 'woman' and 'man'. This approach, however, remains superficial given that associations are only one form of prejudice through which gendered harms arise. Critical scholarship on gender, such as gender performativity theory, emphasizes how harms often arise from the construction of gender itself, such as conflating gender with biological sex. In language models, these issues could lead to the erasure of transgender and gender diverse identities and cause harms in downstream applications, from misgendering users to misdiagnosing patients based on wrong assumptions about their anatomy.\n  For FAccT research on gendered harms to go beyond superficial linguistic associations, we advocate for a broader definition of 'gender bias' in language models. We operationalize insights on the construction of gender through language from gender studies literature and then empirically test how 16 language models of different architectures, training datasets, and model sizes encode gender. We find that language models tend to encode gender as a binary category tied to biological sex, and that gendered terms that do not neatly fall into one of these binary categories are erased and pathologized. Finally, we show that larger models, which achieve better results on performance benchmarks, learn stronger associations between gender and sex, further reinforcing a narrow understanding of gender. Our findings lead us to call for a re-evaluation of how gendered harms in language models are defined and addressed.", "date": "2025-05-20", "categories": []}
{"id": "2505.14388", "title": "Algorithmic Hiring and Diversity: Reducing Human-Algorithm Similarity for Better Outcomes", "abstract": "Algorithmic tools are increasingly used in hiring to improve fairness and diversity, often by enforcing constraints such as gender-balanced candidate shortlists. However, we show theoretically and empirically that enforcing equal representation at the shortlist stage does not necessarily translate into more diverse final hires, even when there is no gender bias in the hiring stage. We identify a crucial factor influencing this outcome: the correlation between the algorithm's screening criteria and the human hiring manager's evaluation criteria -- higher correlation leads to lower diversity in final hires. Using a large-scale empirical analysis of nearly 800,000 job applications across multiple technology firms, we find that enforcing equal shortlists yields limited improvements in hire diversity when the algorithmic screening closely mirrors the hiring manager's preferences. We propose a complementary algorithmic approach designed explicitly to diversify shortlists by selecting candidates likely to be overlooked by managers, yet still competitive according to their evaluation criteria. Empirical simulations show that this approach significantly enhances gender diversity in final hires without substantially compromising hire quality. These findings highlight the importance of algorithmic design choices in achieving organizational diversity goals and provide actionable guidance for practitioners implementing fairness-oriented hiring algorithms.", "date": "2025-05-20", "categories": []}
{"id": "2211.05617", "title": "Debiasing Methods for Fairer Neural Models in Vision and Language Research: A Survey", "abstract": "Despite being responsible for state-of-the-art results in several computer vision and natural language processing tasks, neural networks have faced harsh criticism due to some of their current shortcomings. One of them is that neural networks are correlation machines prone to model biases within the data instead of focusing on actual useful causal relationships. This problem is particularly serious in application domains affected by aspects such as race, gender, and age. To prevent models from incurring on unfair decision-making, the AI community has concentrated efforts in correcting algorithmic biases, giving rise to the research area now widely known as fairness in AI. In this survey paper, we provide an in-depth overview of the main debiasing methods for fairness-aware neural networks in the context of vision and language research. We propose a novel taxonomy to better organize the literature on debiasing methods for fairness, and we discuss the current challenges, trends, and important future work directions for the interested researcher and practitioner.", "date": "2022-11-10", "categories": []}
{"id": "2408.06099", "title": "Approximating Discrimination Within Models When Faced With Several Non-Binary Sensitive Attributes", "abstract": "Discrimination mitigation within machine learning (ML) models could be complicated because multiple factors may be interwoven hierarchically and historically. Yet few existing fairness measures can capture the discrimination level within ML models in the face of multiple sensitive attributes (SAs). To bridge this gap, we propose a fairness measure based on distances between sets from a manifold perspective, named as 'Harmonic Fairness measure via Manifolds (HFM)' with two optional versions, which can deal with a fine-grained discrimination evaluation for several SAs of multiple values. Because directly computing HFM may be costly, to accelerate its subprocedure -- the computation of distances of sets, we further propose two approximation algorithms named 'Approximation of distance between sets for one sensitive attribute with multiple values (ApproxDist)' and 'Approximation of extended distance between sets for several sensitive attributes with multiple values (ExtendDist)' to respectively resolve bias evaluation of one single SA with multiple values and that of several SAs with multiple values. Moreover, we provide an algorithmic effectiveness analysis for ApproxDist under certain assumptions to explain how well it could work. The empirical results demonstrate that our proposed fairness measure HFM is valid and approximation algorithms (i.e. ApproxDist and ExtendDist) are effective and efficient.", "date": "2024-08-12", "categories": []}
{"id": "2503.21563", "title": "Fair PCA, One Component at a Time", "abstract": "The Min-Max Fair PCA problem seeks a low-rank representation of multi-group data such that the the approximation error is as balanced as possible across groups. Existing approaches to this problem return a rank-$d$ fair subspace, but lack the fundamental containment property of standard PCA: each rank-$d$ PCA subspace should contain all lower-rank PCA subspaces. To fill this gap, we define fair principal components as directions that minimize the maximum group-wise reconstruction error, subject to orthogonality with previously selected components, and we introduce an iterative method to compute them. This approach preserves the containment property of standard PCA, and reduces to standard \\pca for data with a single group. We analyze the theoretical properties of our method and show empirically that it outperforms existing approaches to Min-Max Fair PCA.", "date": "2025-03-27", "categories": []}
{"id": "2505.07997", "title": "FairZK: A Scalable System to Prove Machine Learning Fairness in Zero-Knowledge", "abstract": "With the rise of machine learning techniques, ensuring the fairness of decisions made by machine learning algorithms has become of great importance in critical applications. However, measuring fairness often requires full access to the model parameters, which compromises the confidentiality of the models. In this paper, we propose a solution using zero-knowledge proofs, which allows the model owner to convince the public that a machine learning model is fair while preserving the secrecy of the model. To circumvent the efficiency barrier of naively proving machine learning inferences in zero-knowledge, our key innovation is a new approach to measure fairness only with model parameters and some aggregated information of the input, but not on any specific dataset. To achieve this goal, we derive new bounds for the fairness of logistic regression and deep neural network models that are tighter and better reflecting the fairness compared to prior work. Moreover, we develop efficient zero-knowledge proof protocols for common computations involved in measuring fairness, including the spectral norm of matrices, maximum, absolute value, and fixed-point arithmetic.\n  We have fully implemented our system, FairZK, that proves machine learning fairness in zero-knowledge. Experimental results show that FairZK is significantly faster than the naive approach and an existing scheme that use zero-knowledge inferences as a subroutine. The prover time is improved by 3.1x--1789x depending on the size of the model and the dataset. FairZK can scale to a large model with 47 million parameters for the first time, and generates a proof for its fairness in 343 seconds. This is estimated to be 4 orders of magnitude faster than existing schemes, which only scale to small models with hundreds to thousands of parameters.", "date": "2025-05-12", "categories": []}
{"id": "2505.11572", "title": "ASR-FAIRBENCH: Measuring and Benchmarking Equity Across Speech Recognition Systems", "abstract": "Automatic Speech Recognition (ASR) systems have become ubiquitous in everyday applications, yet significant disparities in performance across diverse demographic groups persist. In this work, we introduce the ASR-FAIRBENCH leaderboard which is designed to assess both the accuracy and equity of ASR models in real-time. Leveraging the Meta's Fair-Speech dataset, which captures diverse demographic characteristics, we employ a mixed-effects Poisson regression model to derive an overall fairness score. This score is integrated with traditional metrics like Word Error Rate (WER) to compute the Fairness Adjusted ASR Score (FAAS), providing a comprehensive evaluation framework. Our approach reveals significant performance disparities in SOTA ASR models across demographic groups and offers a benchmark to drive the development of more inclusive ASR technologies.", "date": "2025-05-16", "categories": []}
{"id": "2505.12001", "title": "Interactional Fairness in LLM Multi-Agent Systems: An Evaluation Framework", "abstract": "As large language models (LLMs) are increasingly used in multi-agent systems, questions of fairness should extend beyond resource distribution and procedural design to include the fairness of how agents communicate. Drawing from organizational psychology, we introduce a novel framework for evaluating Interactional fairness encompassing Interpersonal fairness (IF) and Informational fairness (InfF) in LLM-based multi-agent systems (LLM-MAS). We extend the theoretical grounding of Interactional Fairness to non-sentient agents, reframing fairness as a socially interpretable signal rather than a subjective experience. We then adapt established tools from organizational justice research, including Colquitt's Organizational Justice Scale and the Critical Incident Technique, to measure fairness as a behavioral property of agent interaction. We validate our framework through a pilot study using controlled simulations of a resource negotiation task. We systematically manipulate tone, explanation quality, outcome inequality, and task framing (collaborative vs. competitive) to assess how IF influences agent behavior. Results show that tone and justification quality significantly affect acceptance decisions even when objective outcomes are held constant. In addition, the influence of IF vs. InfF varies with context. This work lays the foundation for fairness auditing and norm-sensitive alignment in LLM-MAS.", "date": "2025-05-17", "categories": []}
{"id": "2505.12054", "title": "GenderBench: Evaluation Suite for Gender Biases in LLMs", "abstract": "We present GenderBench -- a comprehensive evaluation suite designed to measure gender biases in LLMs. GenderBench includes 14 probes that quantify 19 gender-related harmful behaviors exhibited by LLMs. We release GenderBench as an open-source and extensible library to improve the reproducibility and robustness of benchmarking across the field. We also publish our evaluation of 12 LLMs. Our measurements reveal consistent patterns in their behavior. We show that LLMs struggle with stereotypical reasoning, equitable gender representation in generated texts, and occasionally also with discriminatory behavior in high-stakes scenarios, such as hiring.", "date": "2025-05-17", "categories": []}
{"id": "2505.12100", "title": "Improving Fairness in LLMs Through Testing-Time Adversaries", "abstract": "Large Language Models (LLMs) push the bound-aries in natural language processing and generative AI, driving progress across various aspects of modern society. Unfortunately, the pervasive issue of bias in LLMs responses (i.e., predictions) poses a significant and open challenge, hindering their application in tasks involving ethical sensitivity and responsible decision-making. In this work, we propose a straightforward, user-friendly and practical method to mitigate such biases, enhancing the reliability and trustworthiness of LLMs. Our method creates multiple variations of a given sentence by modifying specific attributes and evaluates the corresponding prediction behavior compared to the original, unaltered, prediction/sentence. The idea behind this process is that critical ethical predictions often exhibit notable inconsistencies, indicating the presence of bias. Unlike previous approaches, our method relies solely on forward passes (i.e., testing-time adversaries), eliminating the need for training, fine-tuning, or prior knowledge of the training data distribution. Through extensive experiments on the popular Llama family, we demonstrate the effectiveness of our method in improving various fairness metrics, focusing on the reduction of disparities in how the model treats individuals from different racial groups. Specifically, using standard metrics, we improve the fairness in Llama3 in up to 27 percentage points. Overall, our approach significantly enhances fairness, equity, and reliability in LLM-generated results without parameter tuning or training data modifications, confirming its effectiveness in practical scenarios. We believe our work establishes an important step toward enabling the use of LLMs in tasks that require ethical considerations and responsible decision-making.", "date": "2025-05-17", "categories": []}
{"id": "2505.12349", "title": "Wisdom from Diversity: Bias Mitigation Through Hybrid Human-LLM Crowds", "abstract": "Despite their performance, large language models (LLMs) can inadvertently perpetuate biases found in the data they are trained on. By analyzing LLM responses to bias-eliciting headlines, we find that these models often mirror human biases. To address this, we explore crowd-based strategies for mitigating bias through response aggregation. We first demonstrate that simply averaging responses from multiple LLMs, intended to leverage the \"wisdom of the crowd\", can exacerbate existing biases due to the limited diversity within LLM crowds. In contrast, we show that locally weighted aggregation methods more effectively leverage the wisdom of the LLM crowd, achieving both bias mitigation and improved accuracy. Finally, recognizing the complementary strengths of LLMs (accuracy) and humans (diversity), we demonstrate that hybrid crowds containing both significantly enhance performance and further reduce biases across ethnic and gender-related contexts.", "date": "2025-05-18", "categories": []}
{"id": "2505.12718", "title": "Automated Bias Assessment in AI-Generated Educational Content Using CEAT Framework", "abstract": "Recent advances in Generative Artificial Intelligence (GenAI) have transformed educational content creation, particularly in developing tutor training materials. However, biases embedded in AI-generated content--such as gender, racial, or national stereotypes--raise significant ethical and educational concerns. Despite the growing use of GenAI, systematic methods for detecting and evaluating such biases in educational materials remain limited. This study proposes an automated bias assessment approach that integrates the Contextualized Embedding Association Test with a prompt-engineered word extraction method within a Retrieval-Augmented Generation framework. We applied this method to AI-generated texts used in tutor training lessons. Results show a high alignment between the automated and manually curated word sets, with a Pearson correlation coefficient of r = 0.993, indicating reliable and consistent bias assessment. Our method reduces human subjectivity and enhances fairness, scalability, and reproducibility in auditing GenAI-produced educational content.", "date": "2025-05-19", "categories": []}
{"id": "2505.13116", "title": "Continuous Fair SMOTE -- Fairness-Aware Stream Learning from Imbalanced Data", "abstract": "As machine learning is increasingly applied in an online fashion to deal with evolving data streams, the fairness of these algorithms is a matter of growing ethical and legal concern. In many use cases, class imbalance in the data also needs to be dealt with to ensure predictive performance. Current fairness-aware stream learners typically attempt to solve these issues through in- or post-processing by focusing on optimizing one specific discrimination metric, addressing class imbalance in a separate processing step. While C-SMOTE is a highly effective model-agnostic pre-processing approach to mitigate class imbalance, as a side effect of this method, algorithmic bias is often introduced.\n  Therefore, we propose CFSMOTE - a fairness-aware, continuous SMOTE variant - as a pre-processing approach to simultaneously address the class imbalance and fairness concerns by employing situation testing and balancing fairness-relevant groups during oversampling. Unlike other fairness-aware stream learners, CFSMOTE is not optimizing for only one specific fairness metric, therefore avoiding potentially problematic trade-offs. Our experiments show significant improvement on several common group fairness metrics in comparison to vanilla C-SMOTE while maintaining competitive performance, also in comparison to other fairness-aware algorithms.", "date": "2025-05-19", "categories": []}
{"id": "2305.17570", "title": "Auditing Fairness by Betting", "abstract": "We provide practical, efficient, and nonparametric methods for auditing the fairness of deployed classification and regression models. Whereas previous work relies on a fixed-sample size, our methods are sequential and allow for the continuous monitoring of incoming data, making them highly amenable to tracking the fairness of real-world systems. We also allow the data to be collected by a probabilistic policy as opposed to sampled uniformly from the population. This enables auditing to be conducted on data gathered for another purpose. Moreover, this policy may change over time and different policies may be used on different subpopulations. Finally, our methods can handle distribution shift resulting from either changes to the model or changes in the underlying population. Our approach is based on recent progress in anytime-valid inference and game-theoretic statistics-the \"testing by betting\" framework in particular. These connections ensure that our methods are interpretable, fast, and easy to implement. We demonstrate the efficacy of our approach on three benchmark fairness datasets.", "date": "2023-05-27", "categories": []}
{"id": "2505.08792", "title": "A Preliminary Framework for Intersectionality in ML Pipelines", "abstract": "Machine learning (ML) has become a go-to solution for improving how we use, experience, and interact with technology (and the world around us). Unfortunately, studies have repeatedly shown that machine learning technologies may not provide adequate support for societal identities and experiences. Intersectionality is a sociological framework that provides a mechanism for explicitly considering complex social identities, focusing on social justice and power. While the framework of intersectionality can support the development of technologies that acknowledge and support all members of society, it has been adopted and adapted in ways that are not always true to its foundations, thereby weakening its potential for impact. To support the appropriate adoption and use of intersectionality for more equitable technological outcomes, we amplify the foundational intersectionality scholarship--Crenshaw, Combahee, and Collins (three C's), to create a socially relevant preliminary framework in developing machine-learning solutions. We use this framework to evaluate and report on the (mis)alignments of intersectionality application in machine learning literature.", "date": "2025-05-06", "categories": []}
{"id": "2505.09471", "title": "Fairness-aware Bayes optimal functional classification", "abstract": "Algorithmic fairness has become a central topic in machine learning, and mitigating disparities across different subpopulations has emerged as a rapidly growing research area. In this paper, we systematically study the classification of functional data under fairness constraints, ensuring the disparity level of the classifier is controlled below a pre-specified threshold. We propose a unified framework for fairness-aware functional classification, tackling an infinite-dimensional functional space, addressing key challenges from the absence of density ratios and intractability of posterior probabilities, and discussing unique phenomena in functional classification. We further design a post-processing algorithm, Fair Functional Linear Discriminant Analysis classifier (Fair-FLDA), which targets at homoscedastic Gaussian processes and achieves fairness via group-wise thresholding. Under weak structural assumptions on eigenspace, theoretical guarantees on fairness and excess risk controls are established. As a byproduct, our results cover the excess risk control of the standard FLDA as a special case, which, to the best of our knowledge, is first time seen. Our theoretical findings are complemented by extensive numerical experiments on synthetic and real datasets, highlighting the practicality of our designed algorithm.", "date": "2025-05-14", "categories": []}
{"id": "2505.07870", "title": "Efficient Fairness Testing in Large Language Models: Prioritizing Metamorphic Relations for Bias Detection", "abstract": "Large Language Models (LLMs) are increasingly deployed in various applications, raising critical concerns about fairness and potential biases in their outputs. This paper explores the prioritization of metamorphic relations (MRs) in metamorphic testing as a strategy to efficiently detect fairness issues within LLMs. Given the exponential growth of possible test cases, exhaustive testing is impractical; therefore, prioritizing MRs based on their effectiveness in detecting fairness violations is crucial. We apply a sentence diversity-based approach to compute and rank MRs to optimize fault detection. Experimental results demonstrate that our proposed prioritization approach improves fault detection rates by 22% compared to random prioritization and 12% compared to distance-based prioritization, while reducing the time to the first failure by 15% and 8%, respectively. Furthermore, our approach performs within 5% of fault-based prioritization in effectiveness, while significantly reducing the computational cost associated with fault labeling. These results validate the effectiveness of diversity-based MR prioritization in enhancing fairness testing for LLMs.", "date": "2025-05-09", "categories": []}
{"id": "2505.08064", "title": "Justified Evidence Collection for Argument-based AI Fairness Assurance", "abstract": "It is well recognised that ensuring fair AI systems is a complex sociotechnical challenge, which requires careful deliberation and continuous oversight across all stages of a system's lifecycle, from defining requirements to model deployment and deprovisioning. Dynamic argument-based assurance cases, which present structured arguments supported by evidence, have emerged as a systematic approach to evaluating and mitigating safety risks and hazards in AI-enabled system development and have also been extended to deal with broader normative goals such as fairness and explainability. This paper introduces a systems-engineering-driven framework, supported by software tooling, to operationalise a dynamic approach to argument-based assurance in two stages. In the first stage, during the requirements planning phase, a multi-disciplinary and multi-stakeholder team define goals and claims to be established (and evidenced) by conducting a comprehensive fairness governance process. In the second stage, a continuous monitoring interface gathers evidence from existing artefacts (e.g. metrics from automated tests), such as model, data, and use case documentation, to support these arguments dynamically. The framework's effectiveness is demonstrated through an illustrative case study in finance, with a focus on supporting fairness-related arguments.", "date": "2025-05-12", "categories": []}
{"id": "2408.05497", "title": "MABR: Multilayer Adversarial Bias Removal Without Prior Bias Knowledge", "abstract": "Models trained on real-world data often mirror and exacerbate existing social biases. Traditional methods for mitigating these biases typically require prior knowledge of the specific biases to be addressed, such as gender or racial biases, and the social groups associated with each instance. In this paper, we introduce a novel adversarial training strategy that operates independently of prior bias-type knowledge and protected attribute labels. Our approach proactively identifies biases during model training by utilizing auxiliary models, which are trained concurrently by predicting the performance of the main model without relying on task labels. Additionally, we implement these auxiliary models at various levels of the feature maps of the main model, enabling the detection of a broader and more nuanced range of bias features. Through experiments on racial and gender biases in sentiment and occupation classification tasks, our method effectively reduces social biases without the need for demographic annotations. Moreover, our approach not only matches but often surpasses the efficacy of methods that require detailed demographic insights, marking a significant advancement in bias mitigation techniques.", "date": "2024-08-10", "categories": []}
{"id": "2410.17500", "title": "Learning Fair and Preferable Allocations through Neural Network", "abstract": "The fair allocation of indivisible resources is a fundamental problem. Existing research has developed various allocation mechanisms or algorithms to satisfy different fairness notions. For example, round robin (RR) was proposed to meet the fairness criterion known as envy-freeness up to one good (EF1). Expert algorithms without mathematical formulations are used in real-world resource allocation problems to find preferable outcomes for users. Therefore, we aim to design mechanisms that strictly satisfy good properties with replicating expert knowledge. However, this problem is challenging because such heuristic rules are often difficult to formalize mathematically, complicating their integration into theoretical frameworks. Additionally, formal algorithms struggle to find preferable outcomes, and directly replicating these implicit rules can result in unfair allocations because human decision-making can introduce biases. In this paper, we aim to learn implicit allocation mechanisms from examples while strictly satisfying fairness constraints, specifically focusing on learning EF1 allocation mechanisms through supervised learning on examples of reported valuations and corresponding allocation outcomes produced by implicit rules. To address this, we developed a neural RR (NRR), a novel neural network that parameterizes RR. NRR is built from a differentiable relaxation of RR and can be trained to learn the agent ordering used for RR. We conducted experiments to learn EF1 allocation mechanisms from examples, demonstrating that our method outperforms baselines in terms of the proximity of predicted allocations and other metrics.", "date": "2024-10-23", "categories": []}
{"id": "2411.03700", "title": "The Root Shapes the Fruit: On the Persistence of Gender-Exclusive Harms in Aligned Language Models", "abstract": "Natural-language assistants are designed to provide users with helpful responses while avoiding harmful outputs, largely achieved through alignment to human preferences. Yet there is limited understanding of whether alignment techniques may inadvertently perpetuate or even amplify harmful biases inherited from their pre-aligned base models. This issue is compounded by the choice of bias evaluation benchmarks in popular preference-finetuned models, which predominantly focus on dominant social categories, such as binary gender, thereby limiting insights into biases affecting underrepresented groups. Towards addressing this gap, we center transgender, nonbinary, and other gender-diverse identities to investigate how alignment procedures interact with pre-existing gender-diverse bias in LLMs. Our key contributions include: 1) a comprehensive survey of bias evaluation modalities across leading preference-finetuned LLMs, highlighting critical gaps in gender-diverse representation, 2) systematic evaluation of gender-diverse biases across 16 models spanning Direct Preference Optimization (DPO) stages, uncovering harms popular bias benchmarks fail to detect, and 3) a flexible framework for measuring harmful biases in implicit reward signals applicable to other social contexts. Our findings reveal that DPO-aligned models are particularly sensitive to supervised finetuning (SFT), and can amplify two forms of real-world gender-diverse harms from their base models: stigmatization and gender non-affirmative language. We conclude with recommendations tailored to DPO and broader alignment practices, advocating for the adoption of community-informed bias evaluation frameworks to more effectively identify and address underrepresented harms in LLMs.", "date": "2024-11-06", "categories": []}
{"id": "2505.04886", "title": "Fairness Perceptions in Regression-based Predictive Models", "abstract": "Regression-based predictive analytics used in modern kidney transplantation is known to inherit biases from training data. This leads to social discrimination and inefficient organ utilization, particularly in the context of a few social groups. Despite this concern, there is limited research on fairness in regression and its impact on organ utilization and placement. This paper introduces three novel divergence-based group fairness notions: (i) independence, (ii) separation, and (iii) sufficiency to assess the fairness of regression-based analytics tools. In addition, fairness preferences are investigated from crowd feedback, in order to identify a socially accepted group fairness criterion for evaluating these tools. A total of 85 participants were recruited from the Prolific crowdsourcing platform, and a Mixed-Logit discrete choice model was used to model fairness feedback and estimate social fairness preferences. The findings clearly depict a strong preference towards the separation and sufficiency fairness notions, and that the predictive analytics is deemed fair with respect to gender and race groups, but unfair in terms of age groups.", "date": "2025-05-08", "categories": []}
{"id": "2505.06259", "title": "Fair Clustering with Clusterlets", "abstract": "Given their widespread usage in the real world, the fairness of clustering methods has become of major interest. Theoretical results on fair clustering show that fairness enjoys transitivity: given a set of small and fair clusters, a trivial centroid-based clustering algorithm yields a fair clustering. Unfortunately, discovering a suitable starting clustering can be computationally expensive, rather complex or arbitrary.\n  In this paper, we propose a set of simple \\emph{clusterlet}-based fuzzy clustering algorithms that match single-class clusters, optimizing fair clustering. Matching leverages clusterlet distance, optimizing for classic clustering objectives, while also regularizing for fairness. Empirical results show that simple matching strategies are able to achieve high fairness, and that appropriate parameter tuning allows to achieve high cohesion and low overlap.", "date": "2025-05-03", "categories": []}
{"id": "2505.07339", "title": "Laypeople's Attitudes Towards Fair, Affirmative, and Discriminatory Decision-Making Algorithms", "abstract": "Affirmative algorithms have emerged as a potential answer to algorithmic discrimination, seeking to redress past harms and rectify the source of historical injustices. We present the results of two experiments ($N$$=$$1193$) capturing laypeople's perceptions of affirmative algorithms -- those which explicitly prioritize the historically marginalized -- in hiring and criminal justice. We contrast these opinions about affirmative algorithms with folk attitudes towards algorithms that prioritize the privileged (i.e., discriminatory) and systems that make decisions independently of demographic groups (i.e., fair). We find that people -- regardless of their political leaning and identity -- view fair algorithms favorably and denounce discriminatory systems. In contrast, we identify disagreements concerning affirmative algorithms: liberals and racial minorities rate affirmative systems as positively as their fair counterparts, whereas conservatives and those from the dominant racial group evaluate affirmative algorithms as negatively as discriminatory systems. We identify a source of these divisions: people have varying beliefs about who (if anyone) is marginalized, shaping their views of affirmative algorithms. We discuss the possibility of bridging these disagreements to bring people together towards affirmative algorithms.", "date": "2025-05-12", "categories": []}
{"id": "2505.05492", "title": "DetoxAI: a Python Toolkit for Debiasing Deep Learning Models in Computer Vision", "abstract": "While machine learning fairness has made significant progress in recent years, most existing solutions focus on tabular data and are poorly suited for vision-based classification tasks, which rely heavily on deep learning. To bridge this gap, we introduce DetoxAI, an open-source Python library for improving fairness in deep learning vision classifiers through post-hoc debiasing. DetoxAI implements state-of-the-art debiasing algorithms, fairness metrics, and visualization tools. It supports debiasing via interventions in internal representations and includes attribution-based visualization tools and quantitative algorithmic fairness metrics to show how bias is mitigated. This paper presents the motivation, design, and use cases of DetoxAI, demonstrating its tangible value to engineers and researchers.", "date": "2025-05-02", "categories": []}
{"id": "2403.16101", "title": "Public Perceptions of Fairness Metrics Across Borders", "abstract": "  Which fairness metrics are appropriately applicable in your contexts? There\nmay be instances of discordance regarding the perception of fairness, even when\nthe outcomes comply with established fairness metrics. Several\nquestionnaire-based surveys have been conducted to evaluate fairness metrics\nwith human perceptions of fairness. However, these surveys were limited in\nscope, including only a few hundred participants within a single country. In\nthis study, we conduct an international survey to evaluate public perceptions\nof various fairness metrics in decision-making scenarios. We collected\nresponses from 1,000 participants in each of China, France, Japan, and the\nUnited States, amassing a total of 4,000 participants, to analyze the\npreferences of fairness metrics. Our survey consists of three distinct\nscenarios paired with four fairness metrics. This investigation explores the\nrelationship between personal attributes and the choice of fairness metrics,\nuncovering a significant influence of national context on these preferences.\n", "date": "2024-03-24", "categories": []}
{"id": "2505.05471", "title": "Facets of Disparate Impact: Evaluating Legally Consistent Bias in\n  Machine Learning", "abstract": "  Leveraging current legal standards, we define bias through the lens of\nmarginal benefits and objective testing with the novel metric \"Objective\nFairness Index\". This index combines the contextual nuances of objective\ntesting with metric stability, providing a legally consistent and reliable\nmeasure. Utilizing the Objective Fairness Index, we provide fresh insights into\nsensitive machine learning applications, such as COMPAS (recidivism\nprediction), highlighting the metric's practical and theoretical significance.\nThe Objective Fairness Index allows one to differentiate between discriminatory\ntests and systemic disparities.\n", "date": "2025-05-08", "categories": []}
{"id": "2505.03992", "title": "Algorithmic Accountability in Small Data: Sample-Size-Induced Bias\n  Within Classification Metrics", "abstract": "  Evaluating machine learning models is crucial not only for determining their\ntechnical accuracy but also for assessing their potential societal\nimplications. While the potential for low-sample-size bias in algorithms is\nwell known, we demonstrate the significance of sample-size bias induced by\ncombinatorics in classification metrics. This revelation challenges the\nefficacy of these metrics in assessing bias with high resolution, especially\nwhen comparing groups of disparate sizes, which frequently arise in social\napplications. We provide analyses of the bias that appears in several commonly\napplied metrics and propose a model-agnostic assessment and correction\ntechnique. Additionally, we analyze counts of undefined cases in metric\ncalculations, which can lead to misleading evaluations if improperly handled.\nThis work illuminates the previously unrecognized challenge of combinatorics\nand probability in standard evaluation practices and thereby advances\napproaches for performing fair and trustworthy classification methods.\n", "date": "2025-05-06", "categories": []}
{"id": "2505.03020", "title": "The Multimodal Paradox: How Added and Missing Modalities Shape Bias and\n  Performance in Multimodal AI", "abstract": "  Multimodal learning, which integrates diverse data sources such as images,\ntext, and structured data, has proven superior to unimodal counterparts in\nhigh-stakes decision-making. However, while performance gains remain the gold\nstandard for evaluating multimodal systems, concerns around bias and robustness\nare frequently overlooked. In this context, this paper explores two key\nresearch questions (RQs): (i) RQ1 examines whether adding a modality\ncon-sistently enhances performance and investigates its role in shaping\nfairness measures, assessing whether it mitigates or amplifies bias in\nmultimodal models; (ii) RQ2 investigates the impact of missing modalities at\ninference time, analyzing how multimodal models generalize in terms of both\nperformance and fairness. Our analysis reveals that incorporating new\nmodalities during training consistently enhances the performance of multimodal\nmodels, while fairness trends exhibit variability across different evaluation\nmeasures and datasets. Additionally, the absence of modalities at inference\ndegrades performance and fairness, raising concerns about its robustness in\nreal-world deployment. We conduct extensive experiments using multimodal\nhealthcare datasets containing images, time series, and structured information\nto validate our findings.\n", "date": "2025-05-05", "categories": []}
{"id": "2407.19655", "title": "AI-Driven Healthcare: A Review on Ensuring Fairness and Mitigating Bias", "abstract": "  Artificial intelligence (AI) is rapidly advancing in healthcare, enhancing\nthe efficiency and effectiveness of services across various specialties,\nincluding cardiology, ophthalmology, dermatology, emergency medicine, etc. AI\napplications have significantly improved diagnostic accuracy, treatment\npersonalization, and patient outcome predictions by leveraging technologies\nsuch as machine learning, neural networks, and natural language processing.\nHowever, these advancements also introduce substantial ethical and fairness\nchallenges, particularly related to biases in data and algorithms. These biases\ncan lead to disparities in healthcare delivery, affecting diagnostic accuracy\nand treatment outcomes across different demographic groups. This review paper\nexamines the integration of AI in healthcare, highlighting critical challenges\nrelated to bias and exploring strategies for mitigation. We emphasize the\nnecessity of diverse datasets, fairness-aware algorithms, and regulatory\nframeworks to ensure equitable healthcare delivery. The paper concludes with\nrecommendations for future research, advocating for interdisciplinary\napproaches, transparency in AI decision-making, and the development of\ninnovative and inclusive AI applications.\n", "date": "2024-07-29", "categories": []}
{"id": "2504.15941", "title": "FairTranslate: An English-French Dataset for Gender Bias Evaluation in\n  Machine Translation by Overcoming Gender Binarity", "abstract": "  Large Language Models (LLMs) are increasingly leveraged for translation tasks\nbut often fall short when translating inclusive language -- such as texts\ncontaining the singular 'they' pronoun or otherwise reflecting fair linguistic\nprotocols. Because these challenges span both computational and societal\ndomains, it is imperative to critically evaluate how well LLMs handle inclusive\ntranslation with a well-founded framework.\n  This paper presents FairTranslate, a novel, fully human-annotated dataset\ndesigned to evaluate non-binary gender biases in machine translation systems\nfrom English to French. FairTranslate includes 2418 English-French sentence\npairs related to occupations, annotated with rich metadata such as the\nstereotypical alignment of the occupation, grammatical gender indicator\nambiguity, and the ground-truth gender label (male, female, or inclusive).\n  We evaluate four leading LLMs (Gemma2-2B, Mistral-7B, Llama3.1-8B,\nLlama3.3-70B) on this dataset under different prompting procedures. Our results\nreveal substantial biases in gender representation across LLMs, highlighting\npersistent challenges in achieving equitable outcomes in machine translation.\nThese findings underscore the need for focused strategies and interventions\naimed at ensuring fair and inclusive language usage in LLM-based translation\nsystems.\n  We make the FairTranslate dataset publicly available on Hugging Face, and\ndisclose the code for all experiments on GitHub.\n", "date": "2025-04-22", "categories": []}
{"id": "2505.01652", "title": "Causally Fair Node Classification on Non-IID Graph Data", "abstract": "  Fair machine learning seeks to identify and mitigate biases in predictions\nagainst unfavorable populations characterized by demographic attributes, such\nas race and gender. Recently, a few works have extended fairness to graph data,\nsuch as social networks, but most of them neglect the causal relationships\namong data instances. This paper addresses the prevalent challenge in\nfairness-aware ML algorithms, which typically assume Independent and\nIdentically Distributed (IID) data. We tackle the overlooked domain of non-IID,\ngraph-based settings where data instances are interconnected, influencing the\noutcomes of fairness interventions. We base our research on the Network\nStructural Causal Model (NSCM) framework and posit two main assumptions:\nDecomposability and Graph Independence, which enable the computation of\ninterventional distributions in non-IID settings using the $do$-calculus. Based\non that, we develop the Message Passing Variational Autoencoder for Causal\nInference (MPVA) to compute interventional distributions and facilitate\ncausally fair node classification through estimated interventional\ndistributions. Empirical evaluations on semi-synthetic and real-world datasets\ndemonstrate that MPVA outperforms conventional methods by effectively\napproximating interventional distributions and mitigating bias. The\nimplications of our findings underscore the potential of causality-based\nfairness in complex ML applications, setting the stage for further research\ninto relaxing the initial assumptions to enhance model fairness.\n", "date": "2025-05-03", "categories": []}
{"id": "2505.01699", "title": "Component-Based Fairness in Face Attribute Classification with Bayesian\n  Network-informed Meta Learning", "abstract": "  The widespread integration of face recognition technologies into various\napplications (e.g., access control and personalized advertising) necessitates a\ncritical emphasis on fairness. While previous efforts have focused on\ndemographic fairness, the fairness of individual biological face components\nremains unexplored. In this paper, we focus on face component fairness, a\nfairness notion defined by biological face features. To our best knowledge, our\nwork is the first work to mitigate bias of face attribute prediction at the\nbiological feature level. In this work, we identify two key challenges in\noptimizing face component fairness: attribute label scarcity and attribute\ninter-dependencies, both of which limit the effectiveness of bias mitigation\nfrom previous approaches. To address these issues, we propose \\textbf{B}ayesian\n\\textbf{N}etwork-informed \\textbf{M}eta \\textbf{R}eweighting (BNMR), which\nincorporates a Bayesian Network calibrator to guide an adaptive\nmeta-learning-based sample reweighting process. During the training process of\nour approach, the Bayesian Network calibrator dynamically tracks model bias and\nencodes prior probabilities for face component attributes to overcome the above\nchallenges. To demonstrate the efficacy of our approach, we conduct extensive\nexperiments on a large-scale real-world human face dataset. Our results show\nthat BNMR is able to consistently outperform recent face bias mitigation\nbaselines. Moreover, our results suggest a positive impact of face component\nfairness on the commonly considered demographic fairness (e.g.,\n\\textit{gender}). Our findings pave the way for new research avenues on face\ncomponent fairness, suggesting that face component fairness could serve as a\npotential surrogate objective for demographic fairness. The code for our work\nis publicly\navailable~\\footnote{https://github.com/yliuaa/BNMR-FairCompFace.git}.\n", "date": "2025-05-03", "categories": []}
{"id": "2505.02352", "title": "Social Biases in Knowledge Representations of Wikidata separates Global\n  North from Global South", "abstract": "  Knowledge Graphs have become increasingly popular due to their wide usage in\nvarious downstream applications, including information retrieval, chatbot\ndevelopment, language model construction, and many others. Link prediction (LP)\nis a crucial downstream task for knowledge graphs, as it helps to address the\nproblem of the incompleteness of the knowledge graphs. However, previous\nresearch has shown that knowledge graphs, often created in a (semi) automatic\nmanner, are not free from social biases. These biases can have harmful effects\non downstream applications, especially by leading to unfair behavior toward\nminority groups. To understand this issue in detail, we develop a framework --\nAuditLP -- deploying fairness metrics to identify biased outcomes in LP,\nspecifically how occupations are classified as either male or female-dominated\nbased on gender as a sensitive attribute. We have experimented with the\nsensitive attribute of age and observed that occupations are categorized as\nyoung-biased, old-biased, and age-neutral. We conduct our experiments on a\nlarge number of knowledge triples that belong to 21 different geographies\nextracted from the open-sourced knowledge graph, Wikidata. Our study shows that\nthe variance in the biased outcomes across geographies neatly mirrors the\nsocio-economic and cultural division of the world, resulting in a transparent\npartition of the Global North from the Global South.\n", "date": "2025-05-05", "categories": []}
{"id": "2505.02609", "title": "Study of the influence of a biased database on the prediction of\n  standard algorithms for selecting the best candidate for an interview", "abstract": "  Artificial intelligence is used at various stages of the recruitment process\nto automatically select the best candidate for a position, with companies\nguaranteeing unbiased recruitment. However, the algorithms used are either\ntrained by humans or are based on learning from past experiences that were\nbiased. In this article, we propose to generate data mimicking external\n(discrimination) and internal biases (self-censorship) in order to train five\nclassic algorithms and to study the extent to which they do or do not find the\nbest candidates according to objective criteria. In addition, we study the\ninfluence of the anonymisation of files on the quality of predictions.\n", "date": "2025-05-05", "categories": []}
{"id": "2505.00802", "title": "Explanations as Bias Detectors: A Critical Study of Local Post-hoc XAI\n  Methods for Fairness Exploration", "abstract": "  As Artificial Intelligence (AI) is increasingly used in areas that\nsignificantly impact human lives, concerns about fairness and transparency have\ngrown, especially regarding their impact on protected groups. Recently, the\nintersection of explainability and fairness has emerged as an important area to\npromote responsible AI systems. This paper explores how explainability methods\ncan be leveraged to detect and interpret unfairness. We propose a pipeline that\nintegrates local post-hoc explanation methods to derive fairness-related\ninsights. During the pipeline design, we identify and address critical\nquestions arising from the use of explanations as bias detectors such as the\nrelationship between distributive and procedural fairness, the effect of\nremoving the protected attribute, the consistency and quality of results across\ndifferent explanation methods, the impact of various aggregation strategies of\nlocal explanations on group fairness evaluations, and the overall\ntrustworthiness of explanations as bias detectors. Our results show the\npotential of explanation methods used for fairness while highlighting the need\nto carefully consider the aforementioned critical aspects.\n", "date": "2025-05-01", "categories": []}
{"id": "2505.01070", "title": "Improving Group Fairness in Knowledge Distillation via Laplace\n  Approximation of Early Exits", "abstract": "  Knowledge distillation (KD) has become a powerful tool for training compact\nstudent models using larger, pretrained teacher models, often requiring less\ndata and computational resources. Teacher models typically possess more layers\nand thus exhibit richer feature representations compared to their student\ncounterparts. Furthermore, student models tend to learn simpler, surface-level\nfeatures in their early layers. This discrepancy can increase errors in groups\nwhere labels spuriously correlate with specific input attributes, leading to a\ndecline in group fairness even when overall accuracy remains comparable to the\nteacher. To mitigate these challenges, Early-Exit Neural Networks (EENNs),\nwhich enable predictions at multiple intermediate layers, have been employed.\nConfidence margins derived from these early exits have been utilized to\nreweight both cross-entropy and distillation losses on a per-instance basis. In\nthis paper, we propose that leveraging Laplace approximation-based methods to\nobtain well-calibrated uncertainty estimates can also effectively reweight\nchallenging instances and improve group fairness. We hypothesize that Laplace\napproximation offers a more robust identification of difficult or ambiguous\ninstances compared to margin-based approaches. To validate our claims, we\nbenchmark our approach using a Bert-based model on the MultiNLI dataset.\n", "date": "2025-05-02", "categories": []}
{"id": "2505.01198", "title": "Gender Bias in Explainability: Investigating Performance Disparity in\n  Post-hoc Methods", "abstract": "  While research on applications and evaluations of explanation methods\ncontinues to expand, fairness of the explanation methods concerning disparities\nin their performance across subgroups remains an often overlooked aspect. In\nthis paper, we address this gap by showing that, across three tasks and five\nlanguage models, widely used post-hoc feature attribution methods exhibit\nsignificant gender disparity with respect to their faithfulness, robustness,\nand complexity. These disparities persist even when the models are pre-trained\nor fine-tuned on particularly unbiased datasets, indicating that the\ndisparities we observe are not merely consequences of biased training data. Our\nresults highlight the importance of addressing disparities in explanations when\ndeveloping and applying explainability methods, as these can lead to biased\noutcomes against certain subgroups, with particularly critical implications in\nhigh-stakes contexts. Furthermore, our findings underscore the importance of\nincorporating the fairness of explanations, alongside overall model fairness\nand explainability, as a requirement in regulatory frameworks.\n", "date": "2025-05-02", "categories": []}
{"id": "2402.13393", "title": "Fairness Risks for Group-conditionally Missing Demographics", "abstract": "  Fairness-aware classification models have gained increasing attention in\nrecent years as concerns grow on discrimination against some demographic\ngroups. Most existing models require full knowledge of the sensitive features,\nwhich can be impractical due to privacy, legal issues, and an individual's fear\nof discrimination. The key challenge we will address is the group dependency of\nthe unavailability, e.g., people of some age range may be more reluctant to\nreveal their age. Our solution augments general fairness risks with\nprobabilistic imputations of the sensitive features, while jointly learning the\ngroup-conditionally missing probabilities in a variational auto-encoder. Our\nmodel is demonstrated effective on both image and tabular datasets, achieving\nan improved balance between accuracy and fairness.\n", "date": "2024-02-20", "categories": []}
{"id": "2504.17066", "title": "Whence Is A Model Fair? Fixing Fairness Bugs via Propensity Score\n  Matching", "abstract": "  Fairness-aware learning aims to mitigate discrimination against specific\nprotected social groups (e.g., those categorized by gender, ethnicity, age)\nwhile minimizing predictive performance loss. Despite efforts to improve\nfairness in machine learning, prior studies have shown that many models remain\nunfair when measured against various fairness metrics. In this paper, we\nexamine whether the way training and testing data are sampled affects the\nreliability of reported fairness metrics. Since training and test sets are\noften randomly sampled from the same population, bias present in the training\ndata may still exist in the test data, potentially skewing fairness\nassessments. To address this, we propose FairMatch, a post-processing method\nthat applies propensity score matching to evaluate and mitigate bias. FairMatch\nidentifies control and treatment pairs with similar propensity scores in the\ntest set and adjusts decision thresholds for different subgroups accordingly.\nFor samples that cannot be matched, we perform probabilistic calibration using\nfairness-aware loss functions. Experimental results demonstrate that our\napproach can (a) precisely locate subsets of the test data where the model is\nunbiased, and (b) significantly reduce bias on the remaining data. Overall,\npropensity score matching offers a principled way to improve both fairness\nevaluation and mitigation, without sacrificing predictive performance.\n", "date": "2025-04-23", "categories": []}
{"id": "2406.09495", "title": "FADE: Towards Fairness-aware Generation for Domain Generalization via\n  Classifier-Guided Score-based Diffusion Models", "abstract": "  Fairness-aware domain generalization (FairDG) has emerged as a critical\nchallenge for deploying trustworthy AI systems, particularly in scenarios\ninvolving distribution shifts. Traditional methods for addressing fairness have\nfailed in domain generalization due to their lack of consideration for\ndistribution shifts. Although disentanglement has been used to tackle FairDG,\nit is limited by its strong assumptions. To overcome these limitations, we\npropose Fairness-aware Classifier-Guided Score-based Diffusion Models (FADE) as\na novel approach to effectively address the FairDG issue. Specifically, we\nfirst pre-train a score-based diffusion model (SDM) and two classifiers to\nequip the model with strong generalization capabilities across different\ndomains. Then, we guide the SDM using these pre-trained classifiers to\neffectively eliminate sensitive information from the generated data. Finally,\nthe generated fair data is used to train downstream classifiers, ensuring\nrobust performance under new data distributions. Extensive experiments on three\nreal-world datasets demonstrate that FADE not only enhances fairness but also\nimproves accuracy in the presence of distribution shifts. Additionally, FADE\noutperforms existing methods in achieving the best accuracy-fairness\ntrade-offs.\n", "date": "2024-06-13", "categories": []}
{"id": "2406.16193", "title": "Semi-Variance Reduction for Fair Federated Learning", "abstract": "  Ensuring fairness in a Federated Learning (FL) system, i.e., a satisfactory\nperformance for all of the participating diverse clients, is an important and\nchallenging problem. There are multiple fair FL algorithms in the literature,\nwhich have been relatively successful in providing fairness. However, these\nalgorithms mostly emphasize on the loss functions of worst-off clients to\nimprove their performance, which often results in the suppression of\nwell-performing ones. As a consequence, they usually sacrifice the system's\noverall average performance for achieving fairness. Motivated by this and\ninspired by two well-known risk modeling methods in Finance, Mean-Variance and\nMean-Semi-Variance, we propose and study two new fair FL algorithms, Variance\nReduction (VRed) and Semi-Variance Reduction (SemiVRed). VRed encourages\nequality between clients' loss functions by penalizing their variance. In\ncontrast, SemiVRed penalizes the discrepancy of only the worst-off clients'\nloss functions from the average loss. Through extensive experiments on multiple\nvision and language datasets, we show that, SemiVRed achieves SoTA performance\nin scenarios with heterogeneous data distributions and improves both fairness\nand system overall average performance.\n", "date": "2024-06-23", "categories": []}
{"id": "2504.21296", "title": "Fairness in Graph Learning Augmented with Machine Learning: A Survey", "abstract": "  Augmenting specialised machine learning techniques into traditional graph\nlearning models has achieved notable success across various domains, including\nfederated graph learning, dynamic graph learning, and graph transformers.\nHowever, the intricate mechanisms of these specialised techniques introduce\nsignificant challenges in maintaining model fairness, potentially resulting in\ndiscriminatory outcomes in high-stakes applications such as recommendation\nsystems, disaster response, criminal justice, and loan approval. This paper\nsystematically examines the unique fairness challenges posed by Graph Learning\naugmented with Machine Learning (GL-ML). It highlights the complex interplay\nbetween graph learning mechanisms and machine learning techniques, emphasising\nhow the augmentation of machine learning both enhances and complicates\nfairness. Additionally, we explore four critical techniques frequently employed\nto improve fairness in GL-ML methods. By thoroughly investigating the root\ncauses and broader implications of fairness challenges in this rapidly evolving\nfield, this work establishes a robust foundation for future research and\ninnovation in GL-ML fairness.\n", "date": "2025-04-30", "categories": []}
{"id": "2504.21400", "title": "Who Gets the Callback? Generative AI and Gender Bias", "abstract": "  Generative artificial intelligence (AI), particularly large language models\n(LLMs), is being rapidly deployed in recruitment and for candidate\nshortlisting. We audit several mid-sized open-source LLMs for gender bias using\na dataset of 332,044 real-world online job postings. For each posting, we\nprompt the model to recommend whether an equally qualified male or female\ncandidate should receive an interview callback. We find that most models tend\nto favor men, especially for higher-wage roles. Mapping job descriptions to the\nStandard Occupational Classification system, we find lower callback rates for\nwomen in male-dominated occupations and higher rates in female-associated ones,\nindicating occupational segregation. A comprehensive analysis of linguistic\nfeatures in job ads reveals strong alignment of model recommendations with\ntraditional gender stereotypes. To examine the role of recruiter identity, we\nsteer model behavior by infusing Big Five personality traits and simulating the\nperspectives of historical figures. We find that less agreeable personas reduce\nstereotyping, consistent with an agreeableness bias in LLMs. Our findings\nhighlight how AI-driven hiring may perpetuate biases in the labor market and\nhave implications for fairness and diversity within firms.\n", "date": "2025-04-30", "categories": []}
{"id": "2504.21634", "title": "Quantitative Auditing of AI Fairness with Differentially Private\n  Synthetic Data", "abstract": "  Fairness auditing of AI systems can identify and quantify biases. However,\ntraditional auditing using real-world data raises security and privacy\nconcerns. It exposes auditors to security risks as they become custodians of\nsensitive information and targets for cyberattacks. Privacy risks arise even\nwithout direct breaches, as data analyses can inadvertently expose confidential\ninformation. To address these, we propose a framework that leverages\ndifferentially private synthetic data to audit the fairness of AI systems. By\napplying privacy-preserving mechanisms, it generates synthetic data that\nmirrors the statistical properties of the original dataset while ensuring\nprivacy. This method balances the goal of rigorous fairness auditing and the\nneed for strong privacy protections. Through experiments on real datasets like\nAdult, COMPAS, and Diabetes, we compare fairness metrics of synthetic and real\ndata. By analyzing the alignment and discrepancies between these metrics, we\nassess the capacity of synthetic data to preserve the fairness properties of\nreal data. Our results demonstrate the framework's ability to enable meaningful\nfairness evaluations while safeguarding sensitive information, proving its\napplicability across critical and sensitive domains.\n", "date": "2025-04-30", "categories": []}
{"id": "2504.20642", "title": "Decision-centric fairness: Evaluation and optimization for resource\n  allocation problems", "abstract": "  Data-driven decision support tools play an increasingly central role in\ndecision-making across various domains. In this work, we focus on binary\nclassification models for predicting positive-outcome scores and deciding on\nresource allocation, e.g., credit scores for granting loans or churn propensity\nscores for targeting customers with a retention campaign. Such models may\nexhibit discriminatory behavior toward specific demographic groups through\ntheir predicted scores, potentially leading to unfair resource allocation. We\nfocus on demographic parity as a fairness metric to compare the proportions of\ninstances that are selected based on their positive outcome scores across\ngroups. In this work, we propose a decision-centric fairness methodology that\ninduces fairness only within the decision-making region -- the range of\nrelevant decision thresholds on the score that may be used to decide on\nresource allocation -- as an alternative to a global fairness approach that\nseeks to enforce parity across the entire score distribution. By restricting\nthe induction of fairness to the decision-making region, the proposed\ndecision-centric approach avoids imposing overly restrictive constraints on the\nmodel, which may unnecessarily degrade the quality of the predicted scores. We\nempirically compare our approach to a global fairness approach on multiple\n(semi-synthetic) datasets to identify scenarios in which focusing on fairness\nwhere it truly matters, i.e., decision-centric fairness, proves beneficial.\n", "date": "2025-04-29", "categories": []}
{"id": "2504.20902", "title": "Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for\n  Visual Classifiers", "abstract": "  A person downloading a pre-trained model from the web should be aware of its\nbiases. Existing approaches for bias identification rely on datasets containing\nlabels for the task of interest, something that a non-expert may not have\naccess to, or may not have the necessary resources to collect: this greatly\nlimits the number of tasks where model biases can be identified. In this work,\nwe present Classifier-to-Bias (C2B), the first bias discovery framework that\nworks without access to any labeled data: it only relies on a textual\ndescription of the classification task to identify biases in the target\nclassification model. This description is fed to a large language model to\ngenerate bias proposals and corresponding captions depicting biases together\nwith task-specific target labels. A retrieval model collects images for those\ncaptions, which are then used to assess the accuracy of the model w.r.t. the\ngiven biases. C2B is training-free, does not require any annotations, has no\nconstraints on the list of biases, and can be applied to any pre-trained model\non any classification task. Experiments on two publicly available datasets show\nthat C2B discovers biases beyond those of the original datasets and outperforms\na recent state-of-the-art bias detection baseline that relies on task-specific\nannotations, being a promising first step toward addressing task-agnostic\nunsupervised bias detection.\n", "date": "2025-04-29", "categories": []}
{"id": "2402.07745", "title": "Predictive Churn with the Set of Good Models", "abstract": "  Issues can arise when research focused on fairness, transparency, or safety\nis conducted separately from research driven by practical deployment concerns\nand vice versa. This separation creates a growing need for translational work\nthat bridges the gap between independently studied concepts that may be\nfundamentally related. This paper explores connections between two seemingly\nunrelated concepts of predictive inconsistency that share intriguing parallels.\nThe first, known as predictive multiplicity, occurs when models that perform\nsimilarly (e.g., nearly equivalent training loss) produce conflicting\npredictions for individual samples. This concept is often emphasized in\nalgorithmic fairness research as a means of promoting transparency in ML model\ndevelopment. The second concept, predictive churn, examines the differences in\nindividual predictions before and after model updates, a key challenge in\ndeploying ML models in consumer-facing applications. We present theoretical and\nempirical results that uncover links between these previously disconnected\nconcepts.\n", "date": "2024-02-12", "categories": []}
{"id": "2411.09804", "title": "Fair Resource Allocation in Weakly Coupled Markov Decision Processes", "abstract": "  We consider fair resource allocation in sequential decision-making\nenvironments modeled as weakly coupled Markov decision processes, where\nresource constraints couple the action spaces of $N$ sub-Markov decision\nprocesses (sub-MDPs) that would otherwise operate independently. We adopt a\nfairness definition using the generalized Gini function instead of the\ntraditional utilitarian (total-sum) objective. After introducing a general but\ncomputationally prohibitive solution scheme based on linear programming, we\nfocus on the homogeneous case where all sub-MDPs are identical. For this case,\nwe show for the first time that the problem reduces to optimizing the\nutilitarian objective over the class of \"permutation invariant\" policies. This\nresult is particularly useful as we can exploit Whittle index policies in the\nrestless bandits setting while, for the more general setting, we introduce a\ncount-proportion-based deep reinforcement learning approach. Finally, we\nvalidate our theoretical findings with comprehensive experiments, confirming\nthe effectiveness of our proposed method in achieving fairness.\n", "date": "2024-11-14", "categories": []}
{"id": "2412.04914", "title": "Achieving Group Fairness through Independence in Predictive Process\n  Monitoring", "abstract": "  Predictive process monitoring focuses on forecasting future states of ongoing\nprocess executions, such as predicting the outcome of a particular case. In\nrecent years, the application of machine learning models in this domain has\ngarnered significant scientific attention. When using historical execution\ndata, which may contain biases or exhibit unfair behavior, these biases may be\nencoded into the trained models. Consequently, when such models are deployed to\nmake decisions or guide interventions for new cases, they risk perpetuating\nthis unwanted behavior. This work addresses group fairness in predictive\nprocess monitoring by investigating independence, i.e. ensuring predictions are\nunaffected by sensitive group membership. We explore independence through\nmetrics for demographic parity such as $\\Delta$DP, as well as recently\nintroduced, threshold-independent distribution-based alternatives.\nAdditionally, we propose a composite loss function existing of binary\ncross-entropy and a distribution-based loss (Wasserstein) to train models that\nbalance predictive performance and fairness, and allow for customizable\ntrade-offs. The effectiveness of both the fairness metrics and the composite\nloss functions is validated through a controlled experimental setup.\n", "date": "2024-12-06", "categories": []}
{"id": "2504.19370", "title": "Mitigating Bias in Facial Recognition Systems: Centroid Fairness Loss\n  Optimization", "abstract": "  The urging societal demand for fair AI systems has put pressure on the\nresearch community to develop predictive models that are not only globally\naccurate but also meet new fairness criteria, reflecting the lack of disparate\nmistreatment with respect to sensitive attributes ($\\textit{e.g.}$ gender,\nethnicity, age). In particular, the variability of the errors made by certain\nFacial Recognition (FR) systems across specific segments of the population\ncompromises the deployment of the latter, and was judged unacceptable by\nregulatory authorities. Designing fair FR systems is a very challenging\nproblem, mainly due to the complex and functional nature of the performance\nmeasure used in this domain ($\\textit{i.e.}$ ROC curves) and because of the\nhuge heterogeneity of the face image datasets usually available for training.\nIn this paper, we propose a novel post-processing approach to improve the\nfairness of pre-trained FR models by optimizing a regression loss which acts on\ncentroid-based scores. Beyond the computational advantages of the method, we\npresent numerical experiments providing strong empirical evidence of the gain\nin fairness and of the ability to preserve global accuracy.\n", "date": "2025-04-27", "categories": []}
{"id": "2504.19621", "title": "AI Alignment in Medical Imaging: Unveiling Hidden Biases Through\n  Counterfactual Analysis", "abstract": "  Machine learning (ML) systems for medical imaging have demonstrated\nremarkable diagnostic capabilities, but their susceptibility to biases poses\nsignificant risks, since biases may negatively impact generalization\nperformance. In this paper, we introduce a novel statistical framework to\nevaluate the dependency of medical imaging ML models on sensitive attributes,\nsuch as demographics. Our method leverages the concept of counterfactual\ninvariance, measuring the extent to which a model's predictions remain\nunchanged under hypothetical changes to sensitive attributes. We present a\npractical algorithm that combines conditional latent diffusion models with\nstatistical hypothesis testing to identify and quantify such biases without\nrequiring direct access to counterfactual data. Through experiments on\nsynthetic datasets and large-scale real-world medical imaging datasets,\nincluding \\textsc{cheXpert} and MIMIC-CXR, we demonstrate that our approach\naligns closely with counterfactual fairness principles and outperforms standard\nbaselines. This work provides a robust tool to ensure that ML diagnostic\nsystems generalize well, e.g., across demographic groups, offering a critical\nstep towards AI safety in healthcare. Code:\nhttps://github.com/Neferpitou3871/AI-Alignment-Medical-Imaging.\n", "date": "2025-04-28", "categories": []}
{"id": "2504.18262", "title": "Local Statistical Parity for the Estimation of Fair Decision Trees", "abstract": "  Given the high computational complexity of decision tree estimation,\nclassical methods construct a tree by adding one node at a time in a recursive\nway. To facilitate promoting fairness, we propose a fairness criterion local to\nthe tree nodes. We prove how it is related to the Statistical Parity criterion,\npopular in the Algorithmic Fairness literature, and show how to incorporate it\ninto standard recursive tree estimation algorithms.\n  We present a tree estimation algorithm called Constrained Logistic Regression\nTree (C-LRT), which is a modification of the standard CART algorithm using\nlocally linear classifiers and imposing restrictions as done in Constrained\nLogistic Regression.\n  Finally, we evaluate the performance of trees estimated with C-LRT on\ndatasets commonly used in the Algorithmic Fairness literature, using various\nclassification and fairness metrics. The results confirm that C-LRT\nsuccessfully allows to control and balance accuracy and fairness.\n", "date": "2025-04-25", "categories": []}
{"id": "2504.18353", "title": "Testing Individual Fairness in Graph Neural Networks", "abstract": "  The biases in artificial intelligence (AI) models can lead to automated\ndecision-making processes that discriminate against groups and/or individuals\nbased on sensitive properties such as gender and race. While there are many\nstudies on diagnosing and mitigating biases in various AI models, there is\nlittle research on individual fairness in Graph Neural Networks (GNNs). Unlike\ntraditional models, which treat data features independently and overlook their\ninter-relationships, GNNs are designed to capture graph-based structure where\nnodes are interconnected. This relational approach enables GNNs to model\ncomplex dependencies, but it also means that biases can propagate through these\nconnections, complicating the detection and mitigation of individual fairness\nviolations. This PhD project aims to develop a testing framework to assess and\nensure individual fairness in GNNs. It first systematically reviews the\nliterature on individual fairness, categorizing existing approaches to define,\nmeasure, test, and mitigate model biases, creating a taxonomy of individual\nfairness. Next, the project will develop a framework for testing and ensuring\nfairness in GNNs by adapting and extending current fairness testing and\nmitigation techniques. The framework will be evaluated through industrial case\nstudies, focusing on graph-based large language models.\n", "date": "2025-04-25", "categories": []}
{"id": "2504.17279", "title": "Evaluating and Mitigating Bias in AI-Based Medical Text Generation", "abstract": "  Artificial intelligence (AI) systems, particularly those based on deep\nlearning models, have increasingly achieved expert-level performance in medical\napplications. However, there is growing concern that such AI systems may\nreflect and amplify human bias, and reduce the quality of their performance in\nhistorically under-served populations. The fairness issue has attracted\nconsiderable research interest in the medical imaging classification field, yet\nit remains understudied in the text generation domain. In this study, we\ninvestigate the fairness problem in text generation within the medical field\nand observe significant performance discrepancies across different races,\nsexes, and age groups, including intersectional groups, various model scales,\nand different evaluation metrics. To mitigate this fairness issue, we propose\nan algorithm that selectively optimizes those underperformed groups to reduce\nbias. The selection rules take into account not only word-level accuracy but\nalso the pathology accuracy to the target reference, while ensuring that the\nentire process remains fully differentiable for effective model training. Our\nevaluations across multiple backbones, datasets, and modalities demonstrate\nthat our proposed algorithm enhances fairness in text generation without\ncompromising overall performance. Specifically, the disparities among various\ngroups across different metrics were diminished by more than 30% with our\nalgorithm, while the relative change in text generation accuracy was typically\nwithin 2%. By reducing the bias generated by deep learning models, our proposed\napproach can potentially alleviate concerns about the fairness and reliability\nof text generation diagnosis in medical domain.\n  Our code is publicly available to facilitate further research at\nhttps://github.com/iriscxy/GenFair.\n", "date": "2025-04-24", "categories": []}
{"id": "2504.16238", "title": "General Post-Processing Framework for Fairness Adjustment of Machine\n  Learning Models", "abstract": "  As machine learning increasingly influences critical domains such as credit\nunderwriting, public policy, and talent acquisition, ensuring compliance with\nfairness constraints is both a legal and ethical imperative. This paper\nintroduces a novel framework for fairness adjustments that applies to diverse\nmachine learning tasks, including regression and classification, and\naccommodates a wide range of fairness metrics. Unlike traditional approaches\ncategorized as pre-processing, in-processing, or post-processing, our method\nadapts in-processing techniques for use as a post-processing step. By\ndecoupling fairness adjustments from the model training process, our framework\npreserves model performance on average while enabling greater flexibility in\nmodel development. Key advantages include eliminating the need for custom loss\nfunctions, enabling fairness tuning using different datasets, accommodating\nproprietary models as black-box systems, and providing interpretable insights\ninto the fairness adjustments. We demonstrate the effectiveness of this\napproach by comparing it to Adversarial Debiasing, showing that our framework\nachieves a comparable fairness/accuracy tradeoff on real-world datasets.\n", "date": "2025-04-22", "categories": []}
{"id": "2504.16255", "title": "FairPlay: A Collaborative Approach to Mitigate Bias in Datasets for\n  Improved AI Fairness", "abstract": "  The issue of fairness in decision-making is a critical one, especially given\nthe variety of stakeholder demands for differing and mutually incompatible\nversions of fairness. Adopting a strategic interaction of perspectives provides\nan alternative to enforcing a singular standard of fairness. We present a\nweb-based software application, FairPlay, that enables multiple stakeholders to\ndebias datasets collaboratively. With FairPlay, users can negotiate and arrive\nat a mutually acceptable outcome without a universally agreed-upon theory of\nfairness. In the absence of such a tool, reaching a consensus would be highly\nchallenging due to the lack of a systematic negotiation process and the\ninability to modify and observe changes. We have conducted user studies that\ndemonstrate the success of FairPlay, as users could reach a consensus within\nabout five rounds of gameplay, illustrating the application's potential for\nenhancing fairness in AI systems.\n", "date": "2025-04-22", "categories": []}
{"id": "2405.18780", "title": "Certifying Counterfactual Bias in LLMs", "abstract": "  Large Language Models (LLMs) can produce biased responses that can cause\nrepresentational harms. However, conventional studies are insufficient to\nthoroughly evaluate biases across LLM responses for different demographic\ngroups (a.k.a. counterfactual bias), as they do not scale to large number of\ninputs and do not provide guarantees. Therefore, we propose the first\nframework, LLMCert-B that certifies LLMs for counterfactual bias on\ndistributions of prompts. A certificate consists of high-confidence bounds on\nthe probability of unbiased LLM responses for any set of counterfactual prompts\n- prompts differing by demographic groups, sampled from a distribution. We\nillustrate counterfactual bias certification for distributions of\ncounterfactual prompts created by applying prefixes sampled from prefix\ndistributions, to a given set of prompts. We consider prefix distributions\nconsisting random token sequences, mixtures of manual jailbreaks, and\nperturbations of jailbreaks in LLM's embedding space. We generate non-trivial\ncertificates for SOTA LLMs, exposing their vulnerabilities over distributions\nof prompts generated from computationally inexpensive prefix distributions.\n", "date": "2024-05-29", "categories": []}
{"id": "2408.16683", "title": "A Catalog of Fairness-Aware Practices in Machine Learning Engineering", "abstract": "  Machine learning's widespread adoption in decision-making processes raises\nconcerns about fairness, particularly regarding the treatment of sensitive\nfeatures and potential discrimination against minorities. The software\nengineering community has responded by developing fairness-oriented metrics,\nempirical studies, and approaches. However, there remains a gap in\nunderstanding and categorizing practices for engineering fairness throughout\nthe machine learning lifecycle. This paper presents a novel catalog of\npractices for addressing fairness in machine learning derived from a systematic\nmapping study. The study identifies and categorizes 28 practices from existing\nliterature, mapping them onto different stages of the machine learning\nlifecycle. From this catalog, the authors extract actionable items and\nimplications for both researchers and practitioners in software engineering.\nThis work aims to provide a comprehensive resource for integrating fairness\nconsiderations into the development and deployment of machine learning systems,\nenhancing their reliability, accountability, and credibility.\n", "date": "2024-08-29", "categories": []}
{"id": "2504.15924", "title": "Achieving Distributive Justice in Federated Learning via Uncertainty\n  Quantification", "abstract": "  Client-level fairness metrics for federated learning are used to ensure that\nall clients in a federation either: a) have similar final performance on their\nlocal data distributions (i.e., client parity), or b) obtain final performance\non their local data distributions relative to their contribution to the\nfederated learning process (i.e., contribution fairness). While a handful of\nworks that propose either client-parity or contribution-based fairness metrics\nground their definitions and decisions in social theories of equality -- such\nas distributive justice -- most works arbitrarily choose what notion of\nfairness to align with which makes it difficult for practitioners to choose\nwhich fairness metric aligns best with their fairness ethics. In this work, we\npropose UDJ-FL (Uncertainty-based Distributive Justice for Federated Learning),\na flexible federated learning framework that can achieve multiple distributive\njustice-based client-level fairness metrics. Namely, by utilizing techniques\ninspired by fair resource allocation, in conjunction with performing aleatoric\nuncertainty-based client weighing, our UDJ-FL framework is able to achieve\negalitarian, utilitarian, Rawls' difference principle, or desert-based\nclient-level fairness. We empirically show the ability of UDJ-FL to achieve all\nfour defined distributive justice-based client-level fairness metrics in\naddition to providing fairness equivalent to (or surpassing) other popular fair\nfederated learning works. Further, we provide justification for why aleatoric\nuncertainty weighing is necessary to the construction of our UDJ-FL framework\nas well as derive theoretical guarantees for the generalization bounds of\nUDJ-FL. Our code is publicly available at\nhttps://github.com/alycia-noel/UDJ-FL.\n", "date": "2025-04-22", "categories": []}
{"id": "2411.04761", "title": "Mining the Minoria: Unknown, Under-represented, and Under-performing\n  Minority Groups", "abstract": "  Due to a variety of reasons, such as privacy, data in the wild often misses\nthe grouping information required for identifying minorities. On the other\nhand, it is known that machine learning models are only as good as the data\nthey are trained on and, hence, may underperform for the under-represented\nminority groups. The missing grouping information presents a dilemma for\nresponsible data scientists who find themselves in an unknown-unknown\nsituation, where not only do they not have access to the grouping attributes\nbut do not also know what groups to consider.\n  This paper is an attempt to address this dilemma. Specifically, we propose a\nminority mining problem, where we find vectors in the attribute space that\nreveal potential groups that are under-represented and under-performing.\nTechnically speaking, we propose a geometric transformation of data into a dual\nspace and use notions such as the arrangement of hyperplanes to design an\nefficient algorithm for the problem in lower dimensions. Generalizing our\nsolution to the higher dimensions is cursed by dimensionality. Therefore, we\npropose a solution based on smart exploration of the search space for such\ncases. We conduct comprehensive experiments using real-world and synthetic\ndatasets alongside the theoretical analysis. Our experiment results demonstrate\nthe effectiveness of our proposed solutions in mining the unknown,\nunder-represented, and under-performing minorities.\n", "date": "2024-11-07", "categories": []}
{"id": "2501.04683", "title": "Toward Sufficient Statistical Power in Algorithmic Bias Assessment: A\n  Test for ABROCA", "abstract": "  Algorithmic bias is a pressing concern in educational data mining (EDM), as\nit risks amplifying inequities in learning outcomes. The Area Between ROC\nCurves (ABROCA) metric is frequently used to measure discrepancies in model\nperformance across demographic groups to quantify overall model fairness.\nHowever, its skewed distribution--especially when class or group imbalances\nexist--makes significance testing challenging. This study investigates ABROCA's\ndistributional properties and contributes robust methods for its significance\ntesting. Specifically, we address (1) whether ABROCA follows any known\ndistribution, (2) how to reliably test for algorithmic bias using ABROCA, and\n(3) the statistical power achievable with ABROCA-based bias assessments under\ntypical EDM sample specifications. Simulation results confirm that ABROCA does\nnot match standard distributions, including those suited to accommodate\nskewness. We propose nonparametric randomization tests for ABROCA and\ndemonstrate that reliably detecting bias with ABROCA requires large sample\nsizes or substantial effect sizes, particularly in imbalanced settings.\nFindings suggest that ABROCA-based bias evaluations based on sample sizes\ncommon in EDM tend to be underpowered, undermining the reliability of\nconclusions about model fairness. By offering open-source code to simulate\npower and statistically test ABROCA, this paper aims to foster more reliable\nstatistical testing in EDM research. It supports broader efforts toward\nreplicability and equity in educational modeling.\n", "date": "2025-01-08", "categories": []}
{"id": "2504.11504", "title": "Counterfactual Fairness Evaluation of Machine Learning Models on\n  Educational Datasets", "abstract": "  As machine learning models are increasingly used in educational settings,\nfrom detecting at-risk students to predicting student performance, algorithmic\nbias and its potential impacts on students raise critical concerns about\nalgorithmic fairness. Although group fairness is widely explored in education,\nworks on individual fairness in a causal context are understudied, especially\non counterfactual fairness. This paper explores the notion of counterfactual\nfairness for educational data by conducting counterfactual fairness analysis of\nmachine learning models on benchmark educational datasets. We demonstrate that\ncounterfactual fairness provides meaningful insight into the causality of\nsensitive attributes and causal-based individual fairness in education.\n", "date": "2025-04-15", "categories": []}
{"id": "2504.12323", "title": "The Other Side of the Coin: Exploring Fairness in Retrieval-Augmented\n  Generation", "abstract": "  Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by\nretrieving relevant document from external knowledge sources. By referencing\nthis external knowledge, RAG effectively reduces the generation of factually\nincorrect content and addresses hallucination issues within LLMs. Recently,\nthere has been growing attention to improving the performance and efficiency of\nRAG systems from various perspectives. While these advancements have yielded\nsignificant results, the application of RAG in domains with considerable\nsocietal implications raises a critical question about fairness: What impact\ndoes the introduction of the RAG paradigm have on the fairness of LLMs? To\naddress this question, we conduct extensive experiments by varying the LLMs,\nretrievers, and retrieval sources. Our experimental analysis reveals that the\nscale of the LLMs plays a significant role in influencing fairness outcomes\nwithin the RAG framework. When the model scale is smaller than 8B, the\nintegration of retrieval mechanisms often exacerbates unfairness in small-scale\nLLMs (e.g., LLaMA3.2-1B, Mistral-7B, and LLaMA3-8B). To mitigate the fairness\nissues introduced by RAG for small-scale LLMs, we propose two approaches,\nFairFT and FairFilter. Specifically, in FairFT, we align the retriever with the\nLLM in terms of fairness, enabling it to retrieve documents that facilitate\nfairer model outputs. In FairFilter, we propose a fairness filtering mechanism\nto filter out biased content after retrieval. Finally, we validate our proposed\napproaches on real-world datasets, demonstrating their effectiveness in\nimproving fairness while maintaining performance.\n", "date": "2025-04-11", "categories": []}
{"id": "2504.14388", "title": "Balancing Fairness and Performance in Healthcare AI: A Gradient\n  Reconciliation Approach", "abstract": "  The rapid growth of healthcare data and advances in computational power have\naccelerated the adoption of artificial intelligence (AI) in medicine. However,\nAI systems deployed without explicit fairness considerations risk exacerbating\nexisting healthcare disparities, potentially leading to inequitable resource\nallocation and diagnostic disparities across demographic subgroups. To address\nthis challenge, we propose FairGrad, a novel gradient reconciliation framework\nthat automatically balances predictive performance and multi-attribute fairness\noptimization in healthcare AI models. Our method resolves conflicting\noptimization objectives by projecting each gradient vector onto the orthogonal\nplane of the others, thereby regularizing the optimization trajectory to ensure\nequitable consideration of all objectives. Evaluated on diverse real-world\nhealthcare datasets and predictive tasks - including Substance Use Disorder\n(SUD) treatment and sepsis mortality - FairGrad achieved statistically\nsignificant improvements in multi-attribute fairness metrics (e.g., equalized\nodds) while maintaining competitive predictive accuracy. These results\ndemonstrate the viability of harmonizing fairness and utility in\nmission-critical medical AI applications.\n", "date": "2025-04-19", "categories": []}
{"id": "2209.01410", "title": "Closed-Loop View of the Regulation of AI: Equal Impact across Repeated\n  Interactions", "abstract": "  There has been much recent interest in the regulation of AI. We argue for a\nview based on civil-rights legislation, built on the notions of equal treatment\nand equal impact. In a closed-loop view of the AI system and its users, the\nequal treatment concerns one pass through the loop. Equal impact, in our view,\nconcerns the long-run average behaviour across repeated interactions. In order\nto establish the existence of the average and its properties, one needs to\nstudy the ergodic properties of the closed-loop and its unique stationary\nmeasure.\n", "date": "2022-09-03", "categories": []}
{"id": "2403.19419", "title": "Fairness in Ranking: Robustness through Randomization without the\n  Protected Attribute", "abstract": "  There has been great interest in fairness in machine learning, especially in\nrelation to classification problems. In ranking-related problems, such as in\nonline advertising, recommender systems, and HR automation, much work on\nfairness remains to be done. Two complications arise: first, the protected\nattribute may not be available in many applications. Second, there are multiple\nmeasures of fairness of rankings, and optimization-based methods utilizing a\nsingle measure of fairness of rankings may produce rankings that are unfair\nwith respect to other measures. In this work, we propose a randomized method\nfor post-processing rankings, which do not require the availability of the\nprotected attribute. In an extensive numerical study, we show the robustness of\nour methods with respect to P-Fairness and effectiveness with respect to\nNormalized Discounted Cumulative Gain (NDCG) from the baseline ranking,\nimproving on previously proposed methods.\n", "date": "2024-03-28", "categories": []}
{"id": "2504.12458", "title": "M$^2$FGB: A Min-Max Gradient Boosting Framework for Subgroup Fairness", "abstract": "  In recent years, fairness in machine learning has emerged as a critical\nconcern to ensure that developed and deployed predictive models do not have\ndisadvantageous predictions for marginalized groups. It is essential to\nmitigate discrimination against individuals based on protected attributes such\nas gender and race. In this work, we consider applying subgroup justice\nconcepts to gradient-boosting machines designed for supervised learning\nproblems. Our approach expanded gradient-boosting methodologies to explore a\nbroader range of objective functions, which combines conventional losses such\nas the ones from classification and regression and a min-max fairness term. We\nstudy relevant theoretical properties of the solution of the min-max\noptimization problem. The optimization process explored the primal-dual\nproblems at each boosting round. This generic framework can be adapted to\ndiverse fairness concepts. The proposed min-max primal-dual gradient boosting\nalgorithm was theoretically shown to converge under mild conditions and\nempirically shown to be a powerful and flexible approach to address binary and\nsubgroup fairness.\n", "date": "2025-04-16", "categories": []}
{"id": "2504.12587", "title": "Software Engineering Principles for Fairer Systems: Experiments with\n  GroupCART", "abstract": "  Discrimination-aware classification aims to make accurate predictions while\nsatisfying fairness constraints. Traditional decision tree learners typically\noptimize for information gain in the target attribute alone, which can result\nin models that unfairly discriminate against protected social groups (e.g.,\ngender, ethnicity). Motivated by these shortcomings, we propose GroupCART, a\ntree-based ensemble optimizer that avoids bias during model construction by\noptimizing not only for decreased entropy in the target attribute but also for\nincreased entropy in protected attributes. Our experiments show that GroupCART\nachieves fairer models without data transformation and with minimal performance\ndegradation. Furthermore, the method supports customizable weighting, offering\na smooth and flexible trade-off between predictive performance and fairness\nbased on user requirements. These results demonstrate that algorithmic bias in\ndecision tree models can be mitigated through multi-task, fairness-aware\nlearning. All code and datasets used in this study are available at:\nhttps://github.com/anonymous12138/groupCART.\n", "date": "2025-04-17", "categories": []}
{"id": "2504.12767", "title": "Out of Sight Out of Mind, Out of Sight Out of Mind: Measuring Bias in\n  Language Models Against Overlooked Marginalized Groups in Regional Contexts", "abstract": "  We know that language models (LMs) form biases and stereotypes of minorities,\nleading to unfair treatments of members of these groups, thanks to research\nmainly in the US and the broader English-speaking world. As the negative\nbehavior of these models has severe consequences for society and individuals,\nindustry and academia are actively developing methods to reduce the bias in\nLMs. However, there are many under-represented groups and languages that have\nbeen overlooked so far. This includes marginalized groups that are specific to\nindividual countries and regions in the English speaking and Western world, but\ncrucially also almost all marginalized groups in the rest of the world. The UN\nestimates, that between 600 million to 1.2 billion people worldwide are members\nof marginalized groups and in need for special protection. If we want to\ndevelop inclusive LMs that work for everyone, we have to broaden our\nunderstanding to include overlooked marginalized groups and low-resource\nlanguages and dialects.\n  In this work, we contribute to this effort with the first study investigating\noffensive stereotyping bias in 23 LMs for 270 marginalized groups from Egypt,\nthe remaining 21 Arab countries, Germany, the UK, and the US. Additionally, we\ninvestigate the impact of low-resource languages and dialects on the study of\nbias in LMs, demonstrating the limitations of current bias metrics, as we\nmeasure significantly higher bias when using the Egyptian Arabic dialect versus\nModern Standard Arabic. Our results show, LMs indeed show higher bias against\nmany marginalized groups in comparison to dominant groups. However, this is not\nthe case for Arabic LMs, where the bias is high against both marginalized and\ndominant groups in relation to religion and ethnicity.\n  Our results also show higher intersectional bias against Non-binary, LGBTQIA+\nand Black women.\n", "date": "2025-04-17", "categories": []}
{"id": "2302.05508", "title": "FairPy: A Toolkit for Evaluation of Prediction Biases and their\n  Mitigation in Large Language Models", "abstract": "  Recent studies have demonstrated that large pretrained language models (LLMs)\nsuch as BERT and GPT-2 exhibit biases in token prediction, often inherited from\nthe data distributions present in their training corpora. In response, a number\nof mathematical frameworks have been proposed to quantify, identify, and\nmitigate these the likelihood of biased token predictions. In this paper, we\npresent a comprehensive survey of such techniques tailored towards widely used\nLLMs such as BERT, GPT-2, etc. We additionally introduce Fairpy, a modular and\nextensible toolkit that provides plug-and-play interfaces for integrating these\nmathematical tools, enabling users to evaluate both pretrained and custom\nlanguage models. Fairpy supports the implementation of existing debiasing\nalgorithms. The toolkit is open-source and publicly available at:\n\\href{https://github.com/HrishikeshVish/Fairpy}{https://github.com/HrishikeshVish/Fairpy}\n", "date": "2023-02-10", "categories": []}
{"id": "2412.10575", "title": "Who's the (Multi-)Fairest of Them All: Rethinking Interpolation-Based\n  Data Augmentation Through the Lens of Multicalibration", "abstract": "  Data augmentation methods, especially SoTA interpolation-based methods such\nas Fair Mixup, have been widely shown to increase model fairness. However, this\nfairness is evaluated on metrics that do not capture model uncertainty and on\ndatasets with only one, relatively large, minority group. As a remedy,\nmulticalibration has been introduced to measure fairness while accommodating\nuncertainty and accounting for multiple minority groups. However, existing\nmethods of improving multicalibration involve reducing initial training data to\ncreate a holdout set for post-processing, which is not ideal when minority\ntraining data is already sparse. This paper uses multicalibration to more\nrigorously examine data augmentation for classification fairness. We\nstress-test four versions of Fair Mixup on two structured data classification\nproblems with up to 81 marginalized groups, evaluating multicalibration\nviolations and balanced accuracy. We find that on nearly every experiment, Fair\nMixup \\textit{worsens} baseline performance and fairness, but the simple\nvanilla Mixup \\textit{outperforms} both Fair Mixup and the baseline, especially\nwhen calibrating on small groups. \\textit{Combining} vanilla Mixup with\nmulticalibration post-processing, which enforces multicalibration through\npost-processing on a holdout set, further increases fairness.\n", "date": "2024-12-13", "categories": []}
{"id": "2504.09210", "title": "FairACE: Achieving Degree Fairness in Graph Neural Networks via\n  Contrastive and Adversarial Group-Balanced Training", "abstract": "  Fairness has been a significant challenge in graph neural networks (GNNs)\nsince degree biases often result in un-equal prediction performance among nodes\nwith varying degrees. Existing GNN models focus on prediction accuracy,\nfrequently overlooking fairness across different degree groups. To addressthis\nissue, we propose a novel GNN framework, namely Fairness- Aware Asymmetric\nContrastive Ensemble (FairACE), which inte-grates asymmetric contrastive\nlearning with adversarial training to improve degree fairness. FairACE captures\none-hop local neighborhood information and two-hop monophily similarity to\ncreate fairer node representations and employs a degree fairness regulator to\nbalance performance between high-degree and low-degree nodes. During model\ntraining, a novel group-balanced fairness loss is proposed to minimize\nclassification disparities across degree groups. In addition, we also propose a\nnovel fairness metric, the Accuracy Distribution Gap (ADG), which can\nquantitatively assess and ensure equitable performance across different\ndegree-based node groups. Experimental results on both synthetic and real-world\ndatasets demonstrate that FairACE significantly improves degree fairness\nmetrics while maintaining competitive accuracy in comparison to the\nstate-of-the-art GNN models.\n", "date": "2025-04-12", "categories": []}
{"id": "2504.10797", "title": "Name of Thrones: Evaluating How LLMs Rank Student Names, Race, and\n  Gender in Status Hierarchies", "abstract": "  Across cultures, names tell a lot about their bearers as they carry deep\npersonal and cultural significance. Names also serve as powerful signals of\ngender, race, and status in the social hierarchy - a pecking order in which\nindividual positions shape others' expectations on their perceived competence\nand worth. With the widespread adoption of LLMs and as names are often an input\nfor LLMs, it is crucial to evaluate whether LLMs may sort people into status\npositions based on first and last names and, if so, whether it is in an unfair,\nbiased fashion. While prior work has primarily investigated biases in first\nnames, little attention has been paid to last names and even less to the\ncombined effects of first and last names. In this study, we conduct a\nlarge-scale analysis of name variations across 5 ethnicities to examine how AI\nexhibits name biases. Our study investigates three key characteristics of\ninequality and finds that LLMs reflect and reinforce status hierarchies based\non names that signal gender and ethnicity as they encode differential\nexpectations of competence, leadership, and economic potential. Contrary to the\ncommon assumption that AI tends to favor Whites, we show that East and, in some\ncontexts, South Asian names receive higher rankings. We also disaggregate\nAsians, a population projected to be the largest immigrant group in the U.S. by\n2055. Our results challenge the monolithic Asian model minority assumption,\nillustrating a more complex and stratified model of bias. Gender moderates\nbiases, with girls facing unfair disadvantages in certain racial groups.\nAdditionally, spanning cultural categories by adopting Western first names\nimproves AI-perceived status for East and Southeast Asian students,\nparticularly for girls. Our findings underscore the importance of\nintersectional and more nuanced understandings of race, gender, and mixed\nidentities in the evaluation of LLMs.\n", "date": "2025-04-15", "categories": []}
{"id": "2504.11104", "title": "Using LLMs as prompt modifier to avoid biases in AI image generators", "abstract": "  This study examines how Large Language Models (LLMs) can reduce biases in\ntext-to-image generation systems by modifying user prompts. We define bias as a\nmodel's unfair deviation from population statistics given neutral prompts. Our\nexperiments with Stable Diffusion XL, 3.5 and Flux demonstrate that\nLLM-modified prompts significantly increase image diversity and reduce bias\nwithout the need to change the image generators themselves. While occasionally\nproducing results that diverge from original user intent for elaborate prompts,\nthis approach generally provides more varied interpretations of underspecified\nrequests rather than superficial variations. The method works particularly well\nfor less advanced image generators, though limitations persist for certain\ncontexts like disability representation. All prompts and generated images are\navailable at https://iisys-hof.github.io/llm-prompt-img-gen/\n", "date": "2025-04-15", "categories": []}
{"id": "2504.11431", "title": "Masculine Defaults via Gendered Discourse in Podcasts and Large Language\n  Models", "abstract": "  Masculine defaults are widely recognized as a significant type of gender\nbias, but they are often unseen as they are under-researched. Masculine\ndefaults involve three key parts: (i) the cultural context, (ii) the masculine\ncharacteristics or behaviors, and (iii) the reward for, or simply acceptance\nof, those masculine characteristics or behaviors. In this work, we study\ndiscourse-based masculine defaults, and propose a twofold framework for (i) the\nlarge-scale discovery and analysis of gendered discourse words in spoken\ncontent via our Gendered Discourse Correlation Framework (GDCF); and (ii) the\nmeasurement of the gender bias associated with these gendered discourse words\nin LLMs via our Discourse Word-Embedding Association Test (D-WEAT). We focus\nour study on podcasts, a popular and growing form of social media, analyzing\n15,117 podcast episodes. We analyze correlations between gender and discourse\nwords -- discovered via LDA and BERTopic -- to automatically form gendered\ndiscourse word lists. We then study the prevalence of these gendered discourse\nwords in domain-specific contexts, and find that gendered discourse-based\nmasculine defaults exist in the domains of business, technology/politics, and\nvideo games. Next, we study the representation of these gendered discourse\nwords from a state-of-the-art LLM embedding model from OpenAI, and find that\nthe masculine discourse words have a more stable and robust representation than\nthe feminine discourse words, which may result in better system performance on\ndownstream tasks for men. Hence, men are rewarded for their discourse patterns\nwith better system performance by one of the state-of-the-art language models\n-- and this embedding disparity is a representational harm and a masculine\ndefault.\n", "date": "2025-04-15", "categories": []}
{"id": "2410.14763", "title": "Enabling Scalable Evaluation of Bias Patterns in Medical LLMs", "abstract": "  Large language models (LLMs) have shown impressive potential in helping with\nnumerous medical challenges. Deploying LLMs in high-stakes applications such as\nmedicine, however, brings in many concerns. One major area of concern relates\nto biased behaviors of LLMs in medical applications, leading to unfair\ntreatment of individuals. To pave the way for the responsible and impactful\ndeployment of Med LLMs, rigorous evaluation is a key prerequisite. Due to the\nhuge complexity and variability of different medical scenarios, existing work\nin this domain has primarily relied on using manually crafted datasets for bias\nevaluation. In this study, we present a new method to scale up such bias\nevaluations by automatically generating test cases based on rigorous medical\nevidence. We specifically target the challenges of a) domain-specificity of\nbias characterization, b) hallucinating while generating the test cases, and c)\nvarious dependencies between the health outcomes and sensitive attributes. To\nthat end, we offer new methods to address these challenges integrated with our\ngenerative pipeline, using medical knowledge graphs, medical ontologies, and\ncustomized general LLM evaluation frameworks in our method. Through a series of\nextensive experiments, we show that the test cases generated by our proposed\nmethod can effectively reveal bias patterns in Med LLMs at larger and more\nflexible scales than human-crafted datasets. We publish a large bias evaluation\ndataset using our pipeline, which is dedicated to a few medical case studies. A\nlive demo of our application for vignette generation is available at\nhttps://vignette.streamlit.app. Our code is also available at\nhttps://github.com/healthylaife/autofair.\n", "date": "2024-10-18", "categories": []}
{"id": "2503.18826", "title": "Interpretable and Fair Mechanisms for Abstaining Classifiers", "abstract": "  Abstaining classifiers have the option to refrain from providing a prediction\nfor instances that are difficult to classify. The abstention mechanism is\ndesigned to trade off the classifier's performance on the accepted data while\nensuring a minimum number of predictions. In this setting, often fairness\nconcerns arise when the abstention mechanism solely reduces errors for the\nmajority groups of the data, resulting in increased performance differences\nacross demographic groups. While there exist a bunch of methods that aim to\nreduce discrimination when abstaining, there is no mechanism that can do so in\nan explainable way. In this paper, we fill this gap by introducing\nInterpretable and Fair Abstaining Classifier IFAC, an algorithm that can reject\npredictions both based on their uncertainty and their unfairness. By rejecting\npossibly unfair predictions, our method reduces error and positive decision\nrate differences across demographic groups of the non-rejected data. Since the\nunfairness-based rejections are based on an interpretable-by-design method,\ni.e., rule-based fairness checks and situation testing, we create a transparent\nprocess that can empower human decision-makers to review the unfair predictions\nand make more just decisions for them. This explainable aspect is especially\nimportant in light of recent AI regulations, mandating that any high-risk\ndecision task should be overseen by human experts to reduce discrimination\nrisks.\n", "date": "2025-03-24", "categories": []}
{"id": "2407.14367", "title": "Thinking Racial Bias in Fair Forgery Detection: Models, Datasets and\n  Evaluations", "abstract": "  Due to the successful development of deep image generation technology,\nforgery detection plays a more important role in social and economic security.\nRacial bias has not been explored thoroughly in the deep forgery detection\nfield. In the paper, we first contribute a dedicated dataset called the Fair\nForgery Detection (FairFD) dataset, where we prove the racial bias of public\nstate-of-the-art (SOTA) methods. Different from existing forgery detection\ndatasets, the self-constructed FairFD dataset contains a balanced racial ratio\nand diverse forgery generation images with the largest-scale subjects.\nAdditionally, we identify the problems with naive fairness metrics when\nbenchmarking forgery detection models. To comprehensively evaluate fairness, we\ndesign novel metrics including Approach Averaged Metric and Utility Regularized\nMetric, which can avoid deceptive results. We also present an effective and\nrobust post-processing technique, Bias Pruning with Fair Activations (BPFA),\nwhich improves fairness without requiring retraining or weight updates.\nExtensive experiments conducted with 12 representative forgery detection models\ndemonstrate the value of the proposed dataset and the reasonability of the\ndesigned fairness metrics. By applying the BPFA to the existing fairest\ndetector, we achieve a new SOTA. Furthermore, we conduct more in-depth analyses\nto offer more insights to inspire researchers in the community.\n", "date": "2024-07-19", "categories": []}
{"id": "2504.07982", "title": "Metamorphic Testing for Fairness Evaluation in Large Language Models:\n  Identifying Intersectional Bias in LLaMA and GPT", "abstract": "  Large Language Models (LLMs) have made significant strides in Natural\nLanguage Processing but remain vulnerable to fairness-related issues, often\nreflecting biases inherent in their training data. These biases pose risks,\nparticularly when LLMs are deployed in sensitive areas such as healthcare,\nfinance, and law. This paper introduces a metamorphic testing approach to\nsystematically identify fairness bugs in LLMs. We define and apply a set of\nfairness-oriented metamorphic relations (MRs) to assess the LLaMA and GPT\nmodel, a state-of-the-art LLM, across diverse demographic inputs. Our\nmethodology includes generating source and follow-up test cases for each MR and\nanalyzing model responses for fairness violations. The results demonstrate the\neffectiveness of MT in exposing bias patterns, especially in relation to tone\nand sentiment, and highlight specific intersections of sensitive attributes\nthat frequently reveal fairness faults. This research improves fairness testing\nin LLMs, providing a structured approach to detect and mitigate biases and\nimprove model robustness in fairness-sensitive applications.\n", "date": "2025-04-04", "categories": []}
{"id": "2504.08151", "title": "Adaptive Bounded Exploration and Intermediate Actions for Data Debiasing", "abstract": "  The performance of algorithmic decision rules is largely dependent on the\nquality of training datasets available to them. Biases in these datasets can\nraise economic and ethical concerns due to the resulting algorithms' disparate\ntreatment of different groups. In this paper, we propose algorithms for\nsequentially debiasing the training dataset through adaptive and bounded\nexploration in a classification problem with costly and censored feedback. Our\nproposed algorithms balance between the ultimate goal of mitigating the impacts\nof data biases -- which will in turn lead to more accurate and fairer\ndecisions, and the exploration risks incurred to achieve this goal.\nSpecifically, we propose adaptive bounds to limit the region of exploration,\nand leverage intermediate actions which provide noisy label information at a\nlower cost. We analytically show that such exploration can help debias data in\ncertain distributions, investigate how {algorithmic fairness interventions} can\nwork in conjunction with our proposed algorithms, and validate the performance\nof these algorithms through numerical experiments on synthetic and real-world\ndata.\n", "date": "2025-04-10", "categories": []}
{"id": "2504.08418", "title": "seeBias: A Comprehensive Tool for Assessing and Visualizing AI Fairness", "abstract": "  Fairness in artificial intelligence (AI) prediction models is increasingly\nemphasized to support responsible adoption in high-stakes domains such as\nhealth care and criminal justice. Guidelines and implementation frameworks\nhighlight the importance of both predictive accuracy and equitable outcomes.\nHowever, current fairness toolkits often evaluate classification performance\ndisparities in isolation, with limited attention to other critical aspects such\nas calibration. To address these gaps, we present seeBias, an R package for\ncomprehensive evaluation of model fairness and predictive performance. seeBias\noffers an integrated evaluation across classification, calibration, and other\nperformance domains, providing a more complete view of model behavior. It\nincludes customizable visualizations to support transparent reporting and\nresponsible AI implementation. Using public datasets from criminal justice and\nhealthcare, we demonstrate how seeBias supports fairness evaluations, and\nuncovers disparities that conventional fairness metrics may overlook. The R\npackage is available on GitHub, and a Python version is under development.\n", "date": "2025-04-11", "categories": []}
{"id": "2402.18206", "title": "Balancing Act: Distribution-Guided Debiasing in Diffusion Models", "abstract": "  Diffusion Models (DMs) have emerged as powerful generative models with\nunprecedented image generation capability. These models are widely used for\ndata augmentation and creative applications. However, DMs reflect the biases\npresent in the training datasets. This is especially concerning in the context\nof faces, where the DM prefers one demographic subgroup vs others (eg. female\nvs male). In this work, we present a method for debiasing DMs without relying\non additional data or model retraining. Specifically, we propose Distribution\nGuidance, which enforces the generated images to follow the prescribed\nattribute distribution. To realize this, we build on the key insight that the\nlatent features of denoising UNet hold rich demographic semantics, and the same\ncan be leveraged to guide debiased generation. We train Attribute Distribution\nPredictor (ADP) - a small mlp that maps the latent features to the distribution\nof attributes. ADP is trained with pseudo labels generated from existing\nattribute classifiers. The proposed Distribution Guidance with ADP enables us\nto do fair generation. Our method reduces bias across single/multiple\nattributes and outperforms the baseline by a significant margin for\nunconditional and text-conditional diffusion models. Further, we present a\ndownstream task of training a fair attribute classifier by rebalancing the\ntraining set with our generated data.\n", "date": "2024-02-28", "categories": []}
{"id": "2504.07118", "title": "Sacred or Secular? Religious Bias in AI-Generated Financial Advice", "abstract": "  This study examines religious biases in AI-generated financial advice,\nfocusing on ChatGPT's responses to financial queries. Using a prompt-based\nmethodology and content analysis, we find that 50% of the financial emails\ngenerated by ChatGPT exhibit religious biases, with explicit biases present in\nboth ingroup and outgroup interactions. While ingroup biases personalize\nresponses based on religious alignment, outgroup biases introduce religious\nframing that may alienate clients or create ideological friction. These\nfindings align with broader research on AI bias and suggest that ChatGPT is not\nmerely reflecting societal biases but actively shaping financial discourse\nbased on perceived religious identity. Using the Critical Algorithm Studies\nframework, we argue that ChatGPT functions as a mediator of financial\nnarratives, selectively reinforcing religious perspectives. This study\nunderscores the need for greater transparency, bias mitigation strategies, and\nregulatory oversight to ensure neutrality in AI-driven financial services.\n", "date": "2025-03-26", "categories": []}
{"id": "2504.07395", "title": "FAIR-SIGHT: Fairness Assurance in Image Recognition via Simultaneous\n  Conformal Thresholding and Dynamic Output Repair", "abstract": "  We introduce FAIR-SIGHT, an innovative post-hoc framework designed to ensure\nfairness in computer vision systems by combining conformal prediction with a\ndynamic output repair mechanism. Our approach calculates a fairness-aware\nnon-conformity score that simultaneously assesses prediction errors and\nfairness violations. Using conformal prediction, we establish an adaptive\nthreshold that provides rigorous finite-sample, distribution-free guarantees.\nWhen the non-conformity score for a new image exceeds the calibrated threshold,\nFAIR-SIGHT implements targeted corrective adjustments, such as logit shifts for\nclassification and confidence recalibration for detection, to reduce both group\nand individual fairness disparities, all without the need for retraining or\nhaving access to internal model parameters. Comprehensive theoretical analysis\nvalidates our method's error control and convergence properties. At the same\ntime, extensive empirical evaluations on benchmark datasets show that\nFAIR-SIGHT significantly reduces fairness disparities while preserving high\npredictive performance.\n", "date": "2025-04-10", "categories": []}
{"id": "2504.07516", "title": "Enhancements for Developing a Comprehensive AI Fairness Assessment\n  Standard", "abstract": "  As AI systems increasingly influence critical sectors like\ntelecommunications, finance, healthcare, and public services, ensuring fairness\nin decision-making is essential to prevent biased or unjust outcomes that\ndisproportionately affect vulnerable entities or result in adverse impacts.\nThis need is particularly pressing as the industry approaches the 6G era, where\nAI will drive complex functions like autonomous network management and\nhyper-personalized services. The TEC Standard for Fairness Assessment and\nRating of AI Systems provides guidelines for evaluating fairness in AI,\nfocusing primarily on tabular data and supervised learning models. However, as\nAI applications diversify, this standard requires enhancement to strengthen its\nimpact and broaden its applicability. This paper proposes an expansion of the\nTEC Standard to include fairness assessments for images, unstructured text, and\ngenerative AI, including large language models, ensuring a more comprehensive\napproach that keeps pace with evolving AI technologies. By incorporating these\ndimensions, the enhanced framework will promote responsible and trustworthy AI\ndeployment across various sectors.\n", "date": "2025-04-10", "categories": []}
{"id": "2502.09765", "title": "Differential Adjusted Parity for Learning Fair Representations", "abstract": "  The development of fair and unbiased machine learning models remains an\nongoing objective for researchers in the field of artificial intelligence. We\nintroduce the Differential Adjusted Parity (DAP) loss to produce unbiased\ninformative representations. It utilises a differentiable variant of the\nadjusted parity metric to create a unified objective function. By combining\ndownstream task classification accuracy and its inconsistency across sensitive\nfeature domains, it provides a single tool to increase performance and mitigate\nbias. A key element in this approach is the use of soft balanced accuracies. In\ncontrast to previous non-adversarial approaches, DAP does not suffer a\ndegeneracy where the metric is satisfied by performing equally poorly across\nall sensitive domains. It outperforms several adversarial models on downstream\ntask accuracy and fairness in our analysis. Specifically, it improves the\ndemographic parity, equalized odds and sensitive feature accuracy by as much as\n22.5\\%, 44.1\\% and 40.1\\%, respectively, when compared to the best performing\nadversarial approaches on these metrics. Overall, the DAP loss and its\nassociated metric can play a significant role in creating more fair machine\nlearning models.\n", "date": "2025-02-13", "categories": []}
{"id": "2504.06470", "title": "Deep Fair Learning: A Unified Framework for Fine-tuning Representations\n  with Sufficient Networks", "abstract": "  Ensuring fairness in machine learning is a critical and challenging task, as\nbiased data representations often lead to unfair predictions. To address this,\nwe propose Deep Fair Learning, a framework that integrates nonlinear sufficient\ndimension reduction with deep learning to construct fair and informative\nrepresentations. By introducing a novel penalty term during fine-tuning, our\nmethod enforces conditional independence between sensitive attributes and\nlearned representations, addressing bias at its source while preserving\npredictive performance. Unlike prior methods, it supports diverse sensitive\nattributes, including continuous, discrete, binary, or multi-group types.\nExperiments on various types of data structure show that our approach achieves\na superior balance between fairness and utility, significantly outperforming\nstate-of-the-art baselines.\n", "date": "2025-04-08", "categories": []}
{"id": "2504.05610", "title": "Fairness in Machine Learning-based Hand Load Estimation: A Case Study on\n  Load Carriage Tasks", "abstract": "  Predicting external hand load from sensor data is essential for ergonomic\nexposure assessments, as obtaining this information typically requires direct\nobservation or supplementary data. While machine learning methods have been\nused to estimate external hand load from worker postures or force exertion\ndata, our findings reveal systematic bias in these predictions due to\nindividual differences such as age and biological sex. To explore this issue,\nwe examined bias in hand load prediction by varying the sex ratio in the\ntraining dataset. We found substantial sex disparity in predictive performance,\nespecially when the training dataset is more sex-imbalanced. To address this\nbias, we developed and evaluated a fair predictive model for hand load\nestimation that leverages a Variational Autoencoder (VAE) with feature\ndisentanglement. This approach is designed to separate sex-agnostic and\nsex-specific latent features, minimizing feature overlap. The disentanglement\ncapability enables the model to make predictions based solely on sex-agnostic\nfeatures of motion patterns, ensuring fair prediction for both biological\nsexes. Our proposed fair algorithm outperformed conventional machine learning\nmethods (e.g., Random Forests) in both fairness and predictive accuracy,\nachieving a lower mean absolute error (MAE) difference across male and female\nsets and improved fairness metrics such as statistical parity (SP) and positive\nand negative residual differences (PRD and NRD), even when trained on\nimbalanced sex datasets. These findings emphasize the importance of\nfairness-aware machine learning algorithms to prevent potential disadvantages\nin workplace health and safety for certain worker populations.\n", "date": "2025-04-08", "categories": []}
{"id": "2504.05891", "title": "To Give or Not to Give? The Impacts of Strategically Withheld Recourse", "abstract": "  Individuals often aim to reverse undesired outcomes in interactions with\nautomated systems, like loan denials, by either implementing system-recommended\nactions (recourse), or manipulating their features. While providing recourse\nbenefits users and enhances system utility, it also provides information about\nthe decision process that can be used for more effective strategic\nmanipulation, especially when the individuals collectively share such\ninformation with each other.\n  We show that this tension leads rational utility-maximizing systems to\nfrequently withhold recourse, resulting in decreased population utility,\nparticularly impacting sensitive groups.\n  To mitigate these effects, we explore the role of recourse subsidies, finding\nthem effective in increasing the provision of recourse actions by rational\nsystems, as well as lowering the potential social cost and mitigating\nunfairness caused by recourse withholding.\n", "date": "2025-04-08", "categories": []}
{"id": "2504.05923", "title": "Uncovering Fairness through Data Complexity as an Early Indicator", "abstract": "  Fairness constitutes a concern within machine learning (ML) applications.\nCurrently, there is no study on how disparities in classification complexity\nbetween privileged and unprivileged groups could influence the fairness of\nsolutions, which serves as a preliminary indicator of potential unfairness. In\nthis work, we investigate this gap, specifically, we focus on synthetic\ndatasets designed to capture a variety of biases ranging from historical bias\nto measurement and representational bias to evaluate how various complexity\nmetrics differences correlate with group fairness metrics. We then apply\nassociation rule mining to identify patterns that link disproportionate\ncomplexity differences between groups with fairness-related outcomes, offering\ndata-centric indicators to guide bias mitigation. Our findings are also\nvalidated by their application in real-world problems, providing evidence that\nquantifying group-wise classification complexity can uncover early indicators\nof potential fairness challenges. This investigation helps practitioners to\nproactively address bias in classification tasks.\n", "date": "2025-04-08", "categories": []}
{"id": "2504.04855", "title": "BIASINSPECTOR: Detecting Bias in Structured Data through LLM Agents", "abstract": "  Detecting biases in structured data is a complex and time-consuming task.\nExisting automated techniques are limited in diversity of data types and\nheavily reliant on human case-by-case handling, resulting in a lack of\ngeneralizability. Currently, large language model (LLM)-based agents have made\nsignificant progress in data science, but their ability to detect data biases\nis still insufficiently explored. To address this gap, we introduce the first\nend-to-end, multi-agent synergy framework, BIASINSPECTOR, designed for\nautomatic bias detection in structured data based on specific user\nrequirements. It first develops a multi-stage plan to analyze user-specified\nbias detection tasks and then implements it with a diverse and well-suited set\nof tools. It delivers detailed results that include explanations and\nvisualizations. To address the lack of a standardized framework for evaluating\nthe capability of LLM agents to detect biases in data, we further propose a\ncomprehensive benchmark that includes multiple evaluation metrics and a large\nset of test cases. Extensive experiments demonstrate that our framework\nachieves exceptional overall performance in structured data bias detection,\nsetting a new milestone for fairer data applications.\n", "date": "2025-04-07", "categories": []}
{"id": "2504.03093", "title": "Post-processing for Fair Regression via Explainable SVD", "abstract": "  This paper presents a post-processing algorithm for training fair neural\nnetwork regression models that satisfy statistical parity, utilizing an\nexplainable singular value decomposition (SVD) of the weight matrix. We propose\na linear transformation of the weight matrix, whereby the singular values\nderived from the SVD of the transformed matrix directly correspond to the\ndifferences in the first and second moments of the output distributions across\ntwo groups. Consequently, we can convert the fairness constraints into\nconstraints on the singular values. We analytically solve the problem of\nfinding the optimal weights under these constraints. Experimental validation on\nvarious datasets demonstrates that our method achieves a similar or superior\nfairness-accuracy trade-off compared to the baselines without using the\nsensitive attribute at the inference time.\n", "date": "2025-04-04", "categories": []}
{"id": "2411.08297", "title": "TowerDebias: A Novel Unfairness Removal Method Based on the Tower\n  Property", "abstract": "  Decision-making processes have increasingly come to rely on sophisticated\nmachine learning tools, raising critical concerns about the fairness of their\npredictions with respect to sensitive groups. The widespread adoption of\ncommercial \"black-box\" models necessitates careful consideration of their legal\nand ethical implications for consumers. When users interact with such black-box\nmodels, a key challenge arises: how can the influence of sensitive attributes,\nsuch as race or gender, be mitigated or removed from its predictions? We\npropose towerDebias (tDB), a novel post-processing method designed to reduce\nthe influence of sensitive attributes in predictions made by black-box models.\nOur tDB approach leverages the Tower Property from probability theory to\nimprove prediction fairness without requiring retraining of the original model.\nThis method is highly versatile, as it requires no prior knowledge of the\noriginal algorithm's internal structure and is adaptable to a diverse range of\napplications. We present a formal fairness improvement theorem for tDB and\nshowcase its effectiveness in both regression and classification tasks using\nmultiple real-world datasets.\n", "date": "2024-11-13", "categories": []}
{"id": "2504.02461", "title": "Am I Being Treated Fairly? A Conceptual Framework for Individuals to\n  Ascertain Fairness", "abstract": "  Current fairness metrics and mitigation techniques provide tools for\npractitioners to asses how non-discriminatory Automatic Decision Making (ADM)\nsystems are. What if I, as an individual facing a decision taken by an ADM\nsystem, would like to know: Am I being treated fairly? We explore how to create\nthe affordance for users to be able to ask this question of ADM. In this paper,\nwe argue for the reification of fairness not only as a property of ADM, but\nalso as an epistemic right of an individual to acquire information about the\ndecisions that affect them and use that information to contest and seek\neffective redress against those decisions, in case they are proven to be\ndiscriminatory. We examine key concepts from existing research not only in\nalgorithmic fairness but also in explainable artificial intelligence,\naccountability, and contestability. Integrating notions from these domains, we\npropose a conceptual framework to ascertain fairness by combining different\ntools that empower the end-users of ADM systems. Our framework shifts the focus\nfrom technical solutions aimed at practitioners to mechanisms that enable\nindividuals to understand, challenge, and verify the fairness of decisions, and\nalso serves as a blueprint for organizations and policymakers, bridging the gap\nbetween technical requirements and practical, user-centered accountability.\n", "date": "2025-04-03", "categories": []}
{"id": "2305.06967", "title": "Data quality dimensions for fair AI", "abstract": "  Artificial Intelligence (AI) systems are not intrinsically neutral and biases\ntrickle in any type of technological tool. In particular when dealing with\npeople, the impact of AI algorithms' technical errors originating with\nmislabeled data is undeniable. As they feed wrong and discriminatory\nclassifications, these systems are not systematically guarded against bias. In\nthis article we consider the problem of bias in AI systems from the point of\nview of data quality dimensions. We highlight the limited model construction of\nbias mitigation tools based on accuracy strategy, illustrating potential\nimprovements of a specific tool in gender classification errors occurring in\ntwo typically difficult contexts: the classification of non-binary individuals,\nfor which the label set becomes incomplete with respect to the dataset; and the\nclassification of transgender individuals, for which the dataset becomes\ninconsistent with respect to the label set. Using formal methods for reasoning\nabout the behavior of the classification system in presence of a changing\nworld, we propose to reconsider the fairness of the classification task in\nterms of completeness, consistency, timeliness and reliability, and offer some\ntheoretical results.\n", "date": "2023-05-11", "categories": []}
{"id": "2410.08407", "title": "What is Left After Distillation? How Knowledge Transfer Impacts Fairness\n  and Bias", "abstract": "  Knowledge Distillation is a commonly used Deep Neural Network (DNN)\ncompression method, which often maintains overall generalization performance.\nHowever, we show that even for balanced image classification datasets, such as\nCIFAR-100, Tiny ImageNet and ImageNet, as many as 41% of the classes are\nstatistically significantly affected by distillation when comparing class-wise\naccuracy (i.e. class bias) between a teacher/distilled student or distilled\nstudent/non-distilled student model. Changes in class bias are not necessarily\nan undesirable outcome when considered outside of the context of a model's\nusage. Using two common fairness metrics, Demographic Parity Difference (DPD)\nand Equalized Odds Difference (EOD) on models trained with the CelebA,\nTrifeature, and HateXplain datasets, our results suggest that increasing the\ndistillation temperature improves the distilled student model's fairness, and\nthe distilled student fairness can even surpass the fairness of the teacher\nmodel at high temperatures. Additionally, we examine individual fairness,\nensuring similar instances receive similar predictions. Our results confirm\nthat higher temperatures also improve the distilled student model's individual\nfairness. This study highlights the uneven effects of distillation on certain\nclasses and its potentially significant role in fairness, emphasizing that\ncaution is warranted when using distilled models for sensitive application\ndomains.\n", "date": "2024-10-10", "categories": []}
{"id": "2501.05396", "title": "FairCoder: Evaluating Social Bias of LLMs in Code Generation", "abstract": "  Large language models (LLMs) have been widely deployed in coding tasks,\ndrawing increasing attention to the evaluation of the quality and safety of\nLLMs' outputs. However, research on bias in code generation remains limited.\nExisting studies typically identify bias by applying malicious prompts or\nreusing tasks and dataset originally designed for discriminative models. Given\nthat prior datasets are not fully optimized for code-related tasks, there is a\npressing need for benchmarks specifically designed for evaluating code models.\nIn this study, we introduce FairCoder, a novel benchmark for evaluating social\nbias in code generation. FairCoder explores the bias issue following the\npipeline in software development, from function implementation to unit test,\nwith diverse real-world scenarios. Additionally, three metrics are designed to\nassess fairness performance on this benchmark. We conduct experiments on widely\nused LLMs and provide a comprehensive analysis of the results. The findings\nreveal that all tested LLMs exhibit social bias.\n", "date": "2025-01-09", "categories": []}
{"id": "2504.01420", "title": "FAIRE: Assessing Racial and Gender Bias in AI-Driven Resume Evaluations", "abstract": "  In an era where AI-driven hiring is transforming recruitment practices,\nconcerns about fairness and bias have become increasingly important. To explore\nthese issues, we introduce a benchmark, FAIRE (Fairness Assessment In Resume\nEvaluation), to test for racial and gender bias in large language models (LLMs)\nused to evaluate resumes across different industries. We use two methods-direct\nscoring and ranking-to measure how model performance changes when resumes are\nslightly altered to reflect different racial or gender identities. Our findings\nreveal that while every model exhibits some degree of bias, the magnitude and\ndirection vary considerably. This benchmark provides a clear way to examine\nthese differences and offers valuable insights into the fairness of AI-based\nhiring tools. It highlights the urgent need for strategies to reduce bias in\nAI-driven recruitment. Our benchmark code and dataset are open-sourced at our\nrepository:\nhttps://github.com/athenawen/FAIRE-Fairness-Assessment-In-Resume-Evaluation.git.\n", "date": "2025-04-02", "categories": []}
{"id": "2503.22762", "title": "The Cost of Local and Global Fairness in Federated Learning", "abstract": "  With the emerging application of Federated Learning (FL) in finance, hiring\nand healthcare, FL models are regulated to be fair, preventing disparities with\nrespect to legally protected attributes such as race or gender. Two concepts of\nfairness are important in FL: global and local fairness. Global fairness\naddresses the disparity across the entire population and local fairness is\nconcerned with the disparity within each client. Prior fair FL frameworks have\nimproved either global or local fairness without considering both. Furthermore,\nwhile the majority of studies on fair FL focuses on binary settings, many\nreal-world applications are multi-class problems. This paper proposes a\nframework that investigates the minimum accuracy lost for enforcing a specified\nlevel of global and local fairness in multi-class FL settings. Our framework\nleads to a simple post-processing algorithm that derives fair outcome\npredictors from the Bayesian optimal score functions. Experimental results show\nthat our algorithm outperforms the current state of the art (SOTA) with regard\nto the accuracy-fairness tradoffs, computational and communication costs. Codes\nare available at:\nhttps://github.com/papersubmission678/The-cost-of-local-and-global-fairness-in-FL .\n", "date": "2025-03-27", "categories": []}
{"id": "2504.00310", "title": "Detecting and Mitigating Bias in LLMs through Knowledge Graph-Augmented\n  Training", "abstract": "  Large language models have revolutionized natural language processing with\ntheir surprising capability to understand and generate human-like text.\nHowever, many of these models inherit and further amplify the biases present in\ntheir training data, raising ethical and fairness concerns. The detection and\nmitigation of such biases are vital to ensuring that LLMs act responsibly and\nequitably across diverse domains. This work investigates Knowledge\nGraph-Augmented Training (KGAT) as a novel method to mitigate bias in LLM.\nUsing structured domain-specific knowledge from real-world knowledge graphs, we\nimprove the understanding of the model and reduce biased output. Public\ndatasets for bias assessment include Gender Shades, Bias in Bios, and FairFace,\nwhile metrics such as demographic parity and equal opportunity facilitate\nrigorous detection. We also performed targeted mitigation strategies to correct\nbiased associations, leading to a significant drop in biased output and\nimproved bias metrics. Equipped with real-world datasets and knowledge graphs,\nour framework is both scalable and effective, paving the way toward responsible\ndeployment in sensitive and high-stakes applications.\n", "date": "2025-04-01", "categories": []}
{"id": "2504.00388", "title": "Using complex prompts to identify fine-grained biases in image\n  generation through ChatGPT-4o", "abstract": "  There are not one but two dimensions of bias that can be revealed through the\nstudy of large AI models: not only bias in training data or the products of an\nAI, but also bias in society, such as disparity in employment or health\noutcomes between different demographic groups. Often training data and AI\noutput is biased for or against certain demographics (i.e. older white people\nare overrepresented in image datasets), but sometimes large AI models\naccurately illustrate biases in the real world (i.e. young black men being\ndisproportionately viewed as threatening). These social disparities often\nappear in image generation AI outputs in the form of 'marked' features, where\nsome feature of an individual or setting is a social marker of disparity, and\nprompts both humans and AI systems to treat subjects that are marked in this\nway as exceptional and requiring special treatment. Generative AI has proven to\nbe very sensitive to such marked features, to the extent of over-emphasising\nthem and thus often exacerbating social biases. I briefly discuss how we can\nuse complex prompts to image generation AI to investigate either dimension of\nbias, emphasising how we can probe the large language models underlying image\ngeneration AI through, for example, automated sentiment analysis of the text\nprompts used to generate images.\n", "date": "2025-04-01", "categories": []}
{"id": "2504.00860", "title": "Investigating the Capabilities and Limitations of Machine Learning for\n  Identifying Bias in English Language Data with Information and Heritage\n  Professionals", "abstract": "  Despite numerous efforts to mitigate their biases, ML systems continue to\nharm already-marginalized people. While predominant ML approaches assume bias\ncan be removed and fair models can be created, we show that these are not\nalways possible, nor desirable, goals. We reframe the problem of ML bias by\ncreating models to identify biased language, drawing attention to a dataset's\nbiases rather than trying to remove them. Then, through a workshop, we\nevaluated the models for a specific use case: workflows of information and\nheritage professionals. Our findings demonstrate the limitations of ML for\nidentifying bias due to its contextual nature, the way in which approaches to\nmitigating it can simultaneously privilege and oppress different communities,\nand its inevitability. We demonstrate the need to expand ML approaches to bias\nand fairness, providing a mixed-methods approach to investigating the\nfeasibility of removing bias or achieving fairness in a given ML use case.\n", "date": "2025-04-01", "categories": []}
{"id": "2504.00874", "title": "P2NIA: Privacy-Preserving Non-Iterative Auditing", "abstract": "  The emergence of AI legislation has increased the need to assess the ethical\ncompliance of high-risk AI systems. Traditional auditing methods rely on\nplatforms' application programming interfaces (APIs), where responses to\nqueries are examined through the lens of fairness requirements. However, such\napproaches put a significant burden on platforms, as they are forced to\nmaintain APIs while ensuring privacy, facing the possibility of data leaks.\nThis lack of proper collaboration between the two parties, in turn, causes a\nsignificant challenge to the auditor, who is subject to estimation bias as they\nare unaware of the data distribution of the platform. To address these two\nissues, we present P2NIA, a novel auditing scheme that proposes a mutually\nbeneficial collaboration for both the auditor and the platform. Extensive\nexperiments demonstrate P2NIA's effectiveness in addressing both issues. In\nsummary, our work introduces a privacy-preserving and non-iterative audit\nscheme that enhances fairness assessments using synthetic or local data,\navoiding the challenges associated with traditional API-based audits.\n", "date": "2025-04-01", "categories": []}
{"id": "2410.13146", "title": "debiaSAE: Benchmarking and Mitigating Vision-Language Model Bias", "abstract": "  As Vision Language Models (VLMs) gain widespread use, their fairness remains\nunder-explored. In this paper, we analyze demographic biases across five models\nand six datasets. We find that portrait datasets like UTKFace and CelebA are\nthe best tools for bias detection, finding gaps in performance and fairness for\nboth LLaVa and CLIP models. Scene-based datasets like PATA and VLStereoSet fail\nto be useful benchmarks for bias due to their text prompts allowing the model\nto guess the answer without a picture. As for pronoun-based datasets like\nVisoGender, we receive mixed signals as only some subsets of the data are\nuseful in providing insights. To alleviate these two problems, we introduce a\nmore rigorous evaluation dataset and a debiasing method based on Sparse\nAutoencoders to help reduce bias in models. We find that our data set generates\nmore meaningful errors than the previous data sets. Furthermore, our debiasing\nmethod improves fairness, gaining 5-15 points in performance over the baseline.\nThis study displays the problems with the current benchmarks for measuring\ndemographic bias in Vision Language Models and introduces both a more effective\ndataset for measuring bias and a novel and interpretable debiasing method based\non Sparse Autoencoders.\n", "date": "2024-10-17", "categories": []}
{"id": "2503.22934", "title": "FairSAM: Fair Classification on Corrupted Data Through Sharpness-Aware\n  Minimization", "abstract": "  Image classification models trained on clean data often suffer from\nsignificant performance degradation when exposed to testing corrupted data,\nsuch as images with impulse noise, Gaussian noise, or environmental noise. This\ndegradation not only impacts overall performance but also disproportionately\naffects various demographic subgroups, raising critical algorithmic bias\nconcerns. Although robust learning algorithms like Sharpness-Aware Minimization\n(SAM) have shown promise in improving overall model robustness and\ngeneralization, they fall short in addressing the biased performance\ndegradation across demographic subgroups. Existing fairness-aware machine\nlearning methods - such as fairness constraints and reweighing strategies - aim\nto reduce performance disparities but hardly maintain robust and equitable\naccuracy across demographic subgroups when faced with data corruption. This\nreveals an inherent tension between robustness and fairness when dealing with\ncorrupted data. To address these challenges, we introduce one novel metric\nspecifically designed to assess performance degradation across subgroups under\ndata corruption. Additionally, we propose \\textbf{FairSAM}, a new framework\nthat integrates \\underline{Fair}ness-oriented strategies into \\underline{SAM}\nto deliver equalized performance across demographic groups under corrupted\nconditions. Our experiments on multiple real-world datasets and various\npredictive tasks show that FairSAM successfully reconciles robustness and\nfairness, offering a structured solution for equitable and resilient image\nclassification in the presence of data corruption.\n", "date": "2025-03-29", "categories": []}
{"id": "2503.23979", "title": "The more the merrier: logical and multistage processors in credit\n  scoring", "abstract": "  Machine Learning algorithms are ubiquitous in key decision-making contexts\nsuch as organizational justice or healthcare, which has spawned a great demand\nfor fairness in these procedures. In this paper we focus on the application of\nfair ML in finance, more concretely on the use of fairness techniques on credit\nscoring. This paper makes two contributions. On the one hand, it addresses the\nexistent gap concerning the application of established methods in the\nliterature to the case of multiple sensitive variables through the use of a new\ntechnique called logical processors (LP). On the other hand, it also explores\nthe novel method of multistage processors (MP) to investigate whether the\ncombination of fairness methods can work synergistically to produce solutions\nwith improved fairness or accuracy. Furthermore, we examine the intersection of\nthese two lines of research by exploring the integration of fairness methods in\nthe multivariate case. The results are very promising and suggest that logical\nprocessors are an appropriate way of handling multiple sensitive variables.\nFurthermore, multistage processors are capable of improving the performance of\nexisting methods.\n", "date": "2025-03-31", "categories": []}
{"id": "2503.22569", "title": "Comparing Methods for Bias Mitigation in Graph Neural Networks", "abstract": "  This paper examines the critical role of Graph Neural Networks (GNNs) in data\npreparation for generative artificial intelligence (GenAI) systems, with a\nparticular focus on addressing and mitigating biases. We present a comparative\nanalysis of three distinct methods for bias mitigation: data sparsification,\nfeature modification, and synthetic data augmentation. Through experimental\nanalysis using the german credit dataset, we evaluate these approaches using\nmultiple fairness metrics, including statistical parity, equality of\nopportunity, and false positive rates. Our research demonstrates that while all\nmethods improve fairness metrics compared to the original dataset, stratified\nsampling and synthetic data augmentation using GraphSAGE prove particularly\neffective in balancing demographic representation while maintaining model\nperformance. The results provide practical insights for developing more\nequitable AI systems while maintaining model performance.\n", "date": "2025-03-28", "categories": []}
{"id": "2409.19804", "title": "Does RAG Introduce Unfairness in LLMs? Evaluating Fairness in\n  Retrieval-Augmented Generation Systems", "abstract": "  Retrieval-Augmented Generation (RAG) has recently gained significant\nattention for its enhanced ability to integrate external knowledge sources into\nopen-domain question answering (QA) tasks. However, it remains unclear how\nthese models address fairness concerns, particularly with respect to sensitive\nattributes such as gender, geographic location, and other demographic factors.\nFirst, as language models evolve to prioritize utility, like improving exact\nmatch accuracy, fairness considerations may have been largely overlooked.\nSecond, the complex, multi-component architecture of RAG methods poses\nchallenges in identifying and mitigating biases, as each component is optimized\nfor distinct objectives. In this paper, we aim to empirically evaluate fairness\nin several RAG methods. We propose a fairness evaluation framework tailored to\nRAG, using scenario-based questions and analyzing disparities across\ndemographic attributes. Our experimental results indicate that, despite recent\nadvances in utility-driven optimization, fairness issues persist in both the\nretrieval and generation stages. These findings underscore the need for\ntargeted interventions to address fairness concerns throughout the RAG\npipeline. The dataset and code used in this study are publicly available at\nthis GitHub Repository https://github.com/elviswxy/RAG_fairness .\n", "date": "2024-09-29", "categories": []}
{"id": "2503.20414", "title": "Active Data Sampling and Generation for Bias Remediation", "abstract": "  Adequate sampling space coverage is the keystone to effectively train\ntrustworthy Machine Learning models. Unfortunately, real data do carry several\ninherent risks due to the many potential biases they exhibit when gathered\nwithout a proper random sampling over the reference population, and most of the\ntimes this is way too expensive or time consuming to be a viable option.\nDepending on how training data have been gathered, unmitigated biases can lead\nto harmful or discriminatory consequences that ultimately hinders large scale\napplicability of pre-trained models and undermine their truthfulness or\nfairness expectations. In this paper, a mixed active sampling and data\ngeneration strategy -- called samplation -- is proposed as a mean to compensate\nduring fine-tuning of a pre-trained classifer the unfair classifications it\nproduces, assuming that the training data come from a non-probabilistic\nsampling schema. Given a pre-trained classifier, first a fairness metric is\nevaluated on a test set, then new reservoirs of labeled data are generated and\nfinally a number of reversely-biased artificial samples are generated for the\nfine-tuning of the model. Using as case study Deep Models for visual semantic\nrole labeling, the proposed method has been able to fully cure a simulated\ngender bias starting from a 90/10 imbalance, with only a small percentage of\nnew data and with a minor effect on accuracy.\n", "date": "2025-03-26", "categories": []}
{"id": "2412.08795", "title": "Coverage-based Fairness in Multi-document Summarization", "abstract": "  Fairness in multi-document summarization (MDS) measures whether a system can\ngenerate a summary fairly representing information from documents with\ndifferent social attribute values. Fairness in MDS is crucial since a fair\nsummary can offer readers a comprehensive view. Previous works focus on\nquantifying summary-level fairness using Proportional Representation, a\nfairness measure based on Statistical Parity. However, Proportional\nRepresentation does not consider redundancy in input documents and overlooks\ncorpus-level unfairness. In this work, we propose a new summary-level fairness\nmeasure, Equal Coverage, which is based on coverage of documents with different\nsocial attribute values and considers the redundancy within documents. To\ndetect the corpus-level unfairness, we propose a new corpus-level measure,\nCoverage Parity. Our human evaluations show that our measures align more with\nour definition of fairness. Using our measures, we evaluate the fairness of\nthirteen different LLMs. We find that Claude3-sonnet is the fairest among all\nevaluated LLMs. We also find that almost all LLMs overrepresent different\nsocial attribute values. The code is available at\nhttps://github.com/leehaoyuan/coverage_fairness.\n", "date": "2024-12-11", "categories": []}
{"id": "2503.19182", "title": "Evaluating Bias in LLMs for Job-Resume Matching: Gender, Race, and\n  Education", "abstract": "  Large Language Models (LLMs) offer the potential to automate hiring by\nmatching job descriptions with candidate resumes, streamlining recruitment\nprocesses, and reducing operational costs. However, biases inherent in these\nmodels may lead to unfair hiring practices, reinforcing societal prejudices and\nundermining workplace diversity. This study examines the performance and\nfairness of LLMs in job-resume matching tasks within the English language and\nU.S. context. It evaluates how factors such as gender, race, and educational\nbackground influence model decisions, providing critical insights into the\nfairness and reliability of LLMs in HR applications. Our findings indicate that\nwhile recent models have reduced biases related to explicit attributes like\ngender and race, implicit biases concerning educational background remain\nsignificant. These results highlight the need for ongoing evaluation and the\ndevelopment of advanced bias mitigation strategies to ensure equitable hiring\npractices when using LLMs in industry settings.\n", "date": "2025-03-24", "categories": []}
{"id": "2412.18138", "title": "What Constitutes a Less Discriminatory Algorithm?", "abstract": "  Disparate impact doctrine offers an important legal apparatus for targeting\ndiscriminatory data-driven algorithmic decisions. A recent body of work has\nfocused on conceptualizing one particular construct from this doctrine: the\nless discriminatory alternative, an alternative policy that reduces disparities\nwhile meeting the same business needs of a status quo or baseline policy.\nHowever, attempts to operationalize this construct in the algorithmic setting\nmust grapple with some thorny challenges and ambiguities. In this paper, we\nattempt to raise and resolve important questions about less discriminatory\nalgorithms (LDAs). How should we formally define LDAs, and how does this\ninteract with different societal goals they might serve? And how feasible is it\nfor firms or plaintiffs to computationally search for candidate LDAs? We find\nthat formal LDA definitions face fundamental challenges when they attempt to\nevaluate and compare predictive models in the absence of held-out data. As a\nresult, we argue that LDA definitions cannot be purely quantitative, and must\nrely on standards of \"reasonableness.\" We then identify both mathematical and\ncomputational constraints on firms' ability to efficiently conduct a proactive\nsearch for LDAs, but we provide evidence that these limits are \"weak\" in a\nformal sense. By defining LDAs formally, we put forward a framework in which\nboth firms and plaintiffs can search for alternative models that comport with\nsocietal goals.\n", "date": "2024-12-24", "categories": []}
{"id": "2501.09687", "title": "U-Fair: Uncertainty-based Multimodal Multitask Learning for Fairer\n  Depression Detection", "abstract": "  Machine learning bias in mental health is becoming an increasingly pertinent\nchallenge. Despite promising efforts indicating that multitask approaches often\nwork better than unitask approaches, there is minimal work investigating the\nimpact of multitask learning on performance and fairness in depression\ndetection nor leveraged it to achieve fairer prediction outcomes. In this work,\nwe undertake a systematic investigation of using a multitask approach to\nimprove performance and fairness for depression detection. We propose a novel\ngender-based task-reweighting method using uncertainty grounded in how the\nPHQ-8 questionnaire is structured. Our results indicate that, although a\nmultitask approach improves performance and fairness compared to a unitask\napproach, the results are not always consistent and we see evidence of negative\ntransfer and a reduction in the Pareto frontier, which is concerning given the\nhigh-stake healthcare setting. Our proposed approach of gender-based\nreweighting with uncertainty improves performance and fairness and alleviates\nboth challenges to a certain extent. Our findings on each PHQ-8 subitem task\ndifficulty are also in agreement with the largest study conducted on the PHQ-8\nsubitem discrimination capacity, thus providing the very first tangible\nevidence linking ML findings with large-scale empirical population studies\nconducted on the PHQ-8.\n", "date": "2025-01-16", "categories": []}
{"id": "2503.17956", "title": "On the Origins of Sampling Bias: Implications on Fairness Measurement\n  and Mitigation", "abstract": "  Accurately measuring discrimination is crucial to faithfully assessing\nfairness of trained machine learning (ML) models. Any bias in measuring\ndiscrimination leads to either amplification or underestimation of the existing\ndisparity. Several sources of bias exist and it is assumed that bias resulting\nfrom machine learning is born equally by different groups (e.g. females vs\nmales, whites vs blacks, etc.). If, however, bias is born differently by\ndifferent groups, it may exacerbate discrimination against specific\nsub-populations. Sampling bias, in particular, is inconsistently used in the\nliterature to describe bias due to the sampling procedure. In this paper, we\nattempt to disambiguate this term by introducing clearly defined variants of\nsampling bias, namely, sample size bias (SSB) and underrepresentation bias\n(URB). Through an extensive set of experiments on benchmark datasets and using\nmainstream learning algorithms, we expose relevant observations in several\nmodel training scenarios. The observations are finally framed as actionable\nrecommendations for practitioners.\n", "date": "2025-03-23", "categories": []}
{"id": "2309.14345", "title": "Bias Testing and Mitigation in LLM-based Code Generation", "abstract": "  As the adoption of LLMs becomes more widespread in software coding\necosystems, a pressing issue has emerged: does the generated code contain\nsocial bias and unfairness, such as those related to age, gender, and race?\nThis issue concerns the integrity, fairness, and ethical foundation of software\napplications that depend on the code generated by these models but are\nunderexplored in the literature. This paper presents a novel bias testing\nframework that is specifically designed for code generation tasks. Based on\nthis framework, we conduct an extensive empirical study on the biases in code\ngenerated by five widely studied LLMs (i.e., PALM-2-CodeChat-bison,\nClaude-instant-1, GPT-3.5-turbo, GPT-4-turbo, and GPT-4). Our findings reveal\nthat biases are prevalent. For example, 13.47% to 49.10% of the codes generated\nby these LLMs have biased behaviors towards gender. Moreover, we study five\nbias mitigation prompt strategies that are commonly used in current code\ngeneration scenarios, i.e., zero-shot, one-shot, few-shot, and two\nChain-of-Thought (CoT) prompts, with and without provided feedback-driven\nrefinement. Our evaluation results illustrate that using direct prompt\nengineering strategies has limited effectiveness in mitigating bias, but our\ntest execution feedback can help to reduce the ratio of code biases to a large\nextent (e.g., from 59.88% to 4.79% for GPT-4).\n", "date": "2023-09-03", "categories": []}
{"id": "2503.16836", "title": "A Flexible Fairness Framework with Surrogate Loss Reweighting for\n  Addressing Sociodemographic Disparities", "abstract": "  This paper presents a new algorithmic fairness framework called\n$\\boldsymbol{\\alpha}$-$\\boldsymbol{\\beta}$ Fair Machine Learning\n($\\boldsymbol{\\alpha}$-$\\boldsymbol{\\beta}$ FML), designed to optimize fairness\nlevels across sociodemographic attributes. Our framework employs a new family\nof surrogate loss functions, paired with loss reweighting techniques, allowing\nprecise control over fairness-accuracy trade-offs through tunable\nhyperparameters $\\boldsymbol{\\alpha}$ and $\\boldsymbol{\\beta}$. To efficiently\nsolve the learning objective, we propose Parallel Stochastic Gradient Descent\nwith Surrogate Loss (P-SGD-S) and establish convergence guarantees for both\nconvex and nonconvex loss functions. Experimental results demonstrate that our\nframework improves overall accuracy while reducing fairness violations,\noffering a smooth trade-off between standard empirical risk minimization and\nstrict minimax fairness. Results across multiple datasets confirm its\nadaptability, ensuring fairness improvements without excessive performance\ndegradation.\n", "date": "2025-03-21", "categories": []}
{"id": "2503.17231", "title": "LoGoFair: Post-Processing for Local and Global Fairness in Federated\n  Learning", "abstract": "  Federated learning (FL) has garnered considerable interest for its capability\nto learn from decentralized data sources. Given the increasing application of\nFL in decision-making scenarios, addressing fairness issues across different\nsensitive groups (e.g., female, male) in FL is crucial. Current research often\nfocuses on facilitating fairness at each client's data (local fairness) or\nwithin the entire dataset across all clients (global fairness). However,\nexisting approaches that focus exclusively on either local or global fairness\nfail to address two key challenges: (\\textbf{CH1}) Under statistical\nheterogeneity, global fairness does not imply local fairness, and vice versa.\n(\\textbf{CH2}) Achieving fairness under model-agnostic setting. To tackle the\naforementioned challenges, this paper proposes a novel post-processing\nframework for achieving both Local and Global Fairness in the FL context,\nnamely LoGoFair. To address CH1, LoGoFair endeavors to seek the Bayes optimal\nclassifier under local and global fairness constraints, which strikes the\noptimal accuracy-fairness balance in the probabilistic sense. To address CH2,\nLoGoFair employs a model-agnostic federated post-processing procedure that\nenables clients to collaboratively optimize global fairness while ensuring\nlocal fairness, thereby achieving the optimal fair classifier within FL.\nExperimental results on three real-world datasets further illustrate the\neffectiveness of the proposed LoGoFair framework.\n", "date": "2025-03-21", "categories": []}
{"id": "2409.00553", "title": "Multi-Output Distributional Fairness via Post-Processing", "abstract": "  The post-processing approaches are becoming prominent techniques to enhance\nmachine learning models' fairness because of their intuitiveness, low\ncomputational cost, and excellent scalability. However, most existing\npost-processing methods are designed for task-specific fairness measures and\nare limited to single-output models. In this paper, we introduce a\npost-processing method for multi-output models, such as the ones used for\nmulti-task/multi-class classification and representation learning, to enhance a\nmodel's distributional parity, a task-agnostic fairness measure. Existing\nmethods for achieving distributional parity rely on the (inverse) cumulative\ndensity function of a model's output, restricting their applicability to\nsingle-output models. Extending previous works, we propose to employ optimal\ntransport mappings to move a model's outputs across different groups towards\ntheir empirical Wasserstein barycenter. An approximation technique is applied\nto reduce the complexity of computing the exact barycenter and a kernel\nregression method is proposed to extend this process to out-of-sample data. Our\nempirical studies evaluate the proposed approach against various baselines on\nmulti-task/multi-class classification and representation learning tasks,\ndemonstrating the effectiveness of the proposed approach.\n", "date": "2024-08-31", "categories": []}
{"id": "2207.04053", "title": "On the Need and Applicability of Causality for Fairness: A Unified\n  Framework for AI Auditing and Legal Analysis", "abstract": "  As Artificial Intelligence (AI) increasingly influences decisions in critical\nsocietal sectors, understanding and establishing causality becomes essential\nfor evaluating the fairness of automated systems. This article explores the\nsignificance of causal reasoning in addressing algorithmic discrimination,\nemphasizing both legal and societal perspectives. By reviewing landmark cases\nand regulatory frameworks, particularly within the European Union, we\nillustrate the challenges inherent in proving causal claims when confronted\nwith opaque AI decision-making processes. The discussion outlines practical\nobstacles and methodological limitations in applying causal inference to\nreal-world fairness scenarios, proposing actionable solutions to enhance\ntransparency, accountability, and fairness in algorithm-driven decisions.\n", "date": "2022-07-08", "categories": []}
{"id": "2503.15163", "title": "Global Group Fairness in Federated Learning via Function Tracking", "abstract": "  We investigate group fairness regularizers in federated learning, aiming to\ntrain a globally fair model in a distributed setting. Ensuring global fairness\nin distributed training presents unique challenges, as fairness regularizers\ntypically involve probability metrics between distributions across all clients\nand are not naturally separable by client. To address this, we introduce a\nfunction-tracking scheme for the global fairness regularizer based on a Maximum\nMean Discrepancy (MMD), which incurs a small communication overhead. This\nscheme seamlessly integrates into most federated learning algorithms while\npreserving rigorous convergence guarantees, as demonstrated in the context of\nFedAvg. Additionally, when enforcing differential privacy, the kernel-based MMD\nregularization enables straightforward analysis through a change of kernel,\nleveraging an intuitive interpretation of kernel convolution. Numerical\nexperiments confirm our theoretical insights.\n", "date": "2025-03-19", "categories": []}
{"id": "2208.06648", "title": "Imputation Strategies Under Clinical Presence: Impact on Algorithmic\n  Fairness", "abstract": "  Machine learning risks reinforcing biases present in data and, as we argue in\nthis work, in what is absent from data. In healthcare, societal and decision\nbiases shape patterns in missing data, yet the algorithmic fairness\nimplications of group-specific missingness are poorly understood. The way we\naddress missingness in healthcare can have detrimental impacts on downstream\nalgorithmic fairness. Our work questions current recommendations and practices\naimed at handling missing data with a focus on their effect on algorithmic\nfairness, and offers a path forward. Specifically, we consider the theoretical\nunderpinnings of existing recommendations as well as their empirical predictive\nperformance and corresponding algorithmic fairness measured through subgroup\nperformances. Our results show that current practices for handling missingness\nlack principled foundations, are disconnected from the realities of missingness\nmechanisms in healthcare, and can be counterproductive. For example, we show\nthat favouring group-specific imputation strategy can be misguided and\nexacerbate prediction disparities. We then build on our findings to propose a\nframework for empirically guiding imputation choices, and an accompanying\nreporting framework. Our work constitutes an important contribution to recent\nefforts by regulators and practitioners to grapple with the realities of\nreal-world data, and to foster the responsible and transparent deployment of\nmachine learning systems. We demonstrate the practical utility of the proposed\nframework through experimentation on widely used datasets, where we show how\nthe proposed framework can guide the selection of imputation strategies,\nallowing us to choose among strategies that yield equal overall predictive\nperformance but present different algorithmic fairness properties.\n", "date": "2022-08-13", "categories": []}
{"id": "2306.00636", "title": "Unfair Utilities and First Steps Towards Improving Them", "abstract": "  Many fairness criteria constrain the policy or choice of predictors, which\ncan have unwanted consequences, in particular, when optimizing the policy under\nsuch constraints. Here, we advocate to instead focus on the utility function\nthe policy is optimizing for. We define value of information fairness and\npropose to not use utility functions that violate this criterion. This\nprinciple suggests to modify these utility functions such that they satisfy\nvalue of information fairness. We describe how this can be done and discuss\nconsequences for the corresponding optimal policies. We apply our framework to\nthought experiments and the COMPAS data. Focussing on the utility function\nprovides better answers than existing fairness notions: We are not aware of any\nintuitively fair policy that is disallowed by value of information fairness,\nand when we find that value of information fairness recommends an intuitively\nunfair policy, no existing fairness notion finds an intuitively fair policy.\n", "date": "2023-06-01", "categories": []}
{"id": "2410.17263", "title": "An Effective Theory of Bias Amplification", "abstract": "  Machine learning models can capture and amplify biases present in data,\nleading to disparate test performance across social groups. To better\nunderstand, evaluate, and mitigate these biases, a deeper theoretical\nunderstanding of how model design choices and data distribution properties\ncontribute to bias is needed. In this work, we contribute a precise analytical\ntheory in the context of ridge regression, both with and without random\nprojections, where the former models feedforward neural networks in a\nsimplified regime. Our theory offers a unified and rigorous explanation of\nmachine learning bias, providing insights into phenomena such as bias\namplification and minority-group bias in various feature and parameter regimes.\nFor example, we observe that there may be an optimal regularization penalty or\ntraining time to avoid bias amplification, and there can be differences in test\nerror between groups that are not alleviated with increased parameterization.\nImportantly, our theoretical predictions align with empirical observations\nreported in the literature on machine learning bias. We extensively empirically\nvalidate our theory on synthetic and semi-synthetic datasets.\n", "date": "2024-10-07", "categories": []}
{"id": "2411.07521", "title": "Fair Summarization: Bridging Quality and Diversity in Extractive\n  Summaries", "abstract": "  Fairness in multi-document summarization of user-generated content remains a\ncritical challenge in natural language processing (NLP). Existing summarization\nmethods often fail to ensure equitable representation across different social\ngroups, leading to biased outputs. In this paper, we introduce two novel\nmethods for fair extractive summarization: FairExtract, a clustering-based\napproach, and FairGPT, which leverages GPT-3.5-turbo with fairness constraints.\nWe evaluate these methods using Divsumm summarization dataset of White-aligned,\nHispanic, and African-American dialect tweets and compare them against relevant\nbaselines. The results obtained using a comprehensive set of summarization\nquality metrics such as SUPERT, BLANC, SummaQA, BARTScore, and UniEval, as well\nas a fairness metric F, demonstrate that FairExtract and FairGPT achieve\nsuperior fairness while maintaining competitive summarization quality.\nAdditionally, we introduce composite metrics (e.g., SUPERT+F, BLANC+F) that\nintegrate quality and fairness into a single evaluation framework, offering a\nmore nuanced understanding of the trade-offs between these objectives. Our code\nis available online.\n", "date": "2024-11-12", "categories": []}
{"id": "2412.16769", "title": "Does calibration mean what they say it means; or, the reference class\n  problem rises again", "abstract": "  Discussions of statistical criteria for fairness commonly convey the\nnormative significance of calibration within groups by invoking what risk\nscores \"mean.\" On the Same Meaning picture, group-calibrated scores \"mean the\nsame thing\" (on average) across individuals from different groups and\naccordingly, guard against disparate treatment of individuals based on group\nmembership. My contention is that calibration guarantees no such thing. Since\nconcrete actual people belong to many groups, calibration cannot ensure the\nkind of consistent score interpretation that the Same Meaning picture implies\nmatters for fairness, unless calibration is met within every group to which an\nindividual belongs. Alas only perfect predictors may meet this bar. The Same\nMeaning picture thus commits a reference class fallacy by inferring from\ncalibration within some group to the \"meaning\" or evidential value of an\nindividual's score, because they are a member of that group. The reference\nclass answer it presumes does not only lack justification; it is very likely\nwrong. I then show that the reference class problem besets not just calibration\nbut other group statistical criteria that claim a close connection to fairness.\nReflecting on the origins of this oversight opens a wider lens onto the\npredominant methodology in algorithmic fairness based on stylized cases.\n", "date": "2024-12-21", "categories": []}
{"id": "2503.12536", "title": "Debiasing Diffusion Model: Enhancing Fairness through Latent\n  Representation Learning in Stable Diffusion Model", "abstract": "  Image generative models, particularly diffusion-based models, have surged in\npopularity due to their remarkable ability to synthesize highly realistic\nimages. However, since these models are data-driven, they inherit biases from\nthe training datasets, frequently leading to disproportionate group\nrepresentations that exacerbate societal inequities. Traditionally, efforts to\ndebiase these models have relied on predefined sensitive attributes,\nclassifiers trained on such attributes, or large language models to steer\noutputs toward fairness. However, these approaches face notable drawbacks:\npredefined attributes do not adequately capture complex and continuous\nvariations among groups. To address these issues, we introduce the Debiasing\nDiffusion Model (DDM), which leverages an indicator to learn latent\nrepresentations during training, promoting fairness through balanced\nrepresentations without requiring predefined sensitive attributes. This\napproach not only demonstrates its effectiveness in scenarios previously\naddressed by conventional techniques but also enhances fairness without relying\non predefined sensitive attributes as conditions. In this paper, we discuss the\nlimitations of prior bias mitigation techniques in diffusion-based models,\nelaborate on the architecture of the DDM, and validate the effectiveness of our\napproach through experiments.\n", "date": "2025-03-16", "categories": []}
{"id": "2410.19314", "title": "Revealing and Reducing Gender Biases in Vision and Language Assistants\n  (VLAs)", "abstract": "  Pre-trained large language models (LLMs) have been reliably integrated with\nvisual input for multimodal tasks. The widespread adoption of instruction-tuned\nimage-to-text vision-language assistants (VLAs) like LLaVA and InternVL\nnecessitates evaluating gender biases. We study gender bias in 22 popular\nopen-source VLAs with respect to personality traits, skills, and occupations.\nOur results show that VLAs replicate human biases likely present in the data,\nsuch as real-world occupational imbalances. Similarly, they tend to attribute\nmore skills and positive personality traits to women than to men, and we see a\nconsistent tendency to associate negative personality traits with men. To\neliminate the gender bias in these models, we find that fine-tuning-based\ndebiasing methods achieve the best trade-off between debiasing and retaining\nperformance on downstream tasks. We argue for pre-deploying gender bias\nassessment in VLAs and motivate further development of debiasing strategies to\nensure equitable societal outcomes. Code is available at\nhttps://github.com/ExplainableML/vla-gender-bias.\n", "date": "2024-10-25", "categories": []}
{"id": "2503.09763", "title": "BiasConnect: Investigating Bias Interactions in Text-to-Image Models", "abstract": "  The biases exhibited by Text-to-Image (TTI) models are often treated as if\nthey are independent, but in reality, they may be deeply interrelated.\nAddressing bias along one dimension, such as ethnicity or age, can\ninadvertently influence another dimension, like gender, either mitigating or\nexacerbating existing disparities. Understanding these interdependencies is\ncrucial for designing fairer generative models, yet measuring such effects\nquantitatively remains a challenge. In this paper, we aim to address these\nquestions by introducing BiasConnect, a novel tool designed to analyze and\nquantify bias interactions in TTI models. Our approach leverages a\ncounterfactual-based framework to generate pairwise causal graphs that reveals\nthe underlying structure of bias interactions for the given text prompt.\nAdditionally, our method provides empirical estimates that indicate how other\nbias dimensions shift toward or away from an ideal distribution when a given\nbias is modified. Our estimates have a strong correlation (+0.69) with the\ninterdependency observations post bias mitigation. We demonstrate the utility\nof BiasConnect for selecting optimal bias mitigation axes, comparing different\nTTI models on the dependencies they learn, and understanding the amplification\nof intersectional societal biases in TTI models.\n", "date": "2025-03-12", "categories": []}
{"id": "2503.09866", "title": "EquiPy: Sequential Fairness using Optimal Transport in Python", "abstract": "  Algorithmic fairness has received considerable attention due to the failures\nof various predictive AI systems that have been found to be unfairly biased\nagainst subgroups of the population. Many approaches have been proposed to\nmitigate such biases in predictive systems, however, they often struggle to\nprovide accurate estimates and transparent correction mechanisms in the case\nwhere multiple sensitive variables, such as a combination of gender and race,\nare involved. This paper introduces a new open source Python package, EquiPy,\nwhich provides a easy-to-use and model agnostic toolbox for efficiently\nachieving fairness across multiple sensitive variables. It also offers\ncomprehensive graphic utilities to enable the user to interpret the influence\nof each sensitive variable within a global context. EquiPy makes use of\ntheoretical results that allow the complexity arising from the use of multiple\nvariables to be broken down into easier-to-solve sub-problems. We demonstrate\nthe ease of use for both mitigation and interpretation on publicly available\ndata derived from the US Census and provide sample code for its use.\n", "date": "2025-03-12", "categories": []}
{"id": "2211.15072", "title": "FaiREE: Fair Classification with Finite-Sample and Distribution-Free\n  Guarantee", "abstract": "  Algorithmic fairness plays an increasingly critical role in machine learning\nresearch. Several group fairness notions and algorithms have been proposed.\nHowever, the fairness guarantee of existing fair classification methods mainly\ndepends on specific data distributional assumptions, often requiring large\nsample sizes, and fairness could be violated when there is a modest number of\nsamples, which is often the case in practice. In this paper, we propose FaiREE,\na fair classification algorithm that can satisfy group fairness constraints\nwith finite-sample and distribution-free theoretical guarantees. FaiREE can be\nadapted to satisfy various group fairness notions (e.g., Equality of\nOpportunity, Equalized Odds, Demographic Parity, etc.) and achieve the optimal\naccuracy. These theoretical guarantees are further supported by experiments on\nboth synthetic and real data. FaiREE is shown to have favorable performance\nover state-of-the-art algorithms.\n", "date": "2022-11-28", "categories": []}
{"id": "2407.14210", "title": "Fair Overlap Number of Balls (Fair-ONB): A Data-Morphology-based\n  Undersampling Method for Bias Reduction", "abstract": "  One of the key issues regarding classification problems in Trustworthy\nArtificial Intelligence is ensuring Fairness in the prediction of different\nclasses when protected (sensitive) features are present. Data quality is\ncritical in these cases, as biases in training data can be reflected in machine\nlearning, impacting human lives and failing to comply with current regulations.\nOne strategy to improve data quality and avoid these problems is preprocessing\nthe dataset. Instance selection via undersampling can foster balanced learning\nof classes and protected feature values. Performing undersampling in class\noverlap areas close to the decision boundary should bolster the impact on the\nclassifier. This work proposes Fair Overlap Number of Balls (Fair-ONB), an\nundersampling method that harnesses the data morphology of the different data\ngroups (obtained from the combination of classes and protected feature values)\nto perform guided undersampling in overlap areas. It employs attributes of the\nball coverage of the groups, such as the radius, number of covered instances\nand density, to select the most suitable areas for undersampling and reduce\nbias. Results show that the Fair-ONB method improves model Fairness with low\nimpact on the classifier's predictive performance.\n", "date": "2024-07-19", "categories": []}
{"id": "2501.06795", "title": "Bridging the Fairness Gap: Enhancing Pre-trained Models with\n  LLM-Generated Sentences", "abstract": "  Pre-trained language models (PLMs) are trained on data that inherently\ncontains gender biases, leading to undesirable impacts. Traditional debiasing\nmethods often rely on external corpora, which may lack quality, diversity, or\ndemographic balance, affecting the effectiveness of debiasing. With the rise of\nlarge language models and their extensive knowledge, we propose enhancing\nfairness (Fair-Gender) in PLMs by absorbing coherent, attribute-balanced, and\nsemantically rich sentences. However, these sentences cannot be directly used\nfor debiasing due to alignment issues and the risk of negative transfer. We\naddress this by applying causal analysis to estimate causal effects, filtering\nout unaligned sentences, and identifying aligned ones for incorporation into\nPLMs, thereby ensuring positive transfer. Experiments show that our approach\nsignificantly reduces gender biases in PLMs while preserving their language\nexpressiveness.\n", "date": "2025-01-12", "categories": []}
{"id": "2503.09587", "title": "Fair Federated Medical Image Classification Against Quality Shift via\n  Inter-Client Progressive State Matching", "abstract": "  Despite the potential of federated learning in medical applications,\ninconsistent imaging quality across institutions-stemming from lower-quality\ndata from a minority of clients-biases federated models toward more common\nhigh-quality images. This raises significant fairness concerns. Existing fair\nfederated learning methods have demonstrated some effectiveness in solving this\nproblem by aligning a single 0th- or 1st-order state of convergence (e.g.,\ntraining loss or sharpness). However, we argue in this work that fairness based\non such a single state is still not an adequate surrogate for fairness during\ntesting, as these single metrics fail to fully capture the convergence\ncharacteristics, making them suboptimal for guiding fair learning. To address\nthis limitation, we develop a generalized framework. Specifically, we propose\nassessing convergence using multiple states, defined as sharpness or perturbed\nloss computed at varying search distances. Building on this comprehensive\nassessment, we propose promoting fairness for these states across clients to\nachieve our ultimate fairness objective. This is accomplished through the\nproposed method, FedISM+. In FedISM+, the search distance evolves over time,\nprogressively focusing on different states. We then incorporate two components\nin local training and global aggregation to ensure cross-client fairness for\neach state. This gradually makes convergence equitable for all states, thereby\nimproving fairness during testing. Our empirical evaluations, performed on the\nwell-known RSNA ICH and ISIC 2019 datasets, demonstrate the superiority of\nFedISM+ over existing state-of-the-art methods for fair federated learning. The\ncode is available at https://github.com/wnn2000/FFL4MIA.\n", "date": "2025-03-12", "categories": []}
{"id": "2407.19114", "title": "To which reference class do you belong? Measuring racial fairness of\n  reference classes with normative modeling", "abstract": "  Reference classes in healthcare establish healthy norms, such as pediatric\ngrowth charts of height and weight, and are used to chart deviations from these\nnorms which represent potential clinical risk. How the demographics of the\nreference class influence clinical interpretation of deviations is unknown.\nUsing normative modeling, a method for building reference classes, we evaluate\nthe fairness (racial bias) in reference models of structural brain images that\nare widely used in psychiatry and neurology. We test whether including race in\nthe model creates fairer models. We predict self-reported race using the\ndeviation scores from three different reference class normative models, to\nbetter understand bias in an integrated, multivariate sense. Across all of\nthese tasks, we uncover racial disparities that are not easily addressed with\nexisting data or commonly used modeling techniques. Our work suggests that\ndeviations from the norm could be due to demographic mismatch with the\nreference class, and assigning clinical meaning to these deviations should be\ndone with caution. Our approach also suggests that acquiring more\nrepresentative samples is an urgent research priority.\n", "date": "2024-07-26", "categories": []}
{"id": "2407.19345", "title": "Inference-Time Selective Debiasing to Enhance Fairness in Text\n  Classification Models", "abstract": "  We propose selective debiasing -- an inference-time safety mechanism designed\nto enhance the overall model quality in terms of prediction performance and\nfairness, especially in scenarios where retraining the model is impractical.\nThe method draws inspiration from selective classification, where at inference\ntime, predictions with low quality, as indicated by their uncertainty scores,\nare discarded. In our approach, we identify the potentially biased model\npredictions and, instead of discarding them, we remove bias from these\npredictions using LEACE -- a post-processing debiasing method. To select\nproblematic predictions, we propose a bias quantification approach based on KL\ndivergence, which achieves better results than standard uncertainty\nquantification methods. Experiments on text classification datasets with\nencoder-based classification models demonstrate that selective debiasing helps\nto reduce the performance gap between post-processing methods and debiasing\ntechniques from the at-training and pre-processing categories.\n", "date": "2024-07-27", "categories": []}
{"id": "2503.07806", "title": "Towards Large Language Models that Benefit for All: Benchmarking Group\n  Fairness in Reward Models", "abstract": "  As Large Language Models (LLMs) become increasingly powerful and accessible\nto human users, ensuring fairness across diverse demographic groups, i.e.,\ngroup fairness, is a critical ethical concern. However, current fairness and\nbias research in LLMs is limited in two aspects. First, compared to traditional\ngroup fairness in machine learning classification, it requires that the\nnon-sensitive attributes, in this case, the prompt questions, be the same\nacross different groups. In many practical scenarios, different groups,\nhowever, may prefer different prompt questions and this requirement becomes\nimpractical. Second, it evaluates group fairness only for the LLM's final\noutput without identifying the source of possible bias. Namely, the bias in\nLLM's output can result from both the pretraining and the finetuning. For\nfinetuning, the bias can result from both the RLHF procedure and the learned\nreward model. Arguably, evaluating the group fairness of each component in the\nLLM pipeline could help develop better methods to mitigate the possible bias.\nRecognizing those two limitations, this work benchmarks the group fairness of\nlearned reward models. By using expert-written text from arXiv, we are able to\nbenchmark the group fairness of reward models without requiring the same prompt\nquestions across different demographic groups. Surprisingly, our results\ndemonstrate that all the evaluated reward models (e.g., Nemotron-4-340B-Reward,\nArmoRM-Llama3-8B-v0.1, and GRM-llama3-8B-sftreg) exhibit statistically\nsignificant group unfairness. We also observed that top-performing reward\nmodels (w.r.t. canonical performance metrics) tend to demonstrate better group\nfairness.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.07817", "title": "Group Fairness in Multi-Task Reinforcement Learning", "abstract": "  This paper addresses a critical societal consideration in the application of\nReinforcement Learning (RL): ensuring equitable outcomes across different\ndemographic groups in multi-task settings. While previous work has explored\nfairness in single-task RL, many real-world applications are multi-task in\nnature and require policies to maintain fairness across all tasks. We introduce\na novel formulation of multi-task group fairness in RL and propose a\nconstrained optimization algorithm that explicitly enforces fairness\nconstraints across multiple tasks simultaneously. We have shown that our\nproposed algorithm does not violate fairness constraints with high probability\nand with sublinear regret in the finite-horizon episodic setting. Through\nexperiments in RiverSwim and MuJoCo environments, we demonstrate that our\napproach better ensures group fairness across multiple tasks compared to\nprevious methods that lack explicit multi-task fairness constraints in both the\nfinite-horizon setting and the infinite-horizon setting. Our results show that\nthe proposed algorithm achieves smaller fairness gaps while maintaining\ncomparable returns across different demographic groups and tasks, suggesting\nits potential for addressing fairness concerns in real-world multi-task RL\napplications.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.08588", "title": "BiasEdit: Debiasing Stereotyped Language Models via Model Editing", "abstract": "  Previous studies have established that language models manifest stereotyped\nbiases. Existing debiasing strategies, such as retraining a model with\ncounterfactual data, representation projection, and prompting often fail to\nefficiently eliminate bias or directly alter the models' biased internal\nrepresentations. To address these issues, we propose BiasEdit, an efficient\nmodel editing method to remove stereotypical bias from language models through\nlightweight networks that act as editors to generate parameter updates.\nBiasEdit employs a debiasing loss guiding editor networks to conduct local\nedits on partial parameters of a language model for debiasing while preserving\nthe language modeling abilities during editing through a retention loss.\nExperiments on StereoSet and Crows-Pairs demonstrate the effectiveness,\nefficiency, and robustness of BiasEdit in eliminating bias compared to\ntangental debiasing baselines and little to no impact on the language models'\ngeneral capabilities. In addition, we conduct bias tracing to probe bias in\nvarious modules and explore bias editing impacts on different components of\nlanguage models.\n", "date": "2025-03-11", "categories": []}
{"id": "2502.11995", "title": "Presumed Cultural Identity: How Names Shape LLM Responses", "abstract": "  Names are deeply tied to human identity. They can serve as markers of\nindividuality, cultural heritage, and personal history. However, using names as\na core indicator of identity can lead to over-simplification of complex\nidentities. When interacting with LLMs, user names are an important point of\ninformation for personalisation. Names can enter chatbot conversations through\ndirect user input (requested by chatbots), as part of task contexts such as CV\nreviews, or as built-in memory features that store user information for\npersonalisation. We study biases associated with names by measuring cultural\npresumptions in the responses generated by LLMs when presented with common\nsuggestion-seeking queries, which might involve making assumptions about the\nuser. Our analyses demonstrate strong assumptions about cultural identity\nassociated with names present in LLM generations across multiple cultures. Our\nwork has implications for designing more nuanced personalisation systems that\navoid reinforcing stereotypes while maintaining meaningful customisation.\n", "date": "2025-02-17", "categories": []}
{"id": "2503.05769", "title": "Effect of Gender Fair Job Description on Generative AI Images", "abstract": "  STEM fields are traditionally male-dominated, with gender biases shaping\nperceptions of job accessibility. This study analyzed gender representation in\nSTEM occupation images generated by OpenAI DALL-E 3 \\& Black Forest FLUX.1\nusing 150 prompts in three linguistic forms: German generic masculine, German\npair form, and English. As control, 20 pictures of social occupations were\ngenerated as well. Results revealed significant male bias across all forms,\nwith the German pair form showing reduced bias but still overrepresenting men\nfor the STEM-Group and mixed results for the Group of Social Occupations. These\nfindings highlight generative AI's role in reinforcing societal biases,\nemphasizing the need for further discussion on diversity (in AI). Further\naspects analyzed are age-distribution and ethnic diversity.\n", "date": "2025-02-25", "categories": []}
{"id": "2503.07066", "title": "You Only Debias Once: Towards Flexible Accuracy-Fairness Trade-offs at\n  Inference Time", "abstract": "  Deep neural networks are prone to various bias issues, jeopardizing their\napplications for high-stake decision-making. Existing fairness methods\ntypically offer a fixed accuracy-fairness trade-off, since the weight of the\nwell-trained model is a fixed point (fairness-optimum) in the weight space.\nNevertheless, more flexible accuracy-fairness trade-offs at inference time are\npractically desired since: 1) stakes of the same downstream task can vary for\ndifferent individuals, and 2) different regions have diverse laws or\nregularization for fairness. If using the previous fairness methods, we have to\ntrain multiple models, each offering a specific level of accuracy-fairness\ntrade-off. This is often computationally expensive, time-consuming, and\ndifficult to deploy, making it less practical for real-world applications. To\naddress this problem, we propose You Only Debias Once (YODO) to achieve in-situ\nflexible accuracy-fairness trade-offs at inference time, using a single model\nthat trained only once. Instead of pursuing one individual fixed point\n(fairness-optimum) in the weight space, we aim to find a \"line\" in the weight\nspace that connects the accuracy-optimum and fairness-optimum points using a\nsingle model. Points (models) on this line implement varying levels of\naccuracy-fairness trade-offs. At inference time, by manually selecting the\nspecific position of the learned \"line\", our proposed method can achieve\narbitrary accuracy-fairness trade-offs for different end-users and scenarios.\nExperimental results on tabular and image datasets show that YODO achieves\nflexible trade-offs between model accuracy and fairness, at ultra-low\noverheads. For example, if we need $100$ levels of trade-off on the \\acse\ndataset, YODO takes $3.53$ seconds while training $100$ fixed models consumes\n$425$ seconds. The code is available at https://github.com/ahxt/yodo.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.07326", "title": "AI Biases as Asymmetries: A Review to Guide Practice", "abstract": "  The understanding of bias in AI is currently undergoing a revolution.\nInitially understood as errors or flaws, biases are increasingly recognized as\nintegral to AI systems and sometimes preferable to less biased alternatives. In\nthis paper, we review the reasons for this changed understanding and provide\nnew guidance on two questions: First, how should we think about and measure\nbiases in AI systems, consistent with the new understanding? Second, what kinds\nof bias in an AI system should we accept or even amplify, and what kinds should\nwe minimize or eliminate, and why? The key to answering both questions, we\nargue, is to understand biases as \"violations of a symmetry standard\"\n(following Kelly). We distinguish three main types of asymmetry in AI\nsystems-error biases, inequality biases, and process biases-and highlight\nplaces in the pipeline of AI development and application where bias of each\ntype is likely to be good, bad, or inevitable.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.05093", "title": "Visual Cues of Gender and Race are Associated with Stereotyping in\n  Vision-Language Models", "abstract": "  Current research on bias in Vision Language Models (VLMs) has important\nlimitations: it is focused exclusively on trait associations while ignoring\nother forms of stereotyping, it examines specific contexts where biases are\nexpected to appear, and it conceptualizes social categories like race and\ngender as binary, ignoring the multifaceted nature of these identities. Using\nstandardized facial images that vary in prototypicality, we test four VLMs for\nboth trait associations and homogeneity bias in open-ended contexts. We find\nthat VLMs consistently generate more uniform stories for women compared to men,\nwith people who are more gender prototypical in appearance being represented\nmore uniformly. By contrast, VLMs represent White Americans more uniformly than\nBlack Americans. Unlike with gender prototypicality, race prototypicality was\nnot related to stronger uniformity. In terms of trait associations, we find\nlimited evidence of stereotyping-Black Americans were consistently linked with\nbasketball across all models, while other racial associations (i.e., art,\nhealthcare, appearance) varied by specific VLM. These findings demonstrate that\nVLM stereotyping manifests in ways that go beyond simple group membership,\nsuggesting that conventional bias mitigation strategies may be insufficient to\naddress VLM stereotyping and that homogeneity bias persists even when trait\nassociations are less apparent in model outputs.\n", "date": "2025-03-07", "categories": []}
{"id": "2503.05665", "title": "AIM-Fair: Advancing Algorithmic Fairness via Selectively Fine-Tuning\n  Biased Models with Contextual Synthetic Data", "abstract": "  Recent advances in generative models have sparked research on improving model\nfairness with AI-generated data. However, existing methods often face\nlimitations in the diversity and quality of synthetic data, leading to\ncompromised fairness and overall model accuracy. Moreover, many approaches rely\non the availability of demographic group labels, which are often costly to\nannotate. This paper proposes AIM-Fair, aiming to overcome these limitations\nand harness the potential of cutting-edge generative models in promoting\nalgorithmic fairness. We investigate a fine-tuning paradigm starting from a\nbiased model initially trained on real-world data without demographic\nannotations. This model is then fine-tuned using unbiased synthetic data\ngenerated by a state-of-the-art diffusion model to improve its fairness. Two\nkey challenges are identified in this fine-tuning paradigm, 1) the low quality\nof synthetic data, which can still happen even with advanced generative models,\nand 2) the domain and bias gap between real and synthetic data. To address the\nlimitation of synthetic data quality, we propose Contextual Synthetic Data\nGeneration (CSDG) to generate data using a text-to-image diffusion model (T2I)\nwith prompts generated by a context-aware LLM, ensuring both data diversity and\ncontrol of bias in synthetic data. To resolve domain and bias shifts, we\nintroduce a novel selective fine-tuning scheme in which only model parameters\nmore sensitive to bias and less sensitive to domain shift are updated.\nExperiments on CelebA and UTKFace datasets show that our AIM-Fair improves\nmodel fairness while maintaining utility, outperforming both fully and\npartially fine-tuned approaches to model fairness.\n", "date": "2025-03-07", "categories": []}
{"id": "2503.05684", "title": "Fairness-Aware Low-Rank Adaptation Under Demographic Privacy Constraints", "abstract": "  Pre-trained foundation models can be adapted for specific tasks using\nLow-Rank Adaptation (LoRA). However, the fairness properties of these adapted\nclassifiers remain underexplored. Existing fairness-aware fine-tuning methods\nrely on direct access to sensitive attributes or their predictors, but in\npractice, these sensitive attributes are often held under strict consumer\nprivacy controls, and neither the attributes nor their predictors are available\nto model developers, hampering the development of fair models. To address this\nissue, we introduce a set of LoRA-based fine-tuning methods that can be trained\nin a distributed fashion, where model developers and fairness auditors\ncollaborate without sharing sensitive attributes or predictors. In this paper,\nwe evaluate three such methods - sensitive unlearning, adversarial training,\nand orthogonality loss - against a fairness-unaware baseline, using experiments\non the CelebA and UTK-Face datasets with an ImageNet pre-trained ViT-Base\nmodel. We find that orthogonality loss consistently reduces bias while\nmaintaining or improving utility, whereas adversarial training improves False\nPositive Rate Parity and Demographic Parity in some cases, and sensitive\nunlearning provides no clear benefit. In tasks where significant biases are\npresent, distributed fairness-aware fine-tuning methods can effectively\neliminate bias without compromising consumer privacy and, in most cases,\nimprove model utility.\n", "date": "2025-03-07", "categories": []}
{"id": "2401.13097", "title": "Digital Divides in Scene Recognition: Uncovering Socioeconomic Biases in\n  Deep Learning Systems", "abstract": "  Computer-based scene understanding has influenced fields ranging from urban\nplanning to autonomous vehicle performance, yet little is known about how well\nthese technologies work across social differences. We investigate the biases of\ndeep convolutional neural networks (dCNNs) in scene classification, using\nnearly one million images from global and US sources, including user-submitted\nhome photographs and Airbnb listings. We applied statistical models to quantify\nthe impact of socioeconomic indicators such as family income, Human Development\nIndex (HDI), and demographic factors from public data sources (CIA and US\nCensus) on dCNN performance. Our analyses revealed significant socioeconomic\nbias, where pretrained dCNNs demonstrated lower classification accuracy, lower\nclassification confidence, and a higher tendency to assign labels that could be\noffensive when applied to homes (e.g., \"ruin\", \"slum\"), especially in images\nfrom homes with lower socioeconomic status (SES). This trend is consistent\nacross two datasets of international images and within the diverse economic and\nracial landscapes of the United States. This research contributes to\nunderstanding biases in computer vision, emphasizing the need for more\ninclusive and representative training datasets. By mitigating the bias in the\ncomputer vision pipelines, we can ensure fairer and more equitable outcomes for\napplied computer vision, including home valuation and smart home security\nsystems. There is urgency in addressing these biases, which can significantly\nimpact critical decisions in urban development and resource allocation. Our\nfindings also motivate the development of AI systems that better understand and\nserve diverse communities, moving towards technology that equitably benefits\nall sectors of society.\n", "date": "2024-01-23", "categories": []}
{"id": "2503.02865", "title": "FairSense-AI: Responsible AI Meets Sustainability", "abstract": "  In this paper, we introduce FairSense-AI: a multimodal framework designed to\ndetect and mitigate bias in both text and images. By leveraging Large Language\nModels (LLMs) and Vision-Language Models (VLMs), FairSense-AI uncovers subtle\nforms of prejudice or stereotyping that can appear in content, providing users\nwith bias scores, explanatory highlights, and automated recommendations for\nfairness enhancements. In addition, FairSense-AI integrates an AI risk\nassessment component that aligns with frameworks like the MIT AI Risk\nRepository and NIST AI Risk Management Framework, enabling structured\nidentification of ethical and safety concerns. The platform is optimized for\nenergy efficiency via techniques such as model pruning and mixed-precision\ncomputation, thereby reducing its environmental footprint. Through a series of\ncase studies and applications, we demonstrate how FairSense-AI promotes\nresponsible AI use by addressing both the social dimension of fairness and the\npressing need for sustainability in large-scale AI deployments.\nhttps://vectorinstitute.github.io/FairSense-AI,\nhttps://pypi.org/project/fair-sense-ai/ (Sustainability , Responsible AI ,\nLarge Language Models , Vision Language Models , Ethical AI , Green AI)\n", "date": "2025-03-04", "categories": []}
{"id": "2212.14351", "title": "Properties of Group Fairness Metrics for Rankings", "abstract": "  In recent years, several metrics have been developed for evaluating group\nfairness of rankings. Given that these metrics were developed with different\napplication contexts and ranking algorithms in mind, it is not straightforward\nwhich metric to choose for a given scenario. In this paper, we perform a\ncomprehensive comparative analysis of existing group fairness metrics developed\nin the context of fair ranking. By virtue of their diverse application\ncontexts, we argue that such a comparative analysis is not straightforward.\nHence, we take an axiomatic approach whereby we design a set of thirteen\nproperties for group fairness metrics that consider different ranking settings.\nA metric can then be selected depending on whether it satisfies all or a subset\nof these properties. We apply these properties on eleven existing group\nfairness metrics, and through both empirical and theoretical results we\ndemonstrate that most of these metrics only satisfy a small subset of the\nproposed properties. These findings highlight limitations of existing metrics,\nand provide insights into how to evaluate and interpret different fairness\nmetrics in practical deployment. The proposed properties can also assist\npractitioners in selecting appropriate metrics for evaluating fairness in a\nspecific application.\n", "date": "2022-12-29", "categories": []}
{"id": "2407.03153", "title": "An Efficient Framework for Crediting Data Contributors of Diffusion\n  Models", "abstract": "  As diffusion models are deployed in real-world settings, and their\nperformance is driven by training data, appraising the contribution of data\ncontributors is crucial to creating incentives for sharing quality data and to\nimplementing policies for data compensation. Depending on the use case, model\nperformance corresponds to various global properties of the distribution\nlearned by a diffusion model (e.g., overall aesthetic quality). Hence, here we\naddress the problem of attributing global properties of diffusion models to\ndata contributors. The Shapley value provides a principled approach to\nvaluation by uniquely satisfying game-theoretic axioms of fairness. However,\nestimating Shapley values for diffusion models is computationally impractical\nbecause it requires retraining on many training data subsets corresponding to\ndifferent contributors and rerunning inference. We introduce a method to\nefficiently retrain and rerun inference for Shapley value estimation, by\nleveraging model pruning and fine-tuning. We evaluate the utility of our method\nwith three use cases: (i) image quality for a DDPM trained on a CIFAR dataset,\n(ii) demographic diversity for an LDM trained on CelebA-HQ, and (iii) aesthetic\nquality for a Stable Diffusion model LoRA-finetuned on Post-Impressionist\nartworks. Our results empirically demonstrate that our framework can identify\nimportant data contributors across models' global properties, outperforming\nexisting attribution methods for diffusion models.\n", "date": "2024-06-09", "categories": []}
{"id": "2403.08743", "title": "Prompting Fairness: Integrating Causality to Debias Large Language\n  Models", "abstract": "  Large language models (LLMs), despite their remarkable capabilities, are\nsusceptible to generating biased and discriminatory responses. As LLMs\nincreasingly influence high-stakes decision-making (e.g., hiring and\nhealthcare), mitigating these biases becomes critical. In this work, we propose\na causality-guided debiasing framework to tackle social biases, aiming to\nreduce the objectionable dependence between LLMs' decisions and the social\ninformation in the input. Our framework introduces a novel perspective to\nidentify how social information can affect an LLM's decision through different\ncausal pathways. Leveraging these causal insights, we outline principled\nprompting strategies that regulate these pathways through selection mechanisms.\nThis framework not only unifies existing prompting-based debiasing techniques,\nbut also opens up new directions for reducing bias by encouraging the model to\nprioritize fact-based reasoning over reliance on biased social cues. We\nvalidate our framework through extensive experiments on real-world datasets\nacross multiple domains, demonstrating its effectiveness in debiasing LLM\ndecisions, even with only black-box access to the model.\n", "date": "2024-03-13", "categories": []}
{"id": "2410.19803", "title": "First-Person Fairness in Chatbots", "abstract": "  Evaluating chatbot fairness is crucial given their rapid proliferation, yet\ntypical chatbot tasks (e.g., resume writing, entertainment) diverge from the\ninstitutional decision-making tasks (e.g., resume screening) which have\ntraditionally been central to discussion of algorithmic fairness. The\nopen-ended nature and diverse use-cases of chatbots necessitate novel methods\nfor bias assessment. This paper addresses these challenges by introducing a\nscalable counterfactual approach to evaluate \"first-person fairness,\" meaning\nfairness toward chatbot users based on demographic characteristics. Our method\nemploys a Language Model as a Research Assistant (LMRA) to yield quantitative\nmeasures of harmful stereotypes and qualitative analyses of demographic\ndifferences in chatbot responses. We apply this approach to assess biases in\nsix of our language models across millions of interactions, covering sixty-six\ntasks in nine domains and spanning two genders and four races. Independent\nhuman annotations corroborate the LMRA-generated bias evaluations. This study\nrepresents the first large-scale fairness evaluation based on real-world chat\ndata. We highlight that post-training reinforcement learning techniques\nsignificantly mitigate these biases. This evaluation provides a practical\nmethodology for ongoing bias monitoring and mitigation.\n", "date": "2024-10-16", "categories": []}
{"id": "2412.19495", "title": "Disparate Model Performance and Stability in Machine Learning Clinical\n  Support for Diabetes and Heart Diseases", "abstract": "  Machine Learning (ML) algorithms are vital for supporting clinical\ndecision-making in biomedical informatics. However, their predictive\nperformance can vary across demographic groups, often due to the\nunderrepresentation of historically marginalized populations in training\ndatasets. The investigation reveals widespread sex- and age-related inequities\nin chronic disease datasets and their derived ML models. Thus, a novel\nanalytical framework is introduced, combining systematic arbitrariness with\ntraditional metrics like accuracy and data complexity. The analysis of data\nfrom over 25,000 individuals with chronic diseases revealed mild sex-related\ndisparities, favoring predictive accuracy for males, and significant\nage-related differences, with better accuracy for younger patients. Notably,\nolder patients showed inconsistent predictive accuracy across seven datasets,\nlinked to higher data complexity and lower model performance. This highlights\nthat representativeness in training data alone does not guarantee equitable\noutcomes, and model arbitrariness must be addressed before deploying models in\nclinical settings.\n", "date": "2024-12-27", "categories": []}
{"id": "2502.07254", "title": "Fairness in Agentic AI: A Unified Framework for Ethical and Equitable\n  Multi-Agent System", "abstract": "  Ensuring fairness in decentralized multi-agent systems presents significant\nchallenges due to emergent biases, systemic inefficiencies, and conflicting\nagent incentives. This paper provides a comprehensive survey of fairness in\nmulti-agent AI, introducing a novel framework where fairness is treated as a\ndynamic, emergent property of agent interactions. The framework integrates\nfairness constraints, bias mitigation strategies, and incentive mechanisms to\nalign autonomous agent behaviors with societal values while balancing\nefficiency and robustness. Through empirical validation, we demonstrate that\nincorporating fairness constraints results in more equitable decision-making.\nThis work bridges the gap between AI ethics and system design, offering a\nfoundation for accountable, transparent, and socially responsible multi-agent\nAI systems.\n", "date": "2025-02-11", "categories": []}
{"id": "2503.00299", "title": "Hidden Convexity of Fair PCA and Fast Solver via Eigenvalue Optimization", "abstract": "  Principal Component Analysis (PCA) is a foundational technique in machine\nlearning for dimensionality reduction of high-dimensional datasets. However,\nPCA could lead to biased outcomes that disadvantage certain subgroups of the\nunderlying datasets. To address the bias issue, a Fair PCA (FPCA) model was\nintroduced by Samadi et al. (2018) for equalizing the reconstruction loss\nbetween subgroups. The semidefinite relaxation (SDR) based approach proposed by\nSamadi et al. (2018) is computationally expensive even for suboptimal\nsolutions. To improve efficiency, several alternative variants of the FPCA\nmodel have been developed. These variants often shift the focus away from\nequalizing the reconstruction loss. In this paper, we identify a hidden\nconvexity in the FPCA model and introduce an algorithm for convex optimization\nvia eigenvalue optimization. Our approach achieves the desired fairness in\nreconstruction loss without sacrificing performance. As demonstrated in\nreal-world datasets, the proposed FPCA algorithm runs $8\\times$ faster than the\nSDR-based algorithm, and only at most 85% slower than the standard PCA.\n", "date": "2025-03-01", "categories": []}
{"id": "2503.00355", "title": "Structured Reasoning for Fairness: A Multi-Agent Approach to Bias\n  Detection in Textual Data", "abstract": "  From disinformation spread by AI chatbots to AI recommendations that\ninadvertently reinforce stereotypes, textual bias poses a significant challenge\nto the trustworthiness of large language models (LLMs). In this paper, we\npropose a multi-agent framework that systematically identifies biases by\ndisentangling each statement as fact or opinion, assigning a bias intensity\nscore, and providing concise, factual justifications. Evaluated on 1,500\nsamples from the WikiNPOV dataset, the framework achieves 84.9%\naccuracy$\\unicode{x2014}$an improvement of 13.0% over the zero-shot\nbaseline$\\unicode{x2014}$demonstrating the efficacy of explicitly modeling fact\nversus opinion prior to quantifying bias intensity. By combining enhanced\ndetection accuracy with interpretable explanations, this approach sets a\nfoundation for promoting fairness and accountability in modern language models.\n", "date": "2025-03-01", "categories": []}
{"id": "2503.00650", "title": "The Hidden Cost of Waiting for Accurate Predictions", "abstract": "  Algorithmic predictions are increasingly informing societal resource\nallocations by identifying individuals for targeting. Policymakers often build\nthese systems with the assumption that by gathering more observations on\nindividuals, they can improve predictive accuracy and, consequently, allocation\nefficiency. An overlooked yet consequential aspect of prediction-driven\nallocations is that of timing. The planner has to trade off relying on earlier\nand potentially noisier predictions to intervene before individuals experience\nundesirable outcomes, or they may wait to gather more observations to make more\nprecise allocations. We examine this tension using a simple mathematical model,\nwhere the planner collects observations on individuals to improve predictions\nover time. We analyze both the ranking induced by these predictions and optimal\nresource allocation. We show that though individual prediction accuracy\nimproves over time, counter-intuitively, the average ranking loss can worsen.\nAs a result, the planner's ability to improve social welfare can decline. We\nidentify inequality as a driving factor behind this phenomenon. Our findings\nprovide a nuanced perspective and challenge the conventional wisdom that it is\npreferable to wait for more accurate predictions to ensure the most efficient\nallocations.\n", "date": "2025-03-01", "categories": []}
{"id": "2502.20898", "title": "A database to support the evaluation of gender biases in GPT-4o output", "abstract": "  The widespread application of Large Language Models (LLMs) involves ethical\nrisks for users and societies. A prominent ethical risk of LLMs is the\ngeneration of unfair language output that reinforces or exacerbates harm for\nmembers of disadvantaged social groups through gender biases (Weidinger et al.,\n2022; Bender et al., 2021; Kotek et al., 2023). Hence, the evaluation of the\nfairness of LLM outputs with respect to such biases is a topic of rising\ninterest. To advance research in this field, promote discourse on suitable\nnormative bases and evaluation methodologies, and enhance the reproducibility\nof related studies, we propose a novel approach to database construction. This\napproach enables the assessment of gender-related biases in LLM-generated\nlanguage beyond merely evaluating their degree of neutralization.\n", "date": "2025-02-28", "categories": []}
{"id": "2408.11843", "title": "Identifying and Mitigating Social Bias Knowledge in Language Models", "abstract": "  Generating fair and accurate predictions plays a pivotal role in deploying\nlarge language models (LLMs) in the real world. However, existing debiasing\nmethods inevitably generate unfair or incorrect predictions as they are\ndesigned and evaluated to achieve parity across different social groups but\nleave aside individual commonsense facts, resulting in modified knowledge that\nelicits unreasonable or undesired predictions. In this paper, we first\nestablish a new bias mitigation benchmark, BiaScope, which systematically\nassesses performance by leveraging newly constructed datasets and metrics on\nknowledge retention and generalization. Then, we propose a novel debiasing\napproach, Fairness Stamp (FAST), which enables fine-grained calibration of\nindividual social biases. FAST identifies the decisive layer responsible for\nstoring social biases and then calibrates its outputs by integrating a small\nmodular network, considering both bias mitigation and knowledge-preserving\ndemands. Comprehensive experiments demonstrate that FAST surpasses\nstate-of-the-art baselines with superior debiasing performance while not\ncompromising the overall model capability for knowledge retention and\ndownstream predictions. This highlights the potential of fine-grained debiasing\nstrategies to achieve fairness in LLMs.\n", "date": "2024-08-07", "categories": []}
{"id": "2501.01441", "title": "Explanatory Debiasing: Involving Domain Experts in the Data Generation\n  Process to Mitigate Representation Bias in AI Systems", "abstract": "  Representation bias is one of the most common types of biases in artificial\nintelligence (AI) systems, causing AI models to perform poorly on\nunderrepresented data segments. Although AI practitioners use various methods\nto reduce representation bias, their effectiveness is often constrained by\ninsufficient domain knowledge in the debiasing process. To address this gap,\nthis paper introduces a set of generic design guidelines for effectively\ninvolving domain experts in representation debiasing. We instantiated our\nproposed guidelines in a healthcare-focused application and evaluated them\nthrough a comprehensive mixed-methods user study with 35 healthcare experts.\nOur findings show that involving domain experts can reduce representation bias\nwithout compromising model accuracy. Based on our findings, we also offer\nrecommendations for developers to build robust debiasing systems guided by our\ngeneric design guidelines, ensuring more effective inclusion of domain experts\nin the debiasing process.\n", "date": "2024-12-26", "categories": []}
{"id": "2411.08425", "title": "Properties of fairness measures in the context of varying class\n  imbalance and protected group ratios", "abstract": "  Society is increasingly relying on predictive models in fields like criminal\njustice, credit risk management, or hiring. To prevent such automated systems\nfrom discriminating against people belonging to certain groups, fairness\nmeasures have become a crucial component in socially relevant applications of\nmachine learning. However, existing fairness measures have been designed to\nassess the bias between predictions for protected groups without considering\nthe imbalance in the classes of the target variable. Current research on the\npotential effect of class imbalance on fairness focuses on practical\napplications rather than dataset-independent measure properties. In this paper,\nwe study the general properties of fairness measures for changing class and\nprotected group proportions. For this purpose, we analyze the probability mass\nfunctions of six of the most popular group fairness measures. We also measure\nhow the probability of achieving perfect fairness changes for varying class\nimbalance ratios. Moreover, we relate the dataset-independent properties of\nfairness measures described in this paper to classifier fairness in real-life\ntasks. Our results show that measures such as Equal Opportunity and Positive\nPredictive Parity are more sensitive to changes in class imbalance than\nAccuracy Equality. These findings can help guide researchers and practitioners\nin choosing the most appropriate fairness measures for their classification\nproblems.\n", "date": "2024-11-13", "categories": []}
{"id": "2502.17921", "title": "Unmasking Gender Bias in Recommendation Systems and Enhancing\n  Category-Aware Fairness", "abstract": "  Recommendation systems are now an integral part of our daily lives. We rely\non them for tasks such as discovering new movies, finding friends on social\nmedia, and connecting job seekers with relevant opportunities. Given their\nvital role, we must ensure these recommendations are free from societal\nstereotypes. Therefore, evaluating and addressing such biases in recommendation\nsystems is crucial. Previous work evaluating the fairness of recommended items\nfails to capture certain nuances as they mainly focus on comparing performance\nmetrics for different sensitive groups. In this paper, we introduce a set of\ncomprehensive metrics for quantifying gender bias in recommendations.\nSpecifically, we show the importance of evaluating fairness on a more granular\nlevel, which can be achieved using our metrics to capture gender bias using\ncategories of recommended items like genres for movies. Furthermore, we show\nthat employing a category-aware fairness metric as a regularization term along\nwith the main recommendation loss during training can help effectively minimize\nbias in the models' output. We experiment on three real-world datasets, using\nfive baseline models alongside two popular fairness-aware models, to show the\neffectiveness of our metrics in evaluating gender bias. Our metrics help\nprovide an enhanced insight into bias in recommended items compared to previous\nmetrics. Additionally, our results demonstrate how incorporating our\nregularization term significantly improves the fairness in recommendations for\ndifferent categories without substantial degradation in overall recommendation\nperformance.\n", "date": "2025-02-25", "categories": []}
{"id": "2502.18060", "title": "Defining bias in AI-systems: Biased models are fair models", "abstract": "  The debate around bias in AI systems is central to discussions on algorithmic\nfairness. However, the term bias often lacks a clear definition, despite\nfrequently being contrasted with fairness, implying that an unbiased model is\ninherently fair. In this paper, we challenge this assumption and argue that a\nprecise conceptualization of bias is necessary to effectively address fairness\nconcerns. Rather than viewing bias as inherently negative or unfair, we\nhighlight the importance of distinguishing between bias and discrimination. We\nfurther explore how this shift in focus can foster a more constructive\ndiscourse within academic debates on fairness in AI systems.\n", "date": "2025-02-25", "categories": []}
{"id": "2502.18434", "title": "Exploring Gender Disparities in Automatic Speech Recognition Technology", "abstract": "  This study investigates factors influencing Automatic Speech Recognition\n(ASR) systems' fairness and performance across genders, beyond the conventional\nexamination of demographics. Using the LibriSpeech dataset and the Whisper\nsmall model, we analyze how performance varies across different gender\nrepresentations in training data. Our findings suggest a complex interplay\nbetween the gender ratio in training data and ASR performance. Optimal fairness\noccurs at specific gender distributions rather than a simple 50-50 split.\nFurthermore, our findings suggest that factors like pitch variability can\nsignificantly affect ASR accuracy. This research contributes to a deeper\nunderstanding of biases in ASR systems, highlighting the importance of\ncarefully curated training data in mitigating gender bias.\n", "date": "2025-02-25", "categories": []}
{"id": "2405.17814", "title": "FAIntbench: A Holistic and Precise Benchmark for Bias Evaluation in\n  Text-to-Image Models", "abstract": "  The rapid development and reduced barriers to entry for Text-to-Image (T2I)\nmodels have raised concerns about the biases in their outputs, but existing\nresearch lacks a holistic definition and evaluation framework of biases,\nlimiting the enhancement of debiasing techniques. To address this issue, we\nintroduce FAIntbench, a holistic and precise benchmark for biases in T2I\nmodels. In contrast to existing benchmarks that evaluate bias in limited\naspects, FAIntbench evaluate biases from four dimensions: manifestation of\nbias, visibility of bias, acquired attributes, and protected attributes. We\napplied FAIntbench to evaluate seven recent large-scale T2I models and\nconducted human evaluation, whose results demonstrated the effectiveness of\nFAIntbench in identifying various biases. Our study also revealed new research\nquestions about biases, including the side-effect of distillation. The findings\npresented here are preliminary, highlighting the potential of FAIntbench to\nadvance future research aimed at mitigating the biases in T2I models. Our\nbenchmark is publicly available to ensure the reproducibility.\n", "date": "2024-05-28", "categories": []}
{"id": "2407.02408", "title": "CEB: Compositional Evaluation Benchmark for Fairness in Large Language\n  Models", "abstract": "  As Large Language Models (LLMs) are increasingly deployed to handle various\nnatural language processing (NLP) tasks, concerns regarding the potential\nnegative societal impacts of LLM-generated content have also arisen. To\nevaluate the biases exhibited by LLMs, researchers have recently proposed a\nvariety of datasets. However, existing bias evaluation efforts often focus on\nonly a particular type of bias and employ inconsistent evaluation metrics,\nleading to difficulties in comparison across different datasets and LLMs. To\naddress these limitations, we collect a variety of datasets designed for the\nbias evaluation of LLMs, and further propose CEB, a Compositional Evaluation\nBenchmark that covers different types of bias across different social groups\nand tasks. The curation of CEB is based on our newly proposed compositional\ntaxonomy, which characterizes each dataset from three dimensions: bias types,\nsocial groups, and tasks. By combining the three dimensions, we develop a\ncomprehensive evaluation strategy for the bias in LLMs. Our experiments\ndemonstrate that the levels of bias vary across these dimensions, thereby\nproviding guidance for the development of specific bias mitigation methods.\n", "date": "2024-07-02", "categories": []}
{"id": "2408.12494", "title": "GenderCARE: A Comprehensive Framework for Assessing and Reducing Gender\n  Bias in Large Language Models", "abstract": "  Large language models (LLMs) have exhibited remarkable capabilities in\nnatural language generation, but they have also been observed to magnify\nsocietal biases, particularly those related to gender. In response to this\nissue, several benchmarks have been proposed to assess gender bias in LLMs.\nHowever, these benchmarks often lack practical flexibility or inadvertently\nintroduce biases. To address these shortcomings, we introduce GenderCARE, a\ncomprehensive framework that encompasses innovative Criteria, bias Assessment,\nReduction techniques, and Evaluation metrics for quantifying and mitigating\ngender bias in LLMs. To begin, we establish pioneering criteria for gender\nequality benchmarks, spanning dimensions such as inclusivity, diversity,\nexplainability, objectivity, robustness, and realisticity. Guided by these\ncriteria, we construct GenderPair, a novel pair-based benchmark designed to\nassess gender bias in LLMs comprehensively. Our benchmark provides standardized\nand realistic evaluations, including previously overlooked gender groups such\nas transgender and non-binary individuals. Furthermore, we develop effective\ndebiasing techniques that incorporate counterfactual data augmentation and\nspecialized fine-tuning strategies to reduce gender bias in LLMs without\ncompromising their overall performance. Extensive experiments demonstrate a\nsignificant reduction in various gender bias benchmarks, with reductions\npeaking at over 90% and averaging above 35% across 17 different LLMs.\nImportantly, these reductions come with minimal variability in mainstream\nlanguage tasks, remaining below 2%. By offering a realistic assessment and\ntailored reduction of gender biases, we hope that our GenderCARE can represent\na significant step towards achieving fairness and equity in LLMs. More details\nare available at https://github.com/kstanghere/GenderCARE-ccs24.\n", "date": "2024-08-22", "categories": []}
{"id": "2502.16933", "title": "Achieving Fair PCA Using Joint Eigenvalue Decomposition", "abstract": "  Principal Component Analysis (PCA) is a widely used method for dimensionality\nreduction, but it often overlooks fairness, especially when working with data\nthat includes demographic characteristics. This can lead to biased\nrepresentations that disproportionately affect certain groups. To address this\nissue, our approach incorporates Joint Eigenvalue Decomposition (JEVD), a\ntechnique that enables the simultaneous diagonalization of multiple matrices,\nensuring fair and efficient representations. We formally show that the optimal\nsolution of JEVD leads to a fair PCA solution. By integrating JEVD with PCA, we\nstrike an optimal balance between preserving data structure and promoting\nfairness across diverse groups. We demonstrate that our method outperforms\nexisting baseline approaches in fairness and representational quality on\nvarious datasets. It retains the core advantages of PCA while ensuring that\nsensitive demographic attributes do not create disparities in the reduced\nrepresentation.\n", "date": "2025-02-24", "categories": []}
{"id": "2408.16130", "title": "Using Backbone Foundation Model for Evaluating Fairness in Chest\n  Radiography Without Demographic Data", "abstract": "  Ensuring consistent performance across diverse populations and incorporating\nfairness into machine learning models are crucial for advancing medical image\ndiagnostics and promoting equitable healthcare. However, many databases do not\nprovide protected attributes or contain unbalanced representations of\ndemographic groups, complicating the evaluation of model performance across\ndifferent demographics and the application of bias mitigation techniques that\nrely on these attributes. This study aims to investigate the effectiveness of\nusing the backbone of Foundation Models as an embedding extractor for creating\ngroups that represent protected attributes, such as gender and age. We propose\nutilizing these groups in different stages of bias mitigation, including\npre-processing, in-processing, and evaluation. Using databases in and\nout-of-distribution scenarios, it is possible to identify that the method can\ncreate groups that represent gender in both databases and reduce in 4.44% the\ndifference between the gender attribute in-distribution and 6.16% in\nout-of-distribution. However, the model lacks robustness in handling age\nattributes, underscoring the need for more fundamentally fair and robust\nFoundation models. These findings suggest a role in promoting fairness\nassessment in scenarios where we lack knowledge of attributes, contributing to\nthe development of more equitable medical diagnostics.\n", "date": "2024-08-28", "categories": []}
{"id": "2412.04767", "title": "Towards counterfactual fairness through auxiliary variables", "abstract": "  The challenge of balancing fairness and predictive accuracy in machine\nlearning models, especially when sensitive attributes such as race, gender, or\nage are considered, has motivated substantial research in recent years.\nCounterfactual fairness ensures that predictions remain consistent across\ncounterfactual variations of sensitive attributes, which is a crucial concept\nin addressing societal biases. However, existing counterfactual fairness\napproaches usually overlook intrinsic information about sensitive features,\nlimiting their ability to achieve fairness while simultaneously maintaining\nperformance. To tackle this challenge, we introduce EXOgenous Causal reasoning\n(EXOC), a novel causal reasoning framework motivated by exogenous variables. It\nleverages auxiliary variables to uncover intrinsic properties that give rise to\nsensitive attributes. Our framework explicitly defines an auxiliary node and a\ncontrol node that contribute to counterfactual fairness and control the\ninformation flow within the model. Our evaluation, conducted on synthetic and\nreal-world datasets, validates EXOC's superiority, showing that it outperforms\nstate-of-the-art approaches in achieving counterfactual fairness. Our code is\navailable at https://github.com/CASE-Lab-UMD/counterfactual_fairness_2025.\n", "date": "2024-12-06", "categories": []}
{"id": "2405.14891", "title": "Auditing the Fairness of the US COVID-19 Forecast Hub's Case Prediction\n  Models", "abstract": "  The US COVID-19 Forecast Hub, a repository of COVID-19 forecasts from over 50\nindependent research groups, is used by the Centers for Disease Control and\nPrevention (CDC) for their official COVID-19 communications. As such, the\nForecast Hub is a critical centralized resource to promote transparent decision\nmaking. While the Forecast Hub has provided valuable predictions focused on\naccuracy, there is an opportunity to evaluate model performance across social\ndeterminants such as race and urbanization level that have been known to play a\nrole in the COVID-19 pandemic. In this paper, we carry out a comprehensive\nfairness analysis of the Forecast Hub model predictions and we show\nstatistically significant diverse predictive performance across social\ndeterminants, with minority racial and ethnic groups as well as less urbanized\nareas often associated with higher prediction errors. We hope this work will\nencourage COVID-19 modelers and the CDC to report fairness metrics together\nwith accuracy, and to reflect on the potential harms of the models on specific\nsocial groups and contexts.\n", "date": "2024-05-17", "categories": []}
{"id": "2501.11752", "title": "Are generative models fair? A study of racial bias in dermatological\n  image generation", "abstract": "  Racial bias in medicine, such as in dermatology, presents significant ethical\nand clinical challenges. This is likely to happen because there is a\nsignificant underrepresentation of darker skin tones in training datasets for\nmachine learning models. While efforts to address bias in dermatology have\nfocused on improving dataset diversity and mitigating disparities in\ndiscriminative models, the impact of racial bias on generative models remains\nunderexplored. Generative models, such as Variational Autoencoders (VAEs), are\nincreasingly used in healthcare applications, yet their fairness across diverse\nskin tones is currently not well understood. In this study, we evaluate the\nfairness of generative models in clinical dermatology with respect to racial\nbias. For this purpose, we first train a VAE with a perceptual loss to generate\nand reconstruct high-quality skin images across different skin tones. We\nutilize the Fitzpatrick17k dataset to examine how racial bias influences the\nrepresentation and performance of these models. Our findings indicate that VAE\nperformance is, as expected, influenced by representation, i.e. increased skin\ntone representation comes with increased performance on the given skin tone.\nHowever, we also observe, even independently of representation, that the VAE\nperforms better for lighter skin tones. Additionally, the uncertainty estimates\nproduced by the VAE are ineffective in assessing the model's fairness. These\nresults highlight the need for more representative dermatological datasets, but\nalso a need for better understanding the sources of bias in such model, as well\nas improved uncertainty quantification mechanisms to detect and address racial\nbias in generative models for trustworthy healthcare technologies.\n", "date": "2025-01-20", "categories": []}
{"id": "2502.13221", "title": "Two Tickets are Better than One: Fair and Accurate Hiring Under\n  Strategic LLM Manipulations", "abstract": "  In an era of increasingly capable foundation models, job seekers are turning\nto generative AI tools to enhance their application materials. However, unequal\naccess to and knowledge about generative AI tools can harm both employers and\ncandidates by reducing the accuracy of hiring decisions and giving some\ncandidates an unfair advantage. To address these challenges, we introduce a new\nvariant of the strategic classification framework tailored to manipulations\nperformed using large language models, accommodating varying levels of\nmanipulations and stochastic outcomes. We propose a ``two-ticket'' scheme,\nwhere the hiring algorithm applies an additional manipulation to each submitted\nresume and considers this manipulated version together with the original\nsubmitted resume. We establish theoretical guarantees for this scheme, showing\nimprovements for both the fairness and accuracy of hiring decisions when the\ntrue positive rate is maximized subject to a no false positives constraint. We\nfurther generalize this approach to an $n$-ticket scheme and prove that hiring\noutcomes converge to a fixed, group-independent decision, eliminating\ndisparities arising from differential LLM access. Finally, we empirically\nvalidate our framework and the performance of our two-ticket scheme on real\nresumes using an open-source resume screening tool.\n", "date": "2025-02-18", "categories": []}
{"id": "2502.12874", "title": "Testing for Causal Fairness", "abstract": "  Causality is widely used in fairness analysis to prevent discrimination on\nsensitive attributes, such as genders in career recruitment and races in crime\nprediction. However, the current data-based Potential Outcomes Framework (POF)\noften leads to untrustworthy fairness analysis results when handling\nhigh-dimensional data. To address this, we introduce a distribution-based POF\nthat transform fairness analysis into Distributional Closeness Testing (DCT) by\nintervening on sensitive attributes. We define counterfactual closeness\nfairness as the null hypothesis of DCT, where a sensitive attribute is\nconsidered fair if its factual and counterfactual potential outcome\ndistributions are sufficiently close. We introduce the Norm-Adaptive Maximum\nMean Discrepancy Treatment Effect (N-TE) as a statistic for measuring\ndistributional closeness and apply DCT using the empirical estimator of NTE,\nreferred to Counterfactual Fairness-CLOseness Testing ($\\textrm{CF-CLOT}$). To\nensure the trustworthiness of testing results, we establish the testing\nconsistency of N-TE through rigorous theoretical analysis. $\\textrm{CF-CLOT}$\ndemonstrates sensitivity in fairness analysis through the flexibility of the\ncloseness parameter $\\epsilon$. Unfair sensitive attributes have been\nsuccessfully tested by $\\textrm{CF-CLOT}$ in extensive experiments across\nvarious real-world scenarios, which validate the consistency of the testing.\n", "date": "2025-02-18", "categories": []}
{"id": "2307.05333", "title": "Wearable-based Fair and Accurate Pain Assessment Using Multi-Attribute\n  Fairness Loss in Convolutional Neural Networks", "abstract": "  The integration of diverse health data, such as IoT (Internet of Things), EHR\n(Electronic Health Record), and clinical surveys, with scalable AI(Artificial\nIntelligence) has enabled the identification of physical, behavioral, and\npsycho-social indicators of pain. However, the adoption of AI in clinical pain\nevaluation is hindered by challenges like personalization and fairness. Many AI\nmodels, including machine and deep learning, exhibit biases, discriminating\nagainst specific groups based on gender or ethnicity, causing skepticism among\nmedical professionals about their reliability. This paper proposes a\nMulti-attribute Fairness Loss (MAFL) based Convolutional Neural Network (CNN)\nmodel designed to account for protected attributes in data, ensuring fair pain\nstatus predictions while minimizing disparities between privileged and\nunprivileged groups. We evaluate whether a balance between accuracy and\nfairness is achievable by comparing the proposed model with existing mitigation\nmethods. Our findings indicate that the model performs favorably against\nstate-of-the-art techniques. Using the NIH All-Of-US dataset, comprising data\nfrom 868 individuals over 1500 days, we demonstrate our model's effectiveness,\nachieving accuracy rates between 75% and 85%.\n", "date": "2023-07-03", "categories": []}
{"id": "2411.04295", "title": "Fairness with Exponential Weights", "abstract": "  Motivated by the need to remove discrimination in certain applications, we\ndevelop a meta-algorithm that can convert any efficient implementation of an\ninstance of Hedge (or equivalently, an algorithm for discrete bayesian\ninference) into an efficient algorithm for the equivalent contextual bandit\nproblem which guarantees exact statistical parity on every trial. Relative to\nany comparator with statistical parity, the resulting algorithm has the same\nasymptotic regret bound as running the corresponding instance of Exp4 for each\nprotected characteristic independently. Given that our Hedge instance admits\nnon-stationarity we can handle a varying distribution with which to enforce\nstatistical parity with respect to, which is useful when the true population is\nunknown and needs to be estimated from the data received so far. Via\nonline-to-batch conversion we can handle the equivalent batch classification\nproblem with exact statistical parity, giving us results that we believe are\nnovel and important in their own right.\n", "date": "2024-11-06", "categories": []}
{"id": "2502.10562", "title": "Detecting and Monitoring Bias for Subgroups in Breast Cancer Detection\n  AI", "abstract": "  Automated mammography screening plays an important role in early breast\ncancer detection. However, current machine learning models, developed on some\ntraining datasets, may exhibit performance degradation and bias when deployed\nin real-world settings. In this paper, we analyze the performance of\nhigh-performing AI models on two mammography datasets-the Emory Breast Imaging\nDataset (EMBED) and the RSNA 2022 challenge dataset. Specifically, we evaluate\nhow these models perform across different subgroups, defined by six attributes,\nto detect potential biases using a range of classification metrics. Our\nanalysis identifies certain subgroups that demonstrate notable\nunderperformance, highlighting the need for ongoing monitoring of these\nsubgroups' performance. To address this, we adopt a monitoring method designed\nto detect performance drifts over time. Upon identifying a drift, this method\nissues an alert, which can enable timely interventions. This approach not only\nprovides a tool for tracking the performance but also helps ensure that AI\nmodels continue to perform effectively across diverse populations.\n", "date": "2025-02-14", "categories": []}
{"id": "2502.11031", "title": "A Critical Review of Predominant Bias in Neural Networks", "abstract": "  Bias issues of neural networks garner significant attention along with its\npromising advancement. Among various bias issues, mitigating two predominant\nbiases is crucial in advancing fair and trustworthy AI: (1) ensuring neural\nnetworks yields even performance across demographic groups, and (2) ensuring\nalgorithmic decision-making does not rely on protected attributes. However,\nupon the investigation of \\pc papers in the relevant literature, we find that\nthere exists a persistent, extensive but under-explored confusion regarding\nthese two types of biases. Furthermore, the confusion has already significantly\nhampered the clarity of the community and subsequent development of debiasing\nmethodologies. Thus, in this work, we aim to restore clarity by providing two\nmathematical definitions for these two predominant biases and leveraging these\ndefinitions to unify a comprehensive list of papers. Next, we highlight the\ncommon phenomena and the possible reasons for the existing confusion. To\nalleviate the confusion, we provide extensive experiments on synthetic, census,\nand image datasets, to validate the distinct nature of these biases,\ndistinguish their different real-world manifestations, and evaluate the\neffectiveness of a comprehensive list of bias assessment metrics in assessing\nthe mitigation of these biases. Further, we compare these two types of biases\nfrom multiple dimensions including the underlying causes, debiasing methods,\nevaluation protocol, prevalent datasets, and future directions. Last, we\nprovide several suggestions aiming to guide researchers engaged in bias-related\nwork to avoid confusion and further enhance clarity in the community.\n", "date": "2025-02-16", "categories": []}
{"id": "2502.11195", "title": "From Deception to Perception: The Surprising Benefits of Deepfakes for\n  Detecting, Measuring, and Mitigating Bias", "abstract": "  While deepfake technologies have predominantly been criticized for potential\nmisuse, our study demonstrates their significant potential as tools for\ndetecting, measuring, and mitigating biases in key societal domains. By\nemploying deepfake technology to generate controlled facial images, we extend\nthe scope of traditional correspondence studies beyond mere textual\nmanipulations. This enhancement is crucial in scenarios such as pain\nassessments, where subjective biases triggered by sensitive features in facial\nimages can profoundly affect outcomes. Our results reveal that deepfakes not\nonly maintain the effectiveness of correspondence studies but also introduce\ngroundbreaking advancements in bias measurement and correction techniques. This\nstudy emphasizes the constructive role of deepfake technologies as essential\ntools for advancing societal equity and fairness.\n", "date": "2025-02-16", "categories": []}
{"id": "2502.11429", "title": "What's in a Query: Polarity-Aware Distribution-Based Fair Ranking", "abstract": "  Machine learning-driven rankings, where individuals (or items) are ranked in\nresponse to a query, mediate search exposure or attention in a variety of\nsafety-critical settings. Thus, it is important to ensure that such rankings\nare fair. Under the goal of equal opportunity, attention allocated to an\nindividual on a ranking interface should be proportional to their relevance\nacross search queries. In this work, we examine amortized fair ranking -- where\nrelevance and attention are cumulated over a sequence of user queries to make\nfair ranking more feasible in practice. Unlike prior methods that operate on\nexpected amortized attention for each individual, we define new\ndivergence-based measures for attention distribution-based fairness in ranking\n(DistFaiR), characterizing unfairness as the divergence between the\ndistribution of attention and relevance corresponding to an individual over\ntime. This allows us to propose new definitions of unfairness, which are more\nreliable at test time. Second, we prove that group fairness is upper-bounded by\nindividual fairness under this definition for a useful class of divergence\nmeasures, and experimentally show that maximizing individual fairness through\nan integer linear programming-based optimization is often beneficial to group\nfairness. Lastly, we find that prior research in amortized fair ranking ignores\ncritical information about queries, potentially leading to a fairwashing risk\nin practice by making rankings appear more fair than they actually are.\n", "date": "2025-02-17", "categories": []}
{"id": "2502.11603", "title": "DR.GAP: Mitigating Bias in Large Language Models using Gender-Aware\n  Prompting with Demonstration and Reasoning", "abstract": "  Large Language Models (LLMs) exhibit strong natural language processing\ncapabilities but also inherit and amplify societal biases, including gender\nbias, raising fairness concerns. Existing debiasing methods face significant\nlimitations: parameter tuning requires access to model weights, prompt-based\napproaches often degrade model utility, and optimization-based techniques lack\ngeneralizability. To address these challenges, we propose DR.GAP (Demonstration\nand Reasoning for Gender-Aware Prompting), an automated and model-agnostic\napproach that mitigates gender bias while preserving model performance. DR.GAP\nselects bias-revealing examples and generates structured reasoning to guide\nmodels toward more impartial responses. Extensive experiments on coreference\nresolution and QA tasks across multiple LLMs (GPT-3.5, Llama3, and\nLlama2-Alpaca) demonstrate its effectiveness, generalization ability, and\nrobustness. DR.GAP can generalize to vision-language models (VLMs), achieving\nsignificant bias reduction.\n", "date": "2025-02-17", "categories": []}
{"id": "2502.11828", "title": "Intersectional Fairness in Reinforcement Learning with Large State and\n  Constraint Spaces", "abstract": "  In traditional reinforcement learning (RL), the learner aims to solve a\nsingle objective optimization problem: find the policy that maximizes expected\nreward. However, in many real-world settings, it is important to optimize over\nmultiple objectives simultaneously. For example, when we are interested in\nfairness, states might have feature annotations corresponding to multiple\n(intersecting) demographic groups to whom reward accrues, and our goal might be\nto maximize the reward of the group receiving the minimal reward. In this work,\nwe consider a multi-objective optimization problem in which each objective is\ndefined by a state-based reweighting of a single scalar reward function. This\ngeneralizes the problem of maximizing the reward of the minimum reward group.\nWe provide oracle-efficient algorithms to solve these multi-objective RL\nproblems even when the number of objectives is exponentially large-for tabular\nMDPs, as well as for large MDPs when the group functions have additional\nstructure. Finally, we experimentally validate our theoretical results and\ndemonstrate applications on a preferential attachment graph MDP.\n", "date": "2025-02-17", "categories": []}
{"id": "2410.01888", "title": "Conformal Prediction Sets Can Cause Disparate Impact", "abstract": "  Conformal prediction is a statistically rigorous method for quantifying\nuncertainty in models by having them output sets of predictions, with larger\nsets indicating more uncertainty. However, prediction sets are not inherently\nactionable; many applications require a single output to act on, not several.\nTo overcome this limitation, prediction sets can be provided to a human who\nthen makes an informed decision. In any such system it is crucial to ensure the\nfairness of outcomes across protected groups, and researchers have proposed\nthat Equalized Coverage be used as the standard for fairness. By conducting\nexperiments with human participants, we demonstrate that providing prediction\nsets can lead to disparate impact in decisions. Disquietingly, we find that\nproviding sets that satisfy Equalized Coverage actually increases disparate\nimpact compared to marginal coverage. Instead of equalizing coverage, we\npropose to equalize set sizes across groups which empirically leads to lower\ndisparate impact.\n", "date": "2024-10-02", "categories": []}
{"id": "2410.15471", "title": "Generative Models, Humans, Predictive Models: Who Is Worse at\n  High-Stakes Decision Making?", "abstract": "  Despite strong advisory against it, large generative models (LMs) are already\nbeing used for decision making tasks that were previously done by predictive\nmodels or humans. We put popular LMs to the test in a high-stakes decision\nmaking task: recidivism prediction. Studying three closed-access and\nopen-source LMs, we analyze the LMs not exclusively in terms of accuracy, but\nalso in terms of agreement with (imperfect, noisy, and sometimes biased) human\npredictions or existing predictive models. We conduct experiments that assess\nhow providing different types of information, including distractor information\nsuch as photos, can influence LM decisions. We also stress test techniques\ndesigned to either increase accuracy or mitigate bias in LMs, and find that\nsome to have unintended consequences on LM decisions. Our results provide\nadditional quantitative evidence to the wisdom that current LMs are not the\nright tools for these types of tasks.\n", "date": "2024-10-20", "categories": []}
{"id": "2412.17803", "title": "Examining Imbalance Effects on Performance and Demographic Fairness of\n  Clinical Language Models", "abstract": "  Data imbalance is a fundamental challenge in applying language models to\nbiomedical applications, particularly in ICD code prediction tasks where label\nand demographic distributions are uneven. While state-of-the-art language\nmodels have been increasingly adopted in biomedical tasks, few studies have\nsystematically examined how data imbalance affects model performance and\nfairness across demographic groups. This study fills the gap by statistically\nprobing the relationship between data imbalance and model performance in ICD\ncode prediction. We analyze imbalances in a standard benchmark data across\ngender, age, ethnicity, and social determinants of health by state-of-the-art\nbiomedical language models. By deploying diverse performance metrics and\nstatistical analyses, we explore the influence of data imbalance on performance\nvariations and demographic fairness. Our study shows that data imbalance\nsignificantly impacts model performance and fairness, but feature similarity to\nthe majority class may be a more critical factor. We believe this study\nprovides valuable insights for developing more equitable and robust language\nmodels in healthcare applications.\n", "date": "2024-12-23", "categories": []}
{"id": "2502.09831", "title": "Learning Fair Policies for Infectious Diseases Mitigation using Path\n  Integral Control", "abstract": "  Infectious diseases pose major public health challenges to society,\nhighlighting the importance of designing effective policies to reduce economic\nloss and mortality. In this paper, we propose a framework for sequential\ndecision-making under uncertainty to design fairness-aware disease mitigation\npolicies that incorporate various measures of unfairness. Specifically, our\napproach learns equitable vaccination and lockdown strategies based on a\nstochastic multi-group SIR model. To address the challenges of solving the\nresulting sequential decision-making problem, we adopt the path integral\ncontrol algorithm as an efficient solution scheme. Through a case study, we\ndemonstrate that our approach effectively improves fairness compared to\nconventional methods and provides valuable insights for policymakers.\n", "date": "2025-02-14", "categories": []}
{"id": "2502.09319", "title": "Bridging Jensen Gap for Max-Min Group Fairness Optimization in\n  Recommendation", "abstract": "  Group max-min fairness (MMF) is commonly used in fairness-aware recommender\nsystems (RS) as an optimization objective, as it aims to protect marginalized\nitem groups and ensures a fair competition platform. However, our theoretical\nanalysis indicates that integrating MMF constraint violates the assumption of\nsample independence during optimization, causing the loss function to deviate\nfrom linear additivity. Such nonlinearity property introduces the Jensen gap\nbetween the model's convergence point and the optimal point if mini-batch\nsampling is applied. Both theoretical and empirical studies show that as the\nmini-batch size decreases and the group size increases, the Jensen gap will\nwiden accordingly. Some methods using heuristic re-weighting or debiasing\nstrategies have the potential to bridge the Jensen gap. However, they either\nlack theoretical guarantees or suffer from heavy computational costs. To\novercome these limitations, we first theoretically demonstrate that the\nMMF-constrained objective can be essentially reformulated as a group-weighted\noptimization objective. Then we present an efficient and effective algorithm\nnamed FairDual, which utilizes a dual optimization technique to minimize the\nJensen gap. Our theoretical analysis demonstrates that FairDual can achieve a\nsub-linear convergence rate to the globally optimal solution and the Jensen gap\ncan be well bounded under a mini-batch sampling strategy with random shuffle.\nExtensive experiments conducted using six large-scale RS backbone models on\nthree publicly available datasets demonstrate that FairDual outperforms all\nbaselines in terms of both accuracy and fairness. Our data and codes are shared\nat https://github.com/XuChen0427/FairDual.\n", "date": "2025-02-13", "categories": []}
{"id": "2305.16474", "title": "FairDP: Certified Fairness with Differential Privacy", "abstract": "  This paper introduces FairDP, a novel training mechanism designed to provide\ngroup fairness certification for the trained model's decisions, along with a\ndifferential privacy (DP) guarantee to protect training data. The key idea of\nFairDP is to train models for distinct individual groups independently, add\nnoise to each group's gradient for data privacy protection, and progressively\nintegrate knowledge from group models to formulate a comprehensive model that\nbalances privacy, utility, and fairness in downstream tasks. By doing so,\nFairDP ensures equal contribution from each group while gaining control over\nthe amount of DP-preserving noise added to each group's contribution. To\nprovide fairness certification, FairDP leverages the DP-preserving noise to\nstatistically quantify and bound fairness metrics. An extensive theoretical and\nempirical analysis using benchmark datasets validates the efficacy of FairDP\nand improved trade-offs between model utility, privacy, and fairness compared\nwith existing methods. Our empirical results indicate that FairDP can improve\nfairness metrics by more than 65% on average while attaining marginal utility\ndrop (less than 4% on average) under a rigorous DP-preservation across\nbenchmark datasets compared with existing baselines.\n", "date": "2023-05-25", "categories": []}
{"id": "2406.01290", "title": "Resource-constrained Fairness", "abstract": "  Access to resources strongly constrains the decisions we make. While we might\nwish to offer every student a scholarship, or schedule every patient for\nfollow-up meetings with a specialist, limited resources mean that this is not\npossible. When deploying machine learning systems, these resource constraints\nare simply enforced by varying the threshold of a classifier. However, these\nfinite resource limitations are disregarded by most existing tools for fair\nmachine learning, which do not allow the specification of resource limitations\nand do not remain fair when varying thresholds. This makes them ill-suited for\nreal-world deployment. Our research introduces the concept of\n\"resource-constrained fairness\" and quantifies the cost of fairness within this\nframework. We demonstrate that the level of available resources significantly\ninfluences this cost, a factor overlooked in previous evaluations.\n", "date": "2024-06-03", "categories": []}
{"id": "2407.11442", "title": "EARN Fairness: Explaining, Asking, Reviewing, and Negotiating Artificial\n  Intelligence Fairness Metrics Among Stakeholders", "abstract": "  Numerous fairness metrics have been proposed and employed by artificial\nintelligence (AI) experts to quantitatively measure bias and define fairness in\nAI models. Recognizing the need to accommodate stakeholders' diverse fairness\nunderstandings, efforts are underway to solicit their input. However, conveying\nAI fairness metrics to stakeholders without AI expertise, capturing their\npersonal preferences, and seeking a collective consensus remain challenging and\nunderexplored. To bridge this gap, we propose a new framework, EARN Fairness,\nwhich facilitates collective metric decisions among stakeholders without\nrequiring AI expertise. The framework features an adaptable interactive system\nand a stakeholder-centered EARN Fairness process to Explain fairness metrics,\nAsk stakeholders' personal metric preferences, Review metrics collectively, and\nNegotiate a consensus on metric selection. To gather empirical results, we\napplied the framework to a credit rating scenario and conducted a user study\ninvolving 18 decision subjects without AI knowledge. We identify their personal\nmetric preferences and their acceptable level of unfairness in individual\nsessions. Subsequently, we uncovered how they reached metric consensus in team\nsessions. Our work shows that the EARN Fairness framework enables stakeholders\nto express personal preferences and reach consensus, providing practical\nguidance for implementing human-centered AI fairness in high-risk contexts.\nThrough this approach, we aim to harmonize fairness expectations of diverse\nstakeholders, fostering more equitable and inclusive AI fairness.\n", "date": "2024-07-16", "categories": []}
{"id": "2410.04120", "title": "Rethinking Fair Representation Learning for Performance-Sensitive Tasks", "abstract": "  We investigate the prominent class of fair representation learning methods\nfor bias mitigation. Using causal reasoning to define and formalise different\nsources of dataset bias, we reveal important implicit assumptions inherent to\nthese methods. We prove fundamental limitations on fair representation learning\nwhen evaluation data is drawn from the same distribution as training data and\nrun experiments across a range of medical modalities to examine the performance\nof fair representation learning under distribution shifts. Our results explain\napparent contradictions in the existing literature and reveal how rarely\nconsidered causal and statistical aspects of the underlying data affect the\nvalidity of fair representation learning. We raise doubts about current\nevaluation practices and the applicability of fair representation learning\nmethods in performance-sensitive settings. We argue that fine-grained analysis\nof dataset biases should play a key role in the field moving forward.\n", "date": "2024-10-05", "categories": []}
{"id": "2502.05333", "title": "A Tutorial On Intersectionality in Fair Rankings", "abstract": "  We address the critical issue of biased algorithms and unfair rankings, which\nhave permeated various sectors, including search engines, recommendation\nsystems, and workforce management. These biases can lead to discriminatory\noutcomes in a data-driven world, especially against marginalized and\nunderrepresented groups. Efforts towards responsible data science and\nresponsible artificial intelligence aim to mitigate these biases and promote\nfairness, diversity, and transparency. However, most fairness-aware ranking\nmethods singularly focus on protected attributes such as race, gender, or\nsocio-economic status, neglecting the intersectionality of these attributes,\ni.e., the interplay between multiple social identities. Understanding\nintersectionality is crucial to ensure that existing inequalities are not\npreserved by fair rankings. We offer a description of the main ways to\nincorporate intersectionality in fair ranking systems through practical\nexamples and provide a comparative overview of existing literature and a\nsynoptic table summarizing the various methodologies. Our analysis highlights\nthe need for intersectionality to attain fairness, while also emphasizing that\nfairness, alone, does not necessarily imply intersectionality.\n", "date": "2025-02-07", "categories": []}
{"id": "2502.05649", "title": "Gender Bias in Instruction-Guided Speech Synthesis Models", "abstract": "  Recent advancements in controllable expressive speech synthesis, especially\nin text-to-speech (TTS) models, have allowed for the generation of speech with\nspecific styles guided by textual descriptions, known as style prompts. While\nthis development enhances the flexibility and naturalness of synthesized\nspeech, there remains a significant gap in understanding how these models\nhandle vague or abstract style prompts. This study investigates the potential\ngender bias in how models interpret occupation-related prompts, specifically\nexamining their responses to instructions like \"Act like a nurse\". We explore\nwhether these models exhibit tendencies to amplify gender stereotypes when\ninterpreting such prompts. Our experimental results reveal the model's tendency\nto exhibit gender bias for certain occupations. Moreover, models of different\nsizes show varying degrees of this bias across these occupations.\n", "date": "2025-02-08", "categories": []}
{"id": "2502.06341", "title": "Facial Analysis Systems and Down Syndrome", "abstract": "  The ethical, social and legal issues surrounding facial analysis technologies\nhave been widely debated in recent years. Key critics have argued that these\ntechnologies can perpetuate bias and discrimination, particularly against\nmarginalized groups. We contribute to this field of research by reporting on\nthe limitations of facial analysis systems with the faces of people with Down\nsyndrome: this particularly vulnerable group has received very little attention\nin the literature so far. This study involved the creation of a specific\ndataset of face images. An experimental group with faces of people with Down\nsyndrome, and a control group with faces of people who are not affected by the\nsyndrome. Two commercial tools were tested on the dataset, along three tasks:\ngender recognition, age prediction and face labelling. The results show an\noverall lower accuracy of prediction in the experimental group, and other\nspecific patterns of performance differences: i) high error rates in gender\nrecognition in the category of males with Down syndrome; ii) adults with Down\nsyndrome were more often incorrectly labelled as children; iii) social\nstereotypes are propagated in both the control and experimental groups, with\nlabels related to aesthetics more often associated with women, and labels\nrelated to education level and skills more often associated with men. These\nresults, although limited in scope, shed new light on the biases that alter\nface classification when applied to faces of people with Down syndrome. They\nconfirm the structural limitation of the technology, which is inherently\ndependent on the datasets used to train the models.\n", "date": "2025-02-10", "categories": []}
{"id": "2502.06439", "title": "Testing software for non-discrimination: an updated and extended audit\n  in the Italian car insurance domain", "abstract": "  Context. As software systems become more integrated into society's\ninfrastructure, the responsibility of software professionals to ensure\ncompliance with various non-functional requirements increases. These\nrequirements include security, safety, privacy, and, increasingly,\nnon-discrimination.\n  Motivation. Fairness in pricing algorithms grants equitable access to basic\nservices without discriminating on the basis of protected attributes.\n  Method. We replicate a previous empirical study that used black box testing\nto audit pricing algorithms used by Italian car insurance companies, accessible\nthrough a popular online system. With respect to the previous study, we\nenlarged the number of tests and the number of demographic variables under\nanalysis.\n  Results. Our work confirms and extends previous findings, highlighting the\nproblematic permanence of discrimination across time: demographic variables\nsignificantly impact pricing to this day, with birthplace remaining the main\ndiscriminatory factor against individuals not born in Italian cities. We also\nfound that driver profiles can determine the number of quotes available to the\nuser, denying equal opportunities to all.\n  Conclusion. The study underscores the importance of testing for\nnon-discrimination in software systems that affect people's everyday lives.\nPerforming algorithmic audits over time makes it possible to evaluate the\nevolution of such algorithms. It also demonstrates the role that empirical\nsoftware engineering can play in making software systems more accountable.\n", "date": "2025-02-10", "categories": []}
{"id": "2502.06695", "title": "FairDropout: Using Example-Tied Dropout to Enhance Generalization of\n  Minority Groups", "abstract": "  Deep learning models frequently exploit spurious features in training data to\nachieve low training error, often resulting in poor generalization when faced\nwith shifted testing distributions. To address this issue, various methods from\nimbalanced learning, representation learning, and classifier recalibration have\nbeen proposed to enhance the robustness of deep neural networks against\nspurious correlations. In this paper, we observe that models trained with\nempirical risk minimization tend to generalize well for examples from the\nmajority groups while memorizing instances from minority groups. Building on\nrecent findings that show memorization can be localized to a limited number of\nneurons, we apply example-tied dropout as a method we term FairDropout, aimed\nat redirecting this memorization to specific neurons that we subsequently drop\nout during inference. We empirically evaluate FairDropout using the\nsubpopulation benchmark suite encompassing vision, language, and healthcare\ntasks, demonstrating that it significantly reduces reliance on spurious\ncorrelations, and outperforms state-of-the-art methods.\n", "date": "2025-02-10", "categories": []}
{"id": "2502.04218", "title": "Sports and Women's Sports: Gender Bias in Text Generation with Olympic\n  Data", "abstract": "  Large Language Models (LLMs) have been shown to be biased in prior work, as\nthey generate text that is in line with stereotypical views of the world or\nthat is not representative of the viewpoints and values of historically\nmarginalized demographic groups. In this work, we propose using data from\nparallel men's and women's events at the Olympic Games to investigate different\nforms of gender bias in language models. We define three metrics to measure\nbias, and find that models are consistently biased against women when the\ngender is ambiguous in the prompt. In this case, the model frequently retrieves\nonly the results of the men's event with or without acknowledging them as such,\nrevealing pervasive gender bias in LLMs in the context of athletics.\n", "date": "2025-02-06", "categories": []}
{"id": "2502.04281", "title": "DECAF: Learning to be Fair in Multi-agent Resource Allocation", "abstract": "  A wide variety of resource allocation problems operate under resource\nconstraints that are managed by a central arbitrator, with agents who evaluate\nand communicate preferences over these resources. We formulate this broad class\nof problems as Distributed Evaluation, Centralized Allocation (DECA) problems\nand propose methods to learn fair and efficient policies in centralized\nresource allocation. Our methods are applied to learning long-term fairness in\na novel and general framework for fairness in multi-agent systems. We show\nthree different methods based on Double Deep Q-Learning: (1) A joint weighted\noptimization of fairness and utility, (2) a split optimization, learning two\nseparate Q-estimators for utility and fairness, and (3) an online policy\nperturbation to guide existing black-box utility functions toward fair\nsolutions. Our methods outperform existing fair MARL approaches on multiple\nresource allocation domains, even when evaluated using diverse fairness\nfunctions, and allow for flexible online trade-offs between utility and\nfairness.\n", "date": "2025-02-06", "categories": []}
{"id": "2502.04309", "title": "Targeted Learning for Data Fairness", "abstract": "  Data and algorithms have the potential to produce and perpetuate\ndiscrimination and disparate treatment. As such, significant effort has been\ninvested in developing approaches to defining, detecting, and eliminating\nunfair outcomes in algorithms. In this paper, we focus on performing\nstatistical inference for fairness. Prior work in fairness inference has\nlargely focused on inferring the fairness properties of a given predictive\nalgorithm. Here, we expand fairness inference by evaluating fairness in the\ndata generating process itself, referred to here as data fairness. We perform\ninference on data fairness using targeted learning, a flexible framework for\nnonparametric inference. We derive estimators demographic parity, equal\nopportunity, and conditional mutual information. Additionally, we find that our\nestimators for probabilistic metrics exploit double robustness. To validate our\napproach, we perform several simulations and apply our estimators to real data.\n", "date": "2025-02-06", "categories": []}
{"id": "2402.12062", "title": "Causal Equal Protection as Algorithmic Fairness", "abstract": "  By combining the philosophical literature on statistical evidence and the\ninterdisciplinary literature on algorithmic fairness, we revisit recent\nobjections against classification parity in light of causal analyses of\nalgorithmic fairness and the distinction between predictive and diagnostic\nevidence. We focus on trial proceedings as a black-box classification algorithm\nin which defendants are sorted into two groups by convicting or acquitting\nthem. We defend a novel principle, causal equal protection, that combines\nclassification parity with the causal approach. In the do-calculus, causal\nequal protection requires that individuals should not be subject to uneven\nrisks of classification error because of their protected or socially salient\ncharacteristics. The explicit use of protected characteristics, however, may be\nrequired if it equalizes these risks.\n", "date": "2024-02-19", "categories": []}
{"id": "2501.19407", "title": "Algorithmic Inheritance: Surname Bias in AI Decisions Reinforces\n  Intergenerational Inequality", "abstract": "  Surnames often convey implicit markers of social status, wealth, and lineage,\nshaping perceptions in ways that can perpetuate systemic biases and\nintergenerational inequality. This study is the first of its kind to\ninvestigate whether and how surnames influence AI-driven decision-making,\nfocusing on their effects across key areas such as hiring recommendations,\nleadership appointments, and loan approvals. Using 72,000 evaluations of 600\nsurnames from the United States and Thailand, two countries with distinct\nsociohistorical contexts and surname conventions, we classify names into four\ncategories: Rich, Legacy, Normal, and phonetically similar Variant groups. Our\nfindings show that elite surnames consistently increase AI-generated\nperceptions of power, intelligence, and wealth, which in turn influence\nAI-driven decisions in high-stakes contexts. Mediation analysis reveals\nperceived intelligence as a key mechanism through which surname biases\ninfluence AI decision-making process. While providing objective qualifications\nalongside surnames mitigates most of these biases, it does not eliminate them\nentirely, especially in contexts where candidate credentials are low. These\nfindings highlight the need for fairness-aware algorithms and robust policy\nmeasures to prevent AI systems from reinforcing systemic inequalities tied to\nsurnames, an often-overlooked bias compared to more salient characteristics\nsuch as race and gender. Our work calls for a critical reassessment of\nalgorithmic accountability and its broader societal impact, particularly in\nsystems designed to uphold meritocratic principles while counteracting the\nperpetuation of intergenerational privilege.\n", "date": "2025-01-23", "categories": []}
{"id": "2502.02623", "title": "Sample Complexity of Bias Detection with Subsampled Point-to-Subspace\n  Distances", "abstract": "  Sample complexity of bias estimation is a lower bound on the runtime of any\nbias detection method. Many regulatory frameworks require the bias to be tested\nfor all subgroups, whose number grows exponentially with the number of\nprotected attributes. Unless one wishes to run a bias detection with a\ndoubly-exponential run-time, one should like to have polynomial complexity of\nbias detection for a single subgroup. At the same time, the reference data may\nbe based on surveys, and thus come with non-trivial uncertainty.\n  Here, we reformulate bias detection as a point-to-subspace problem on the\nspace of measures and show that, for supremum norm, it can be subsampled\nefficiently. In particular, our probabilistically approximately correct (PAC)\nresults are corroborated by tests on well-known instances.\n", "date": "2025-02-04", "categories": []}
{"id": "2502.02966", "title": "FACTER: Fairness-Aware Conformal Thresholding and Prompt Engineering for\n  Enabling Fair LLM-Based Recommender Systems", "abstract": "  We propose FACTER, a fairness-aware framework for LLM-based recommendation\nsystems that integrates conformal prediction with dynamic prompt engineering.\nBy introducing an adaptive semantic variance threshold and a\nviolation-triggered mechanism, FACTER automatically tightens fairness\nconstraints whenever biased patterns emerge. We further develop an adversarial\nprompt generator that leverages historical violations to reduce repeated\ndemographic biases without retraining the LLM. Empirical results on MovieLens\nand Amazon show that FACTER substantially reduces fairness violations (up to\n95.5%) while maintaining strong recommendation accuracy, revealing semantic\nvariance as a potent proxy of bias.\n", "date": "2025-02-05", "categories": []}
{"id": "2502.03429", "title": "On Fairness of Unified Multimodal Large Language Model for Image\n  Generation", "abstract": "  Unified multimodal large language models (U-MLLMs) have demonstrated\nimpressive performance in visual understanding and generation in an end-to-end\npipeline. Compared with generation-only models (e.g., Stable Diffusion),\nU-MLLMs may raise new questions about bias in their outputs, which can be\naffected by their unified capabilities. This gap is particularly concerning\ngiven the under-explored risk of propagating harmful stereotypes. In this\npaper, we benchmark the latest U-MLLMs and find that most exhibit significant\ndemographic biases, such as gender and race bias. To better understand and\nmitigate this issue, we propose a locate-then-fix strategy, where we audit and\nshow how the individual model component is affected by bias. Our analysis shows\nthat bias originates primarily from the language model. More interestingly, we\nobserve a \"partial alignment\" phenomenon in U-MLLMs, where understanding bias\nappears minimal, but generation bias remains substantial. Thus, we propose a\nnovel balanced preference model to balance the demographic distribution with\nsynthetic data. Experiments demonstrate that our approach reduces demographic\nbias while preserving semantic fidelity. We hope our findings underscore the\nneed for more holistic interpretation and debiasing strategies of U-MLLMs in\nthe future.\n", "date": "2025-02-05", "categories": []}
{"id": "2502.02567", "title": "Fairness in Survival Analysis: A Novel Conditional Mutual Information\n  Augmentation Approach", "abstract": "  Survival analysis, a vital tool for predicting the time to event, has been\nused in many domains such as healthcare, criminal justice, and finance. Like\nclassification tasks, survival analysis can exhibit bias against disadvantaged\ngroups, often due to biases inherent in data or algorithms. Several studies in\nboth the IS and CS communities have attempted to address fairness in survival\nanalysis. However, existing methods often overlook the importance of prediction\nfairness at pre-defined evaluation time points, which is crucial in real-world\napplications where decision making often hinges on specific time frames. To\naddress this critical research gap, we introduce a new fairness concept:\nequalized odds (EO) in survival analysis, which emphasizes prediction fairness\nat pre-defined time points. To achieve the EO fairness in survival analysis, we\npropose a Conditional Mutual Information Augmentation (CMIA) approach, which\nfeatures a novel fairness regularization term based on conditional mutual\ninformation and an innovative censored data augmentation technique. Our CMIA\napproach can effectively balance prediction accuracy and fairness, and it is\napplicable to various survival models. We evaluate the CMIA approach against\nseveral state-of-the-art methods within three different application domains,\nand the results demonstrate that CMIA consistently reduces prediction disparity\nwhile maintaining good accuracy and significantly outperforms the other\ncompeting methods across multiple datasets and survival models (e.g., linear\nCOX, deep AFT).\n", "date": "2025-02-04", "categories": []}
{"id": "2502.00690", "title": "Dissecting Submission Limit in Desk-Rejections: A Mathematical Analysis\n  of Fairness in AI Conference Policies", "abstract": "  As AI research surges in both impact and volume, conferences have imposed\nsubmission limits to maintain paper quality and alleviate organizational\npressure. In this work, we examine the fairness of desk-rejection systems under\nsubmission limits and reveal that existing practices can result in substantial\ninequities. Specifically, we formally define the paper submission limit problem\nand identify a critical dilemma: when the number of authors exceeds three, it\nbecomes impossible to reject papers solely based on excessive submissions\nwithout negatively impacting innocent authors. Thus, this issue may unfairly\naffect early-career researchers, as their submissions may be penalized due to\nco-authors with significantly higher submission counts, while senior\nresearchers with numerous papers face minimal consequences. To address this, we\npropose an optimization-based fairness-aware desk-rejection mechanism and\nformally define two fairness metrics: individual fairness and group fairness.\nWe prove that optimizing individual fairness is NP-hard, whereas group fairness\ncan be efficiently optimized via linear programming. Through case studies, we\ndemonstrate that our proposed system ensures greater equity than existing\nmethods, including those used in CVPR 2025, offering a more socially just\napproach to managing excessive submissions in AI conferences.\n", "date": "2025-02-02", "categories": []}
{"id": "2502.01188", "title": "FairUDT: Fairness-aware Uplift Decision Trees", "abstract": "  Training data used for developing machine learning classifiers can exhibit\nbiases against specific protected attributes. Such biases typically originate\nfrom historical discrimination or certain underlying patterns that\ndisproportionately under-represent minority groups, such as those identified by\ntheir gender, religion, or race. In this paper, we propose a novel approach,\nFairUDT, a fairness-aware Uplift-based Decision Tree for discrimination\nidentification. FairUDT demonstrates how the integration of uplift modeling\nwith decision trees can be adapted to include fair splitting criteria.\nAdditionally, we introduce a modified leaf relabeling approach for removing\ndiscrimination. We divide our dataset into favored and deprived groups based on\na binary sensitive attribute, with the favored dataset serving as the treatment\ngroup and the deprived dataset as the control group. By applying FairUDT and\nour leaf relabeling approach to preprocess three benchmark datasets, we achieve\nan acceptable accuracy-discrimination tradeoff. We also show that FairUDT is\ninherently interpretable and can be utilized in discrimination detection tasks.\nThe code for this project is available https://github.com/ara-25/FairUDT\n", "date": "2025-02-03", "categories": []}
{"id": "2502.01211", "title": "Privilege Scores", "abstract": "  Bias-transforming methods of fairness-aware machine learning aim to correct a\nnon-neutral status quo with respect to a protected attribute (PA). Current\nmethods, however, lack an explicit formulation of what drives non-neutrality.\nWe introduce privilege scores (PS) to measure PA-related privilege by comparing\nthe model predictions in the real world with those in a fair world in which the\ninfluence of the PA is removed. At the individual level, PS can identify\nindividuals who qualify for affirmative action; at the global level, PS can\ninform bias-transforming policies. After presenting estimation methods for PS,\nwe propose privilege score contributions (PSCs), an interpretation method that\nattributes the origin of privilege to mediating features and direct effects. We\nprovide confidence intervals for both PS and PSCs. Experiments on simulated and\nreal-world data demonstrate the broad applicability of our methods and provide\nnovel insights into gender and racial privilege in mortgage and college\nadmissions applications.\n", "date": "2025-02-03", "categories": []}
{"id": "2501.10150", "title": "Dual Debiasing: Remove Stereotypes and Keep Factual Gender for Fair\n  Language Modeling and Translation", "abstract": "  Mitigation of biases, such as language models' reliance on gender\nstereotypes, is a crucial endeavor required for the creation of reliable and\nuseful language technology. The crucial aspect of debiasing is to ensure that\nthe models preserve their versatile capabilities, including their ability to\nsolve language tasks and equitably represent various genders. To address this\nissue, we introduce a streamlined Dual Dabiasing Algorithm through Model\nAdaptation (2DAMA). Novel Dual Debiasing enables robust reduction of\nstereotypical bias while preserving desired factual gender information encoded\nby language models. We show that 2DAMA effectively reduces gender bias in\nEnglish and is one of the first approaches facilitating the mitigation of\nstereotypical tendencies in translation. The proposed method's key advantage is\nthe preservation of factual gender cues, which are useful in a wide range of\nnatural language processing tasks.\n", "date": "2025-01-17", "categories": []}
{"id": "2501.18642", "title": "DebiasPI: Inference-time Debiasing by Prompt Iteration of a\n  Text-to-Image Generative Model", "abstract": "  Ethical intervention prompting has emerged as a tool to counter demographic\nbiases of text-to-image generative AI models. Existing solutions either require\nto retrain the model or struggle to generate images that reflect desired\ndistributions on gender and race. We propose an inference-time process called\nDebiasPI for Debiasing-by-Prompt-Iteration that provides prompt intervention by\nenabling the user to control the distributions of individuals' demographic\nattributes in image generation. DebiasPI keeps track of which attributes have\nbeen generated either by probing the internal state of the model or by using\nexternal attribute classifiers. Its control loop guides the text-to-image model\nto select not yet sufficiently represented attributes, With DebiasPI, we were\nable to create images with equal representations of race and gender that\nvisualize challenging concepts of news headlines. We also experimented with the\nattributes age, body type, profession, and skin tone, and measured how\nattributes change when our intervention prompt targets the distribution of an\nunrelated attribute type. We found, for example, if the text-to-image model is\nasked to balance racial representation, gender representation improves but the\nskin tone becomes less diverse. Attempts to cover a wide range of skin colors\nwith various intervention prompts showed that the model struggles to generate\nthe palest skin tones. We conducted various ablation studies, in which we\nremoved DebiasPI's attribute control, that reveal the model's propensity to\ngenerate young, male characters. It sometimes visualized career success by\ngenerating two-panel images with a pre-success dark-skinned person becoming\nlight-skinned with success, or switching gender from pre-success female to\npost-success male, thus further motivating ethical intervention prompting with\nDebiasPI.\n", "date": "2025-01-28", "categories": []}
{"id": "2409.11149", "title": "SAGED: A Holistic Bias-Benchmarking Pipeline for Language Models with\n  Customisable Fairness Calibration", "abstract": "  The development of unbiased large language models is widely recognized as\ncrucial, yet existing benchmarks fall short in detecting biases due to limited\nscope, contamination, and lack of a fairness baseline. SAGED(bias) is the first\nholistic benchmarking pipeline to address these problems. The pipeline\nencompasses five core stages: scraping materials, assembling benchmarks,\ngenerating responses, extracting numeric features, and diagnosing with\ndisparity metrics. SAGED includes metrics for max disparity, such as impact\nratio, and bias concentration, such as Max Z-scores. Noticing that metric tool\nbias and contextual bias in prompts can distort evaluation, SAGED implements\ncounterfactual branching and baseline calibration for mitigation. For\ndemonstration, we use SAGED on G20 Countries with popular 8b-level models\nincluding Gemma2, Llama3.1, Mistral, and Qwen2. With sentiment analysis, we\nfind that while Mistral and Qwen2 show lower max disparity and higher bias\nconcentration than Gemma2 and Llama3.1, all models are notably biased against\ncountries like Russia and (except for Qwen2) China. With further experiments to\nhave models role-playing U.S. presidents, we see bias amplifies and shifts in\nheterogeneous directions. Moreover, we see Qwen2 and Mistral not engage in\nrole-playing, while Llama3.1 and Gemma2 role-play Trump notably more\nintensively than Biden and Harris, indicating role-playing performance bias in\nthese models.\n", "date": "2024-09-17", "categories": []}
{"id": "2501.18192", "title": "Machine Learning Fairness for Depression Detection using EEG Data", "abstract": "  This paper presents the very first attempt to evaluate machine learning\nfairness for depression detection using electroencephalogram (EEG) data. We\nconduct experiments using different deep learning architectures such as\nConvolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks,\nand Gated Recurrent Unit (GRU) networks across three EEG datasets: Mumtaz,\nMODMA and Rest. We employ five different bias mitigation strategies at the\npre-, in- and post-processing stages and evaluate their effectiveness. Our\nexperimental results show that bias exists in existing EEG datasets and\nalgorithms for depression detection, and different bias mitigation methods\naddress bias at different levels across different fairness measures.\n", "date": "2025-01-30", "categories": []}
{"id": "2501.18277", "title": "Sebra: Debiasing Through Self-Guided Bias Ranking", "abstract": "  Ranking samples by fine-grained estimates of spuriosity (the degree to which\nspurious cues are present) has recently been shown to significantly benefit\nbias mitigation, over the traditional binary biased-\\textit{vs}-unbiased\npartitioning of train sets. However, this spuriosity ranking comes with the\nrequirement of human supervision. In this paper, we propose a debiasing\nframework based on our novel \\ul{Se}lf-Guided \\ul{B}ias \\ul{Ra}nking\n(\\emph{Sebra}), that mitigates biases (spurious correlations) via an automatic\nranking of data points by spuriosity within their respective classes. Sebra\nleverages a key local symmetry in Empirical Risk Minimization (ERM) training --\nthe ease of learning a sample via ERM inversely correlates with its\nspuriousity; the fewer spurious correlations a sample exhibits, the harder it\nis to learn, and vice versa. However, globally across iterations, ERM tends to\ndeviate from this symmetry. Sebra dynamically steers ERM to correct this\ndeviation, facilitating the sequential learning of attributes in increasing\norder of difficulty, \\ie, decreasing order of spuriosity. As a result, the\nsequence in which Sebra learns samples naturally provides spuriousity rankings.\nWe use the resulting fine-grained bias characterization in a contrastive\nlearning framework to mitigate biases from multiple sources. Extensive\nexperiments show that Sebra consistently outperforms previous state-of-the-art\nunsupervised debiasing techniques across multiple standard benchmarks,\nincluding UrbanCars, BAR, CelebA, and ImageNet-1K. Code, pre-trained models,\nand training logs are available at https://kadarsh22.github.io/sebra_iclr25/.\n", "date": "2025-01-30", "categories": []}
{"id": "2501.18331", "title": "Stream-Based Monitoring of Algorithmic Fairness", "abstract": "  Automatic decision and prediction systems are increasingly deployed in\napplications where they significantly impact the livelihood of people, such as\nfor predicting the creditworthiness of loan applicants or the recidivism risk\nof defendants. These applications have given rise to a new class of\nalgorithmic-fairness specifications that require the systems to decide and\npredict without bias against social groups. Verifying these specifications\nstatically is often out of reach for realistic systems, since the systems may,\ne.g., employ complex learning components, and reason over a large input space.\nIn this paper, we therefore propose stream-based monitoring as a solution for\nverifying the algorithmic fairness of decision and prediction systems at\nruntime. Concretely, we present a principled way to formalize algorithmic\nfairness over temporal data streams in the specification language RTLola and\ndemonstrate the efficacy of this approach on a number of benchmarks. Besides\nsynthetic scenarios that particularly highlight its efficiency on streams with\na scaling amount of data, we notably evaluate the monitor on real-world data\nfrom the recidivism prediction tool COMPAS.\n", "date": "2025-01-30", "categories": []}
{"id": "2410.04472", "title": "Collapsed Language Models Promote Fairness", "abstract": "  To mitigate societal biases implicitly encoded in recent successful\npretrained language models, a diverse array of approaches have been proposed to\nencourage model fairness, focusing on prompting, data augmentation, regularized\nfine-tuning, and more. Despite the development, it is nontrivial to reach a\nprincipled understanding of fairness and an effective algorithm that can\nconsistently debias language models. In this work, by rigorous evaluations of\nNeural Collapse -- a learning phenomenon happen in last-layer representations\nand classifiers in deep networks -- on fairness-related words, we find that\ndebiased language models exhibit collapsed alignment between token\nrepresentations and word embeddings. More importantly, this observation\ninspires us to design a principled fine-tuning method that can effectively\nimprove fairness in a wide range of debiasing methods, while still preserving\nthe performance of language models on standard natural language understanding\ntasks. We attach our code at https://github.com/Xujxyang/Fairness-NC-main.\n", "date": "2024-10-06", "categories": []}
{"id": "2501.03112", "title": "LangFair: A Python Package for Assessing Bias and Fairness in Large\n  Language Model Use Cases", "abstract": "  Large Language Models (LLMs) have been observed to exhibit bias in numerous\nways, potentially creating or worsening outcomes for specific groups identified\nby protected attributes such as sex, race, sexual orientation, or age. To help\naddress this gap, we introduce LangFair, an open-source Python package that\naims to equip LLM practitioners with the tools to evaluate bias and fairness\nrisks relevant to their specific use cases. The package offers functionality to\neasily generate evaluation datasets, comprised of LLM responses to\nuse-case-specific prompts, and subsequently calculate applicable metrics for\nthe practitioner's use case. To guide in metric selection, LangFair offers an\nactionable decision framework.\n", "date": "2025-01-06", "categories": []}
{"id": "2501.17420", "title": "Actions Speak Louder than Words: Agent Decisions Reveal Implicit Biases\n  in Language Models", "abstract": "  While advances in fairness and alignment have helped mitigate overt biases\nexhibited by large language models (LLMs) when explicitly prompted, we\nhypothesize that these models may still exhibit implicit biases when simulating\nhuman behavior. To test this hypothesis, we propose a technique to\nsystematically uncover such biases across a broad range of sociodemographic\ncategories by assessing decision-making disparities among agents with\nLLM-generated, sociodemographically-informed personas. Using our technique, we\ntested six LLMs across three sociodemographic groups and four decision-making\nscenarios. Our results show that state-of-the-art LLMs exhibit significant\nsociodemographic disparities in nearly all simulations, with more advanced\nmodels exhibiting greater implicit biases despite reducing explicit biases.\nFurthermore, when comparing our findings to real-world disparities reported in\nempirical studies, we find that the biases we uncovered are directionally\naligned but markedly amplified. This directional alignment highlights the\nutility of our technique in uncovering systematic biases in LLMs rather than\nrandom variations; moreover, the presence and amplification of implicit biases\nemphasizes the need for novel strategies to address these biases.\n", "date": "2025-01-29", "categories": []}
{"id": "2401.12824", "title": "MAPPING: Debiasing Graph Neural Networks for Fair Node Classification\n  with Limited Sensitive Information Leakage", "abstract": "  Despite remarkable success in diverse web-based applications, Graph Neural\nNetworks(GNNs) inherit and further exacerbate historical discrimination and\nsocial stereotypes, which critically hinder their deployments in high-stake\ndomains such as online clinical diagnosis, financial crediting, etc. However,\ncurrent fairness research that primarily craft on i.i.d data, cannot be\ntrivially replicated to non-i.i.d. graph structures with topological dependence\namong samples. Existing fair graph learning typically favors pairwise\nconstraints to achieve fairness but fails to cast off dimensional limitations\nand generalize them into multiple sensitive attributes; besides, most studies\nfocus on in-processing techniques to enforce and calibrate fairness,\nconstructing a model-agnostic debiasing GNN framework at the pre-processing\nstage to prevent downstream misuses and improve training reliability is still\nlargely under-explored. Furthermore, previous work on GNNs tend to enhance\neither fairness or privacy individually but few probe into their interplays. In\nthis paper, we propose a novel model-agnostic debiasing framework named MAPPING\n(\\underline{M}asking \\underline{A}nd \\underline{P}runing and\nMessage-\\underline{P}assing train\\underline{ING}) for fair node classification,\nin which we adopt the distance covariance($dCov$)-based fairness constraints to\nsimultaneously reduce feature and topology biases in arbitrary dimensions, and\ncombine them with adversarial debiasing to confine the risks of attribute\ninference attacks. Experiments on real-world datasets with different GNN\nvariants demonstrate the effectiveness and flexibility of MAPPING. Our results\nshow that MAPPING can achieve better trade-offs between utility and fairness,\nand mitigate privacy risks of sensitive information leakage.\n", "date": "2024-01-23", "categories": []}
{"id": "2501.15318", "title": "A Post-Processing-Based Fair Federated Learning Framework", "abstract": "  Federated Learning (FL) allows collaborative model training among distributed\nparties without pooling local datasets at a central server. However, the\ndistributed nature of FL poses challenges in training fair federated learning\nmodels. The existing techniques are often limited in offering fairness\nflexibility to clients and performance. We formally define and empirically\nanalyze a simple and intuitive post-processing-based framework to improve group\nfairness in FL systems. This framework can be divided into two stages: a\nstandard FL training stage followed by a completely decentralized local\ndebiasing stage. In the first stage, a global model is trained without fairness\nconstraints using a standard federated learning algorithm (e.g. FedAvg). In the\nsecond stage, each client applies fairness post-processing on the global model\nusing their respective local dataset. This allows for customized fairness\nimprovements based on clients' desired and context-guided fairness\nrequirements. We demonstrate two well-established post-processing techniques in\nthis framework: model output post-processing and final layer fine-tuning. We\nevaluate the framework against three common baselines on four different\ndatasets, including tabular, signal, and image data, each with varying levels\nof data heterogeneity across clients. Our work shows that this framework not\nonly simplifies fairness implementation in FL but also provides significant\nfairness improvements with minimal accuracy loss or even accuracy gain, across\ndata modalities and machine learning methods, being especially effective in\nmore heterogeneous settings.\n", "date": "2025-01-25", "categories": []}
{"id": "2501.15430", "title": "Evaluating Simple Debiasing Techniques in RoBERTa-based Hate Speech\n  Detection Models", "abstract": "  The hate speech detection task is known to suffer from bias against African\nAmerican English (AAE) dialect text, due to the annotation bias present in the\nunderlying hate speech datasets used to train these models. This leads to a\ndisparity where normal AAE text is more likely to be misclassified as\nabusive/hateful compared to non-AAE text. Simple debiasing techniques have been\ndeveloped in the past to counter this sort of disparity, and in this work, we\napply and evaluate these techniques in the scope of RoBERTa-based encoders.\nExperimental results suggest that the success of these techniques depends\nheavily on the methods used for training dataset construction, but with proper\nconsideration of representation bias, they can reduce the disparity seen among\ndialect subgroups on the hate speech detection task.\n", "date": "2025-01-26", "categories": []}
{"id": "2501.15634", "title": "Be Intentional About Fairness!: Fairness, Size, and Multiplicity in the\n  Rashomon Set", "abstract": "  When selecting a model from a set of equally performant models, how much\nunfairness can you really reduce? Is it important to be intentional about\nfairness when choosing among this set, or is arbitrarily choosing among the set\nof ''good'' models good enough? Recent work has highlighted that the phenomenon\nof model multiplicity-where multiple models with nearly identical predictive\naccuracy exist for the same task-has both positive and negative implications\nfor fairness, from strengthening the enforcement of civil rights law in AI\nsystems to showcasing arbitrariness in AI decision-making. Despite the enormous\nimplications of model multiplicity, there is little work that explores the\nproperties of sets of equally accurate models, or Rashomon sets, in general. In\nthis paper, we present five main theoretical and methodological contributions\nwhich help us to understand the relatively unexplored properties of the\nRashomon set, in particular with regards to fairness. Our contributions include\nmethods for efficiently sampling models from this set and techniques for\nidentifying the fairest models according to key fairness metrics such as\nstatistical parity. We also derive the probability that an individual's\nprediction will be flipped within the Rashomon set, as well as expressions for\nthe set's size and the distribution of error tolerance used across models.\nThese results lead to policy-relevant takeaways, such as the importance of\nintentionally looking for fair models within the Rashomon set, and\nunderstanding which individuals or groups may be more susceptible to arbitrary\ndecisions.\n", "date": "2025-01-26", "categories": []}
{"id": "2402.14875", "title": "What's in a Name? Auditing Large Language Models for Race and Gender\n  Bias", "abstract": "  We employ an audit design to investigate biases in state-of-the-art large\nlanguage models, including GPT-4. In our study, we prompt the models for advice\ninvolving a named individual across a variety of scenarios, such as during car\npurchase negotiations or election outcome predictions. We find that the advice\nsystematically disadvantages names that are commonly associated with racial\nminorities and women. Names associated with Black women receive the least\nadvantageous outcomes. The biases are consistent across 42 prompt templates and\nseveral models, indicating a systemic issue rather than isolated incidents.\nWhile providing numerical, decision-relevant anchors in the prompt can\nsuccessfully counteract the biases, qualitative details have inconsistent\neffects and may even increase disparities. Our findings underscore the\nimportance of conducting audits at the point of LLM deployment and\nimplementation to mitigate their potential for harm against marginalized\ncommunities.\n", "date": "2024-02-21", "categories": []}
{"id": "2501.14414", "title": "SoK: What Makes Private Learning Unfair?", "abstract": "  Differential privacy has emerged as the most studied framework for\nprivacy-preserving machine learning. However, recent studies show that\nenforcing differential privacy guarantees can not only significantly degrade\nthe utility of the model, but also amplify existing disparities in its\npredictive performance across demographic groups. Although there is extensive\nresearch on the identification of factors that contribute to this phenomenon,\nwe still lack a complete understanding of the mechanisms through which\ndifferential privacy exacerbates disparities. The literature on this problem is\nmuddled by varying definitions of fairness, differential privacy mechanisms,\nand inconsistent experimental settings, often leading to seemingly\ncontradictory results.\n  This survey provides the first comprehensive overview of the factors that\ncontribute to the disparate effect of training models with differential privacy\nguarantees. We discuss their impact and analyze their causal role in such a\ndisparate effect. Our analysis is guided by a taxonomy that categorizes these\nfactors by their position within the machine learning pipeline, allowing us to\ndraw conclusions about their interaction and the feasibility of potential\nmitigation strategies. We find that factors related to the training dataset and\nthe underlying distribution play a decisive role in the occurrence of disparate\nimpact, highlighting the need for research on these factors to address the\nissue.\n", "date": "2025-01-24", "categories": []}
{"id": "2501.14551", "title": "Fairness of Deep Ensembles: On the interplay between per-group task\n  difficulty and under-representation", "abstract": "  Ensembling is commonly regarded as an effective way to improve the general\nperformance of models in machine learning, while also increasing the robustness\nof predictions. When it comes to algorithmic fairness, heterogeneous ensembles,\ncomposed of multiple model types, have been employed to mitigate biases in\nterms of demographic attributes such as sex, age or ethnicity. Moreover, recent\nwork has shown how in multi-class problems even simple homogeneous ensembles\nmay favor performance of the worst-performing target classes. While homogeneous\nensembles are simpler to implement in practice, it is not yet clear whether\ntheir benefits translate to groups defined not in terms of their target class,\nbut in terms of demographic or protected attributes, hence improving fairness.\nIn this work we show how this simple and straightforward method is indeed able\nto mitigate disparities, particularly benefiting under-performing subgroups.\nInterestingly, this can be achieved without sacrificing overall performance,\nwhich is a common trade-off observed in bias mitigation strategies. Moreover,\nwe analyzed the interplay between two factors which may result in biases:\nsub-group under-representation and the inherent difficulty of the task for each\ngroup. These results revealed that, contrary to popular assumptions, having\nbalanced datasets may be suboptimal if the task difficulty varies between\nsubgroups. Indeed, we found that a perfectly balanced dataset may hurt both the\noverall performance and the gap between groups. This highlights the importance\nof considering the interaction between multiple forces at play in fairness.\n", "date": "2025-01-24", "categories": []}
{"id": "2501.10784", "title": "Measuring Fairness in Financial Transaction Machine Learning Models", "abstract": "  Mastercard, a global leader in financial services, develops and deploys\nmachine learning models aimed at optimizing card usage and preventing attrition\nthrough advanced predictive models. These models use aggregated and anonymized\ncard usage patterns, including cross-border transactions and industry-specific\nspending, to tailor bank offerings and maximize revenue opportunities.\nMastercard has established an AI Governance program, based on its Data and Tech\nResponsibility Principles, to evaluate any built and bought AI for efficacy,\nfairness, and transparency. As part of this effort, Mastercard has sought\nexpertise from the Turing Institute through a Data Study Group to better assess\nfairness in more complex AI/ML models. The Data Study Group challenge lies in\ndefining, measuring, and mitigating fairness in these predictions, which can be\ncomplex due to the various interpretations of fairness, gaps in the research\nliterature, and ML-operations challenges.\n", "date": "2025-01-18", "categories": []}
{"id": "2501.13219", "title": "Enhancing Multi-Attribute Fairness in Healthcare Predictive Modeling", "abstract": "  Artificial intelligence (AI) systems in healthcare have demonstrated\nremarkable potential to improve patient outcomes. However, if not designed with\nfairness in mind, they also carry the risks of perpetuating or exacerbating\nexisting health disparities. Although numerous fairness-enhancing techniques\nhave been proposed, most focus on a single sensitive attribute and neglect the\nbroader impact that optimizing fairness for one attribute may have on the\nfairness of other sensitive attributes. In this work, we introduce a novel\napproach to multi-attribute fairness optimization in healthcare AI, tackling\nfairness concerns across multiple demographic attributes concurrently. Our\nmethod follows a two-phase approach: initially optimizing for predictive\nperformance, followed by fine-tuning to achieve fairness across multiple\nsensitive attributes. We develop our proposed method using two strategies,\nsequential and simultaneous. Our results show a significant reduction in\nEqualized Odds Disparity (EOD) for multiple attributes, while maintaining high\npredictive accuracy. Notably, we demonstrate that single-attribute fairness\nmethods can inadvertently increase disparities in non-targeted attributes\nwhereas simultaneous multi-attribute optimization achieves more balanced\nfairness improvements across all attributes. These findings highlight the\nimportance of comprehensive fairness strategies in healthcare AI and offer\npromising directions for future research in this critical area.\n", "date": "2025-01-22", "categories": []}
{"id": "2501.13302", "title": "Watching the AI Watchdogs: A Fairness and Robustness Analysis of AI\n  Safety Moderation Classifiers", "abstract": "  AI Safety Moderation (ASM) classifiers are designed to moderate content on\nsocial media platforms and to serve as guardrails that prevent Large Language\nModels (LLMs) from being fine-tuned on unsafe inputs. Owing to their potential\nfor disparate impact, it is crucial to ensure that these classifiers: (1) do\nnot unfairly classify content belonging to users from minority groups as unsafe\ncompared to those from majority groups and (2) that their behavior remains\nrobust and consistent across similar inputs. In this work, we thus examine the\nfairness and robustness of four widely-used, closed-source ASM classifiers:\nOpenAI Moderation API, Perspective API, Google Cloud Natural Language (GCNL)\nAPI, and Clarifai API. We assess fairness using metrics such as demographic\nparity and conditional statistical parity, comparing their performance against\nASM models and a fair-only baseline. Additionally, we analyze robustness by\ntesting the classifiers' sensitivity to small and natural input perturbations.\nOur findings reveal potential fairness and robustness gaps, highlighting the\nneed to mitigate these issues in future versions of these models.\n", "date": "2025-01-23", "categories": []}
{"id": "2501.13421", "title": "Perceived Fairness of the Machine Learning Development Process: Concept\n  Scale Development", "abstract": "  In machine learning (ML) applications, unfairness is triggered due to bias in\nthe data, the data curation process, erroneous assumptions, and implicit bias\nrendered during the development process. It is also well-accepted by\nresearchers that fairness in ML application development is highly subjective,\nwith a lack of clarity of what it means from an ML development and\nimplementation perspective. Thus, in this research, we investigate and\nformalize the notion of the perceived fairness of ML development from a\nsociotechnical lens. Our goal in this research is to understand the\ncharacteristics of perceived fairness in ML applications. We address this\nresearch goal using a three-pronged strategy: 1) conducting virtual focus\ngroups with ML developers, 2) reviewing existing literature on fairness in ML,\nand 3) incorporating aspects of justice theory relating to procedural and\ndistributive justice. Based on our theoretical exposition, we propose\noperational attributes of perceived fairness to be transparency,\naccountability, and representativeness. These are described in terms of\nmultiple concepts that comprise each dimension of perceived fairness. We use\nthis operationalization to empirically validate the notion of perceived\nfairness of machine learning (ML) applications from both the ML practioners and\nusers perspectives. The multidimensional framework for perceived fairness\noffers a comprehensive understanding of perceived fairness, which can guide the\ncreation of fair ML systems with positive implications for society and\nbusinesses.\n", "date": "2025-01-23", "categories": []}
{"id": "2409.01977", "title": "Counterfactual Fairness by Combining Factual and Counterfactual\n  Predictions", "abstract": "  In high-stake domains such as healthcare and hiring, the role of machine\nlearning (ML) in decision-making raises significant fairness concerns. This\nwork focuses on Counterfactual Fairness (CF), which posits that an ML model's\noutcome on any individual should remain unchanged if they had belonged to a\ndifferent demographic group. Previous works have proposed methods that\nguarantee CF. Notwithstanding, their effects on the model's predictive\nperformance remains largely unclear. To fill in this gap, we provide a\ntheoretical study on the inherent trade-off between CF and predictive\nperformance in a model-agnostic manner. We first propose a simple but effective\nmethod to cast an optimal but potentially unfair predictor into a fair one\nwithout losing the optimality. By analyzing its excess risk in order to achieve\nCF, we quantify this inherent trade-off. Further analysis on our method's\nperformance with access to only incomplete causal knowledge is also conducted.\nBuilt upon it, we propose a performant algorithm that can be applied in such\nscenarios. Experiments on both synthetic and semi-synthetic datasets\ndemonstrate the validity of our analysis and methods.\n", "date": "2024-09-03", "categories": []}
{"id": "2401.13213", "title": "Common-Sense Bias Modeling for Classification Tasks", "abstract": "  Machine learning model bias can arise from dataset composition: correlated\nsensitive features can distort the downstream classification model's decision\nboundary and lead to performance differences along these features. Existing\nde-biasing works tackle the most prominent bias features, such as colors of\ndigits or background of animals. However, real-world datasets often include a\nlarge number of feature correlations that intrinsically manifest in the data as\ncommon sense information. Such spurious visual cues can further reduce model\nrobustness. Thus, domain practitioners desire a comprehensive understanding of\ncorrelations and the flexibility to address relevant biases. To this end, we\npropose a novel framework to extract comprehensive biases in image datasets\nbased on textual descriptions, a common sense-rich modality. Specifically,\nfeatures are constructed by clustering noun phrase embeddings with similar\nsemantics. The presence of each feature across the dataset is inferred, and\ntheir co-occurrence statistics are measured, with spurious correlations\noptionally examined by a human-in-the-loop module. Downstream experiments show\nthat our method uncovers novel model biases in multiple image benchmark\ndatasets. Furthermore, the discovered bias can be mitigated by simple data\nre-weighting to de-correlate the features, outperforming state-of-the-art\nunsupervised bias mitigation methods.\n", "date": "2024-01-24", "categories": []}
{"id": "2405.06687", "title": "Hire Me or Not? Examining Language Model's Behavior with Occupation\n  Attributes", "abstract": "  With the impressive performance in various downstream tasks, large language\nmodels (LLMs) have been widely integrated into production pipelines, like\nrecruitment and recommendation systems. A known issue of models trained on\nnatural language data is the presence of human biases, which can impact the\nfairness of the system. This paper investigates LLMs' behavior with respect to\ngender stereotypes, in the context of occupation decision making. Our framework\nis designed to investigate and quantify the presence of gender stereotypes in\nLLMs' behavior via multi-round question answering. Inspired by prior works, we\nconstruct a dataset by leveraging a standard occupation classification\nknowledge base released by authoritative agencies. We tested three LLMs\n(RoBERTa-large, GPT-3.5-turbo, and Llama2-70b-chat) and found that all models\nexhibit gender stereotypes analogous to human biases, but with different\npreferences. The distinct preferences of GPT-3.5-turbo and Llama2-70b-chat may\nimply the current alignment methods are insufficient for debiasing and could\nintroduce new biases contradicting the traditional gender stereotypes.\n", "date": "2024-05-06", "categories": []}
{"id": "2408.08471", "title": "Fairness Issues and Mitigations in (Differentially Private)\n  Socio-Demographic Data Processes", "abstract": "  Statistical agencies rely on sampling techniques to collect socio-demographic\ndata crucial for policy-making and resource allocation. This paper shows that\nsurveys of important societal relevance introduce sampling errors that unevenly\nimpact group-level estimates, thereby compromising fairness in downstream\ndecisions. To address these issues, this paper introduces an optimization\napproach modeled on real-world survey design processes, ensuring sampling costs\nare optimized while maintaining error margins within prescribed tolerances.\nAdditionally, privacy-preserving methods used to determine sampling rates can\nfurther impact these fairness issues. This paper explores the impact of\ndifferential privacy on the statistics informing the sampling process,\nrevealing a surprising effect: not only is the expected negative effect from\nthe addition of noise for differential privacy negligible, but also this\nprivacy noise can in fact reduce unfairness as it positively biases smaller\ncounts. These findings are validated over an extensive analysis using datasets\ncommonly applied in census statistics.\n", "date": "2024-08-16", "categories": []}
{"id": "2412.17523", "title": "Constructing Fair Latent Space for Intersection of Fairness and\n  Explainability", "abstract": "  As the use of machine learning models has increased, numerous studies have\naimed to enhance fairness. However, research on the intersection of fairness\nand explainability remains insufficient, leading to potential issues in gaining\nthe trust of actual users. Here, we propose a novel module that constructs a\nfair latent space, enabling faithful explanation while ensuring fairness. The\nfair latent space is constructed by disentangling and redistributing labels and\nsensitive attributes, allowing the generation of counterfactual explanations\nfor each type of information. Our module is attached to a pretrained generative\nmodel, transforming its biased latent space into a fair latent space.\nAdditionally, since only the module needs to be trained, there are advantages\nin terms of time and cost savings, without the need to train the entire\ngenerative model. We validate the fair latent space with various fairness\nmetrics and demonstrate that our approach can effectively provide explanations\nfor biased decisions and assurances of fairness.\n", "date": "2024-12-23", "categories": []}
{"id": "2501.10371", "title": "What we learned while automating bias detection in AI hiring systems for\n  compliance with NYC Local Law 144", "abstract": "  Since July 5, 2023, New York City's Local Law 144 requires employers to\nconduct independent bias audits for any automated employment decision tools\n(AEDTs) used in hiring processes. The law outlines a minimum set of bias tests\nthat AI developers and implementers must perform to ensure compliance. Over the\npast few months, we have collected and analyzed audits conducted under this\nlaw, identified best practices, and developed a software tool to streamline\nemployer compliance. Our tool, ITACA_144, tailors our broader bias auditing\nframework to meet the specific requirements of Local Law 144. While automating\nthese legal mandates, we identified several critical challenges that merit\nattention to ensure AI bias regulations and audit methodologies are both\neffective and practical. This document presents the insights gained from\nautomating compliance with NYC Local Law 144. It aims to support other cities\nand states in crafting similar legislation while addressing the limitations of\nthe NYC framework. The discussion focuses on key areas including data\nrequirements, demographic inclusiveness, impact ratios, effective bias,\nmetrics, and data reliability.\n", "date": "2024-12-13", "categories": []}
{"id": "2501.10453", "title": "Uncovering Bias in Foundation Models: Impact, Testing, Harm, and\n  Mitigation", "abstract": "  Bias in Foundation Models (FMs) - trained on vast datasets spanning societal\nand historical knowledge - poses significant challenges for fairness and equity\nacross fields such as healthcare, education, and finance. These biases, rooted\nin the overrepresentation of stereotypes and societal inequalities in training\ndata, exacerbate real-world discrimination, reinforce harmful stereotypes, and\nerode trust in AI systems. To address this, we introduce Trident Probe Testing\n(TriProTesting), a systematic testing method that detects explicit and implicit\nbiases using semantically designed probes. Here we show that FMs, including\nCLIP, ALIGN, BridgeTower, and OWLv2, demonstrate pervasive biases across single\nand mixed social attributes (gender, race, age, and occupation). Notably, we\nuncover mixed biases when social attributes are combined, such as gender x\nrace, gender x age, and gender x occupation, revealing deeper layers of\ndiscrimination. We further propose Adaptive Logit Adjustment\n(AdaLogAdjustment), a post-processing technique that dynamically redistributes\nprobability power to mitigate these biases effectively, achieving significant\nimprovements in fairness without retraining models. These findings highlight\nthe urgent need for ethical AI practices and interdisciplinary solutions to\naddress biases not only at the model level but also in societal structures. Our\nwork provides a scalable and interpretable solution that advances fairness in\nAI systems while offering practical insights for future research on fair AI\ntechnologies.\n", "date": "2025-01-14", "categories": []}
{"id": "2501.11597", "title": "Fairness Testing through Extreme Value Theory", "abstract": "  Data-driven software is increasingly being used as a critical component of\nautomated decision-support systems. Since this class of software learns its\nlogic from historical data, it can encode or amplify discriminatory practices.\nPrevious research on algorithmic fairness has focused on improving average-case\nfairness. On the other hand, fairness at the extreme ends of the spectrum,\nwhich often signifies lasting and impactful shifts in societal attitudes, has\nreceived significantly less emphasis.\n  Leveraging the statistics of extreme value theory (EVT), we propose a novel\nfairness criterion called extreme counterfactual discrimination (ECD). This\ncriterion estimates the worst-case amounts of disadvantage in outcomes for\nindividuals solely based on their memberships in a protected group. Utilizing\ntools from search-based software engineering and generative AI, we present a\nrandomized algorithm that samples a statistically significant set of points\nfrom the tail of ML outcome distributions even if the input dataset lacks a\nsufficient number of relevant samples.\n  We conducted several experiments on four ML models (deep neural networks,\nlogistic regression, and random forests) over 10 socially relevant tasks from\nthe literature on algorithmic fairness. First, we evaluate the generative AI\nmethods and find that they generate sufficient samples to infer valid EVT\ndistribution in 95% of cases. Remarkably, we found that the prevalent bias\nmitigators reduce the average-case discrimination but increase the worst-case\ndiscrimination significantly in 5% of cases. We also observed that even the\ntail-aware mitigation algorithm -- MiniMax-Fairness -- increased the worst-case\ndiscrimination in 30% of cases. We propose a novel ECD-based mitigator that\nimproves fairness in the tail in 90% of cases with no degradation of the\naverage-case discrimination.\n", "date": "2025-01-20", "categories": []}
{"id": "2406.15484", "title": "JobFair: A Framework for Benchmarking Gender Hiring Bias in Large\n  Language Models", "abstract": "  The use of Large Language Models (LLMs) in hiring has led to legislative\nactions to protect vulnerable demographic groups. This paper presents a novel\nframework for benchmarking hierarchical gender hiring bias in Large Language\nModels (LLMs) for resume scoring, revealing significant issues of reverse\ngender hiring bias and overdebiasing. Our contributions are fourfold: Firstly,\nwe introduce a new construct grounded in labour economics, legal principles,\nand critiques of current bias benchmarks: hiring bias can be categorized into\ntwo types: Level bias (difference in the average outcomes between demographic\ncounterfactual groups) and Spread bias (difference in the variance of outcomes\nbetween demographic counterfactual groups); Level bias can be further\nsubdivided into statistical bias (i.e. changing with non-demographic content)\nand taste-based bias (i.e. consistent regardless of non-demographic content).\nSecondly, the framework includes rigorous statistical and computational hiring\nbias metrics, such as Rank After Scoring (RAS), Rank-based Impact Ratio,\nPermutation Test, and Fixed Effects Model. Thirdly, we analyze gender hiring\nbiases in ten state-of-the-art LLMs. Seven out of ten LLMs show significant\nbiases against males in at least one industry. An industry-effect regression\nreveals that the healthcare industry is the most biased against males.\nMoreover, we found that the bias performance remains invariant with resume\ncontent for eight out of ten LLMs. This indicates that the bias performance\nmeasured in this paper might apply to other resume datasets with different\nresume qualities. Fourthly, we provide a user-friendly demo and resume dataset\nto support the adoption and practical use of the framework, which can be\ngeneralized to other social traits and tasks.\n", "date": "2024-06-17", "categories": []}
{"id": "2501.08910", "title": "Lights, Camera, Matching: The Role of Image Illumination in Fair Face\n  Recognition", "abstract": "  Facial brightness is a key image quality factor impacting face recognition\naccuracy differentials across demographic groups. In this work, we aim to\ndecrease the accuracy gap between the similarity score distributions for\nCaucasian and African American female mated image pairs, as measured by d'\nbetween distributions. To balance brightness across demographic groups, we\nconduct three experiments, interpreting brightness in the face skin region\neither as median pixel value or as the distribution of pixel values. Balancing\nbased on median brightness alone yields up to a 46.8% decrease in d', while\nbalancing based on brightness distribution yields up to a 57.6% decrease. In\nall three cases, the similarity scores of the individual distributions improve,\nwith mean scores maximally improving 5.9% for Caucasian females and 3.7% for\nAfrican American females.\n", "date": "2025-01-15", "categories": []}
{"id": "2409.08558", "title": "Fair CoVariance Neural Networks", "abstract": "  Covariance-based data processing is widespread across signal processing and\nmachine learning applications due to its ability to model data\ninterconnectivities and dependencies. However, harmful biases in the data may\nbecome encoded in the sample covariance matrix and cause data-driven methods to\ntreat different subpopulations unfairly. Existing works such as fair principal\ncomponent analysis (PCA) mitigate these effects, but remain unstable in low\nsample regimes, which in turn may jeopardize the fairness goal. To address both\nbiases and instability, we propose Fair coVariance Neural Networks (FVNNs),\nwhich perform graph convolutions on the covariance matrix for both fair and\naccurate predictions. Our FVNNs provide a flexible model compatible with\nseveral existing bias mitigation techniques. In particular, FVNNs allow for\nmitigating the bias in two ways: first, they operate on fair covariance\nestimates that remove biases from their principal components; second, they are\ntrained in an end-to-end fashion via a fairness regularizer in the loss\nfunction so that the model parameters are tailored to solve the task directly\nin a fair manner. We prove that FVNNs are intrinsically fairer than analogous\nPCA approaches thanks to their stability in low sample regimes. We validate the\nrobustness and fairness of our model on synthetic and real-world data,\nshowcasing the flexibility of FVNNs along with the tradeoff between fair and\naccurate performance.\n", "date": "2024-09-13", "categories": []}
{"id": "2501.07885", "title": "Mitigating Algorithmic Bias in Multiclass CNN Classifications Using\n  Causal Modeling", "abstract": "  This study describes a procedure for applying causal modeling to detect and\nmitigate algorithmic bias in a multiclass classification problem. The dataset\nwas derived from the FairFace dataset, supplemented with emotional labels\ngenerated by the DeepFace pre-trained model. A custom Convolutional Neural\nNetwork (CNN) was developed, consisting of four convolutional blocks, followed\nby fully connected layers and dropout layers to mitigate overfitting. Gender\nbias was identified in the CNN model's classifications: Females were more\nlikely to be classified as \"happy\" or \"sad,\" while males were more likely to be\nclassified as \"neutral.\" To address this, the one-vs-all (OvA) technique was\napplied. A causal model was constructed for each emotion class to adjust the\nCNN model's predicted class probabilities. The adjusted probabilities for the\nvarious classes were then aggregated by selecting the class with the highest\nprobability. The resulting debiased classifications demonstrated enhanced\ngender fairness across all classes, with negligible impact--or even a slight\nimprovement--on overall accuracy. This study highlights that algorithmic\nfairness and accuracy are not necessarily trade-offs. All data and code for\nthis study are publicly available for download.\n", "date": "2025-01-14", "categories": []}
{"id": "2501.08155", "title": "FairTTTS: A Tree Test Time Simulation Method for Fairness-Aware\n  Classification", "abstract": "  Algorithmic decision-making has become deeply ingrained in many domains, yet\nbiases in machine learning models can still produce discriminatory outcomes,\noften harming unprivileged groups. Achieving fair classification is inherently\nchallenging, requiring a careful balance between predictive performance and\nethical considerations. We present FairTTTS, a novel post-processing bias\nmitigation method inspired by the Tree Test Time Simulation (TTTS) method.\nOriginally developed to enhance accuracy and robustness against adversarial\ninputs through probabilistic decision-path adjustments, TTTS serves as the\nfoundation for FairTTTS. By building on this accuracy-enhancing technique,\nFairTTTS mitigates bias and improves predictive performance. FairTTTS uses a\ndistance-based heuristic to adjust decisions at protected attribute nodes,\nensuring fairness for unprivileged samples. This fairness-oriented adjustment\noccurs as a post-processing step, allowing FairTTTS to be applied to\npre-trained models, diverse datasets, and various fairness metrics without\nretraining. Extensive evaluation on seven benchmark datasets shows that\nFairTTTS outperforms traditional methods in fairness improvement, achieving a\n20.96% average increase over the baseline compared to 18.78% for related work,\nand further enhances accuracy by 0.55%. In contrast, competing methods\ntypically reduce accuracy by 0.42%. These results confirm that FairTTTS\neffectively promotes more equitable decision-making while simultaneously\nimproving predictive performance.\n", "date": "2025-01-14", "categories": []}
{"id": "2305.15622", "title": "GFairHint: Improving Individual Fairness for Graph Neural Networks via\n  Fairness Hint", "abstract": "  Given the growing concerns about fairness in machine learning and the\nimpressive performance of Graph Neural Networks (GNNs) on graph data learning,\nalgorithmic fairness in GNNs has attracted significant attention. While many\nexisting studies improve fairness at the group level, only a few works promote\nindividual fairness, which renders similar outcomes for similar individuals. A\ndesirable framework that promotes individual fairness should (1) balance\nbetween fairness and performance, (2) accommodate two commonly-used individual\nsimilarity measures (externally annotated and computed from input features),\n(3) generalize across various GNN models, and (4) be computationally efficient.\nUnfortunately, none of the prior work achieves all the desirables. In this\nwork, we propose a novel method, GFairHint, which promotes individual fairness\nin GNNs and achieves all aforementioned desirables. GFairHint learns fairness\nrepresentations through an auxiliary link prediction task, and then\nconcatenates the representations with the learned node embeddings in original\nGNNs as a \"fairness hint\". Through extensive experimental investigations on\nfive real-world graph datasets under three prevalent GNN models covering both\nindividual similarity measures above, GFairHint achieves the best fairness\nresults in almost all combinations of datasets with various backbone models,\nwhile generating comparable utility results, with much less computational cost\ncompared to the previous state-of-the-art (SoTA) method.\n", "date": "2023-05-25", "categories": []}
{"id": "2404.08973", "title": "PraFFL: A Preference-Aware Scheme in Fair Federated Learning", "abstract": "  Fairness in federated learning has emerged as a critical concern, aiming to\ndevelop an unbiased model among groups (e.g., male or female) of diverse\nsensitive features. However, there is a trade-off between model performance and\nfairness, i.e., improving model fairness will decrease model performance.\nExisting approaches have characterized such a trade-off by introducing\nhyperparameters to quantify client's preferences for model fairness and model\nperformance. Nevertheless, these approaches are limited to scenarios where each\nclient has only a single pre-defined preference, and fail to work in practical\nsystems where each client generally has multiple preferences. To this end, we\npropose a Preference-aware scheme in Fair Federated Learning (called PraFFL) to\ngenerate preference-specific models in real time. PraFFL can adaptively adjust\nthe model based on each client's preferences to meet their needs. We\ntheoretically prove that PraFFL can offer the optimal model tailored to an\narbitrary preference of each client, and show its linear convergence.\nExperimental results show that our proposed PraFFL outperforms six fair\nfederated learning algorithms in terms of the model's capability of adapting to\nclients' different preferences. Our implementation is available at\nhttps://github.com/rG223/PraFFL.\n", "date": "2024-04-13", "categories": []}
{"id": "2501.06753", "title": "Procedural Fairness and Its Relationship with Distributive Fairness in\n  Machine Learning", "abstract": "  Fairness in machine learning (ML) has garnered significant attention in\nrecent years. While existing research has predominantly focused on the\ndistributive fairness of ML models, there has been limited exploration of\nprocedural fairness. This paper proposes a novel method to achieve procedural\nfairness during the model training phase. The effectiveness of the proposed\nmethod is validated through experiments conducted on one synthetic and six\nreal-world datasets. Additionally, this work studies the relationship between\nprocedural fairness and distributive fairness in ML models. On one hand, the\nimpact of dataset bias and the procedural fairness of ML model on its\ndistributive fairness is examined. The results highlight a significant\ninfluence of both dataset bias and procedural fairness on distributive\nfairness. On the other hand, the distinctions between optimizing procedural and\ndistributive fairness metrics are analyzed. Experimental results demonstrate\nthat optimizing procedural fairness metrics mitigates biases introduced or\namplified by the decision-making process, thereby ensuring fairness in the\ndecision-making process itself, as well as improving distributive fairness. In\ncontrast, optimizing distributive fairness metrics encourages the ML model's\ndecision-making process to favor disadvantaged groups, counterbalancing the\ninherent preferences for advantaged groups present in the dataset and\nultimately achieving distributive fairness.\n", "date": "2025-01-12", "categories": []}
{"id": "2501.07324", "title": "Foundation Models at Work: Fine-Tuning for Fairness in Algorithmic\n  Hiring", "abstract": "  Foundation models require fine-tuning to ensure their generative outputs\nalign with intended results for specific tasks. Automating this fine-tuning\nprocess is challenging, as it typically needs human feedback that can be\nexpensive to acquire. We present AutoRefine, a method that leverages\nreinforcement learning for targeted fine-tuning, utilizing direct feedback from\nmeasurable performance improvements in specific downstream tasks. We\ndemonstrate the method for a problem arising in algorithmic hiring platforms\nwhere linguistic biases influence a recommendation system. In this setting, a\ngenerative model seeks to rewrite given job specifications to receive more\ndiverse candidate matches from a recommendation engine which matches jobs to\ncandidates. Our model detects and regulates biases in job descriptions to meet\ndiversity and fairness criteria. The experiments on a public hiring dataset and\na real-world hiring platform showcase how large language models can assist in\nidentifying and mitigation biases in the real world.\n", "date": "2025-01-13", "categories": []}
{"id": "2501.01987", "title": "Gender Bias in Text-to-Video Generation Models: A case study of Sora", "abstract": "  The advent of text-to-video generation models has revolutionized content\ncreation as it produces high-quality videos from textual prompts. However,\nconcerns regarding inherent biases in such models have prompted scrutiny,\nparticularly regarding gender representation. Our study investigates the\npresence of gender bias in OpenAI's Sora, a state-of-the-art text-to-video\ngeneration model. We uncover significant evidence of bias by analyzing the\ngenerated videos from a diverse set of gender-neutral and stereotypical\nprompts. The results indicate that Sora disproportionately associates specific\ngenders with stereotypical behaviors and professions, which reflects societal\nprejudices embedded in its training data.\n", "date": "2024-12-30", "categories": []}
{"id": "2409.10676", "title": "Toward Mitigating Sex Bias in Pilot Trainees' Stress and Fatigue\n  Modeling", "abstract": "  While researchers have been trying to understand the stress and fatigue among\npilots, especially pilot trainees, and to develop stress/fatigue models to\nautomate the process of detecting stress/fatigue, they often do not consider\nbiases such as sex in those models. However, in a critical profession like\naviation, where the demographic distribution is disproportionately skewed to\none sex, it is urgent to mitigate biases for fair and safe model predictions.\nIn this work, we investigate the perceived stress/fatigue of 69 college\nstudents, including 40 pilot trainees with around 63% male. We construct models\nwith decision trees first without bias mitigation and then with bias mitigation\nusing a threshold optimizer with demographic parity and equalized odds\nconstraints 30 times with random instances. Using bias mitigation, we achieve\nimprovements of 88.31% (demographic parity difference) and 54.26% (equalized\nodds difference), which are also found to be statistically significant.\n", "date": "2024-09-16", "categories": []}
{"id": "2409.10677", "title": "Mitigating Sex Bias in Audio Data-driven COPD and COVID-19 Breathing\n  Pattern Detection Models", "abstract": "  In the healthcare industry, researchers have been developing machine learning\nmodels to automate diagnosing patients with respiratory illnesses based on\ntheir breathing patterns. However, these models do not consider the demographic\nbiases, particularly sex bias, that often occur when models are trained with a\nskewed patient dataset. Hence, it is essential in such an important industry to\nreduce this bias so that models can make fair diagnoses. In this work, we\nexamine the bias in models used to detect breathing patterns of two major\nrespiratory diseases, i.e., chronic obstructive pulmonary disease (COPD) and\nCOVID-19. Using decision tree models trained with audio recordings of breathing\npatterns obtained from two open-source datasets consisting of 29 COPD and 680\nCOVID-19-positive patients, we analyze the effect of sex bias on the models.\nWith a threshold optimizer and two constraints (demographic parity and\nequalized odds) to mitigate the bias, we witness 81.43% (demographic parity\ndifference) and 71.81% (equalized odds difference) improvements. These findings\nare statistically significant.\n", "date": "2024-09-16", "categories": []}
{"id": "2407.01996", "title": "ViG-Bias: Visually Grounded Bias Discovery and Mitigation", "abstract": "  The proliferation of machine learning models in critical decision making\nprocesses has underscored the need for bias discovery and mitigation\nstrategies. Identifying the reasons behind a biased system is not\nstraightforward, since in many occasions they are associated with hidden\nspurious correlations which are not easy to spot. Standard approaches rely on\nbias audits performed by analyzing model performance in pre-defined subgroups\nof data samples, usually characterized by common attributes like gender or\nethnicity when it comes to people, or other specific attributes defining\nsemantically coherent groups of images. However, it is not always possible to\nknow a-priori the specific attributes defining the failure modes of visual\nrecognition systems. Recent approaches propose to discover these groups by\nleveraging large vision language models, which enable the extraction of\ncross-modal embeddings and the generation of textual descriptions to\ncharacterize the subgroups where a certain model is underperforming. In this\nwork, we argue that incorporating visual explanations (e.g. heatmaps generated\nvia GradCAM or other approaches) can boost the performance of such bias\ndiscovery and mitigation frameworks. To this end, we introduce Visually\nGrounded Bias Discovery and Mitigation (ViG-Bias), a simple yet effective\ntechnique which can be integrated to a variety of existing frameworks to\nimprove both, discovery and mitigation performance. Our comprehensive\nevaluation shows that incorporating visual explanations enhances existing\ntechniques like DOMINO, FACTS and Bias-to-Text, across several challenging\ndatasets, including CelebA, Waterbirds, and NICO++.\n", "date": "2024-07-02", "categories": []}
{"id": "2501.04142", "title": "BiasGuard: Guardrailing Fairness in Machine Learning Production Systems", "abstract": "  As machine learning (ML) systems increasingly impact critical sectors such as\nhiring, financial risk assessments, and criminal justice, the imperative to\nensure fairness has intensified due to potential negative implications. While\nmuch ML fairness research has focused on enhancing training data and processes,\naddressing the outputs of already deployed systems has received less attention.\nThis paper introduces 'BiasGuard', a novel approach designed to act as a\nfairness guardrail in production ML systems. BiasGuard leverages Test-Time\nAugmentation (TTA) powered by Conditional Generative Adversarial Network\n(CTGAN), a cutting-edge generative AI model, to synthesize data samples\nconditioned on inverted protected attribute values, thereby promoting equitable\noutcomes across diverse groups. This method aims to provide equal opportunities\nfor both privileged and unprivileged groups while significantly enhancing the\nfairness metrics of deployed systems without the need for retraining. Our\ncomprehensive experimental analysis across diverse datasets reveals that\nBiasGuard enhances fairness by 31% while only reducing accuracy by 0.09%\ncompared to non-mitigated benchmarks. Additionally, BiasGuard outperforms\nexisting post-processing methods in improving fairness, positioning it as an\neffective tool to safeguard against biases when retraining the model is\nimpractical.\n", "date": "2025-01-07", "categories": []}
{"id": "2402.11512", "title": "From Prejudice to Parity: A New Approach to Debiasing Large Language\n  Model Word Embeddings", "abstract": "  Embeddings play a pivotal role in the efficacy of Large Language Models. They\nare the bedrock on which these models grasp contextual relationships and foster\na more nuanced understanding of language and consequently perform remarkably on\na plethora of complex tasks that require a fundamental understanding of human\nlanguage. Given that these embeddings themselves often reflect or exhibit bias,\nit stands to reason that these models may also inadvertently learn this bias.\nIn this work, we build on the seminal previous work and propose DeepSoftDebias,\nan algorithm that uses a neural network to perform 'soft debiasing'. We\nexhaustively evaluate this algorithm across a variety of SOTA datasets,\naccuracy metrics, and challenging NLP tasks. We find that DeepSoftDebias\noutperforms the current state-of-the-art methods at reducing bias across\ngender, race, and religion.\n", "date": "2024-02-18", "categories": []}
{"id": "2408.12875", "title": "Disentangling, Amplifying, and Debiasing: Learning Disentangled\n  Representations for Fair Graph Neural Networks", "abstract": "  Graph Neural Networks (GNNs) have become essential tools for graph\nrepresentation learning in various domains, such as social media and\nhealthcare. However, they often suffer from fairness issues due to inherent\nbiases in node attributes and graph structure, leading to unfair predictions.\nTo address these challenges, we propose a novel GNN framework, DAB-GNN, that\nDisentangles, Amplifies, and deBiases attribute, structure, and potential\nbiases in the GNN mechanism. DAB-GNN employs a disentanglement and\namplification module that isolates and amplifies each type of bias through\nspecialized disentanglers, followed by a debiasing module that minimizes the\ndistance between subgroup distributions. Extensive experiments on five datasets\ndemonstrate that DAB-GNN significantly outperforms ten state-of-the-art\ncompetitors in terms of achieving an optimal balance between accuracy and\nfairness. The codebase of DAB-GNN is available at\nhttps://github.com/Bigdasgit/DAB-GNN\n", "date": "2024-08-23", "categories": []}
{"id": "2501.02211", "title": "Examining the Robustness of Homogeneity Bias to Hyperparameter\n  Adjustments in GPT-4", "abstract": "  Vision-Language Models trained on massive collections of human-generated data\noften reproduce and amplify societal stereotypes. One critical form of\nstereotyping reproduced by these models is homogeneity bias-the tendency to\nrepresent certain groups as more homogeneous than others. We investigate how\nthis bias responds to hyperparameter adjustments in GPT-4, specifically\nexamining sampling temperature and top p which control the randomness of model\noutputs. By generating stories about individuals from different racial and\ngender groups and comparing their similarities using vector representations, we\nassess both bias robustness and its relationship with hyperparameter values. We\nfind that (1) homogeneity bias persists across most hyperparameter\nconfigurations, with Black Americans and women being represented more\nhomogeneously than White Americans and men, (2) the relationship between\nhyperparameters and group representations shows unexpected non-linear patterns,\nparticularly at extreme values, and (3) hyperparameter adjustments affect\nracial and gender homogeneity bias differently-while increasing temperature or\ndecreasing top p can reduce racial homogeneity bias, these changes show\ndifferent effects on gender homogeneity bias. Our findings suggest that while\nhyperparameter tuning may mitigate certain biases to some extent, it cannot\nserve as a universal solution for addressing homogeneity bias across different\nsocial group dimensions.\n", "date": "2025-01-04", "categories": []}
{"id": "2501.02732", "title": "AFed: Algorithmic Fair Federated Learning", "abstract": "  Federated Learning (FL) has gained significant attention as it facilitates\ncollaborative machine learning among multiple clients without centralizing\ntheir data on a server. FL ensures the privacy of participating clients by\nlocally storing their data, which creates new challenges in fairness.\nTraditional debiasing methods assume centralized access to sensitive\ninformation, rendering them impractical for the FL setting. Additionally, FL is\nmore susceptible to fairness issues than centralized machine learning due to\nthe diverse client data sources that may be associated with group information.\nTherefore, training a fair model in FL without access to client local data is\nimportant and challenging. This paper presents AFed, a straightforward yet\neffective framework for promoting group fairness in FL. The core idea is to\ncircumvent restricted data access by learning the global data distribution.\nThis paper proposes two approaches: AFed-G, which uses a conditional generator\ntrained on the server side, and AFed-GAN, which improves upon AFed-G by\ntraining a conditional GAN on the client side. We augment the client data with\nthe generated samples to help remove bias. Our theoretical analysis justifies\nthe proposed methods, and empirical results on multiple real-world datasets\ndemonstrate a substantial improvement in AFed over several baselines.\n", "date": "2025-01-06", "categories": []}
{"id": "2501.02793", "title": "Fairness Through Matching", "abstract": "  Group fairness requires that different protected groups, characterized by a\ngiven sensitive attribute, receive equal outcomes overall. Typically, the level\nof group fairness is measured by the statistical gap between predictions from\ndifferent protected groups. In this study, we reveal an implicit property of\nexisting group fairness measures, which provides an insight into how the\ngroup-fair models behave. Then, we develop a new group-fair constraint based on\nthis implicit property to learn group-fair models. To do so, we first introduce\na notable theoretical observation: every group-fair model has an implicitly\ncorresponding transport map between the input spaces of each protected group.\nBased on this observation, we introduce a new group fairness measure termed\nMatched Demographic Parity (MDP), which quantifies the averaged gap between\npredictions of two individuals (from different protected groups) matched by a\ngiven transport map. Then, we prove that any transport map can be used in MDP\nto learn group-fair models, and develop a novel algorithm called Fairness\nThrough Matching (FTM), which learns a group-fair model using MDP constraint\nwith an user-specified transport map. We specifically propose two favorable\ntypes of transport maps for MDP, based on the optimal transport theory, and\ndiscuss their advantages. Experiments reveal that FTM successfully trains\ngroup-fair models with certain desirable properties by choosing the transport\nmap accordingly.\n", "date": "2025-01-06", "categories": []}
{"id": "2501.01665", "title": "FairSense: Long-Term Fairness Analysis of ML-Enabled Systems", "abstract": "  Algorithmic fairness of machine learning (ML) models has raised significant\nconcern in the recent years. Many testing, verification, and bias mitigation\ntechniques have been proposed to identify and reduce fairness issues in ML\nmodels. The existing methods are model-centric and designed to detect fairness\nissues under static settings. However, many ML-enabled systems operate in a\ndynamic environment where the predictive decisions made by the system impact\nthe environment, which in turn affects future decision-making. Such a\nself-reinforcing feedback loop can cause fairness violations in the long term,\neven if the immediate outcomes are fair. In this paper, we propose a\nsimulation-based framework called FairSense to detect and analyze long-term\nunfairness in ML-enabled systems. Given a fairness requirement, FairSense\nperforms Monte-Carlo simulation to enumerate evolution traces for each system\nconfiguration. Then, FairSense performs sensitivity analysis on the space of\npossible configurations to understand the impact of design options and\nenvironmental factors on the long-term fairness of the system. We demonstrate\nFairSense's potential utility through three real-world case studies: Loan\nlending, opioids risk scoring, and predictive policing.\n", "date": "2025-01-03", "categories": []}
{"id": "2501.01889", "title": "Exploring Equality: An Investigation into Custom Loss Functions for\n  Fairness Definitions", "abstract": "  This paper explores the complex tradeoffs between various fairness metrics\nsuch as equalized odds, disparate impact, and equal opportunity and predictive\naccuracy within COMPAS by building neural networks trained with custom loss\nfunctions optimized to specific fairness criteria. This paper creates the first\nfairness-driven implementation of the novel Group Accuracy Parity (GAP)\nframework, as theoretically proposed by Gupta et al. (2024), and applies it to\nCOMPAS. To operationalize and accurately compare the fairness of COMPAS models\noptimized to differing fairness ideals, this paper develops and proposes a\ncombinatory analytical procedure that incorporates Pareto front and\nmultivariate analysis, leveraging data visualizations such as violin graphs.\nThis paper concludes that GAP achieves an enhanced equilibrium between fairness\nand accuracy compared to COMPAS's current nationwide implementation and\nalternative implementations of COMPAS optimized to more traditional fairness\ndefinitions. While this paper's algorithmic improvements of COMPAS\nsignificantly augment its fairness, external biases undermine the fairness of\nits implementation. Practices such as predictive policing and issues such as\nthe lack of transparency regarding COMPAS's internal workings have contributed\nto the algorithm's historical injustice. In conjunction with developments\nregarding COMPAS's predictive methodology, legal and institutional changes must\nhappen for COMPAS's just deployment.\n", "date": "2025-01-03", "categories": []}
{"id": "2412.04655", "title": "From Models to Systems: A Comprehensive Fairness Framework for\n  Compositional Recommender Systems", "abstract": "  Fairness research in machine learning often centers on ensuring equitable\nperformance of individual models. However, real-world recommendation systems\nare built on multiple models and even multiple stages, from candidate retrieval\nto scoring and serving, which raises challenges for responsible development and\ndeployment. This system-level view, as highlighted by regulations like the EU\nAI Act, necessitates moving beyond auditing individual models as independent\nentities. We propose a holistic framework for modeling system-level fairness,\nfocusing on the end-utility delivered to diverse user groups, and consider\ninteractions between components such as retrieval and scoring models. We\nprovide formal insights on the limitations of focusing solely on model-level\nfairness and highlight the need for alternative tools that account for\nheterogeneity in user preferences. To mitigate system-level disparities, we\nadapt closed-box optimization tools (e.g., BayesOpt) to jointly optimize\nutility and equity. We empirically demonstrate the effectiveness of our\nproposed framework on synthetic and real datasets, underscoring the need for a\nsystem-level framework.\n", "date": "2024-12-05", "categories": []}
{"id": "2412.10669", "title": "FairGP: A Scalable and Fair Graph Transformer Using Graph Partitioning", "abstract": "  Recent studies have highlighted significant fairness issues in Graph\nTransformer (GT) models, particularly against subgroups defined by sensitive\nfeatures. Additionally, GTs are computationally intensive and memory-demanding,\nlimiting their application to large-scale graphs. Our experiments demonstrate\nthat graph partitioning can enhance the fairness of GT models while reducing\ncomputational complexity. To understand this improvement, we conducted a\ntheoretical investigation into the root causes of fairness issues in GT models.\nWe found that the sensitive features of higher-order nodes disproportionately\ninfluence lower-order nodes, resulting in sensitive feature bias. We propose\nFairness-aware scalable GT based on Graph Partitioning (FairGP), which\npartitions the graph to minimize the negative impact of higher-order nodes. By\noptimizing attention mechanisms, FairGP mitigates the bias introduced by global\nattention, thereby enhancing fairness. Extensive empirical evaluations on six\nreal-world datasets validate the superior performance of FairGP in achieving\nfairness compared to state-of-the-art methods. The codes are available at\nhttps://github.com/LuoRenqiang/FairGP.\n", "date": "2024-12-14", "categories": []}
{"id": "2412.16428", "title": "Data-Driven Fairness Generalization for Deepfake Detection", "abstract": "  Despite the progress made in deepfake detection research, recent studies have\nshown that biases in the training data for these detectors can result in\nvarying levels of performance across different demographic groups, such as race\nand gender. These disparities can lead to certain groups being unfairly\ntargeted or excluded. Traditional methods often rely on fair loss functions to\naddress these issues, but they under-perform when applied to unseen datasets,\nhence, fairness generalization remains a challenge. In this work, we propose a\ndata-driven framework for tackling the fairness generalization problem in\ndeepfake detection by leveraging synthetic datasets and model optimization. Our\napproach focuses on generating and utilizing synthetic data to enhance fairness\nacross diverse demographic groups. By creating a diverse set of synthetic\nsamples that represent various demographic groups, we ensure that our model is\ntrained on a balanced and representative dataset. This approach allows us to\ngeneralize fairness more effectively across different domains. We employ a\ncomprehensive strategy that leverages synthetic data, a loss sharpness-aware\noptimization pipeline, and a multi-task learning framework to create a more\nequitable training environment, which helps maintain fairness across both\nintra-dataset and cross-dataset evaluations. Extensive experiments on benchmark\ndeepfake detection datasets demonstrate the efficacy of our approach,\nsurpassing state-of-the-art approaches in preserving fairness during\ncross-dataset evaluation. Our results highlight the potential of synthetic\ndatasets in achieving fairness generalization, providing a robust solution for\nthe challenges faced in deepfake detection.\n", "date": "2024-12-21", "categories": []}
{"id": "2412.17123", "title": "Fairness in Reinforcement Learning with Bisimulation Metrics", "abstract": "  Ensuring long-term fairness is crucial when developing automated decision\nmaking systems, specifically in dynamic and sequential environments. By\nmaximizing their reward without consideration of fairness, AI agents can\nintroduce disparities in their treatment of groups or individuals. In this\npaper, we establish the connection between bisimulation metrics and group\nfairness in reinforcement learning. We propose a novel approach that leverages\nbisimulation metrics to learn reward functions and observation dynamics,\nensuring that learners treat groups fairly while reflecting the original\nproblem. We demonstrate the effectiveness of our method in addressing\ndisparities in sequential decision making problems through empirical evaluation\non a standard fairness benchmark consisting of lending and college admission\nscenarios.\n", "date": "2024-12-22", "categories": []}
{"id": "2501.00129", "title": "A Data-Centric Approach to Detecting and Mitigating Demographic Bias in\n  Pediatric Mental Health Text: A Case Study in Anxiety Detection", "abstract": "  Introduction: Healthcare AI models often inherit biases from their training\ndata. While efforts have primarily targeted bias in structured data, mental\nhealth heavily depends on unstructured data. This study aims to detect and\nmitigate linguistic differences related to non-biological differences in the\ntraining data of AI models designed to assist in pediatric mental health\nscreening. Our objectives are: (1) to assess the presence of bias by evaluating\noutcome parity across sex subgroups, (2) to identify bias sources through\ntextual distribution analysis, and (3) to develop a de-biasing method for\nmental health text data. Methods: We examined classification parity across\ndemographic groups and assessed how gendered language influences model\npredictions. A data-centric de-biasing method was applied, focusing on\nneutralizing biased terms while retaining salient clinical information. This\nmethodology was tested on a model for automatic anxiety detection in pediatric\npatients. Results: Our findings revealed a systematic under-diagnosis of female\nadolescent patients, with a 4% lower accuracy and a 9% higher False Negative\nRate (FNR) compared to male patients, likely due to disparities in information\ndensity and linguistic differences in patient notes. Notes for male patients\nwere on average 500 words longer, and linguistic similarity metrics indicated\ndistinct word distributions between genders. Implementing our de-biasing\napproach reduced diagnostic bias by up to 27%, demonstrating its effectiveness\nin enhancing equity across demographic groups. Discussion: We developed a\ndata-centric de-biasing framework to address gender-based content disparities\nwithin clinical text. By neutralizing biased language and enhancing focus on\nclinically essential information, our approach demonstrates an effective\nstrategy for mitigating bias in AI healthcare models trained on text.\n", "date": "2024-12-30", "categories": []}
{"id": "2501.00595", "title": "Unbiased GNN Learning via Fairness-Aware Subgraph Diffusion", "abstract": "  Graph Neural Networks (GNNs) have demonstrated remarkable efficacy in\ntackling a wide array of graph-related tasks across diverse domains. However, a\nsignificant challenge lies in their propensity to generate biased predictions,\nparticularly with respect to sensitive node attributes such as age and gender.\nThese biases, inherent in many machine learning models, are amplified in GNNs\ndue to the message-passing mechanism, which allows nodes to influence each\nother, rendering the task of making fair predictions notably challenging. This\nissue is particularly pertinent in critical domains where model fairness holds\nparamount importance. In this paper, we propose a novel generative\nFairness-Aware Subgraph Diffusion (FASD) method for unbiased GNN learning. The\nmethod initiates by strategically sampling small subgraphs from the original\nlarge input graph, and then proceeds to conduct subgraph debiasing via\ngenerative fairness-aware graph diffusion processes based on stochastic\ndifferential equations (SDEs). To effectively diffuse unfairness in the input\ndata, we introduce additional adversary bias perturbations to the subgraphs\nduring the forward diffusion process, and train score-based models to predict\nthese applied perturbations, enabling them to learn the underlying dynamics of\nthe biases present in the data. Subsequently, the trained score-based models\nare utilized to further debias the original subgraph samples through the\nreverse diffusion process. Finally, FASD induces fair node predictions on the\ninput graph by performing standard GNN learning on the debiased subgraphs.\nExperimental results demonstrate the superior performance of the proposed\nmethod over state-of-the-art Fair GNN baselines across multiple benchmark\ndatasets.\n", "date": "2024-12-31", "categories": []}
{"id": "2501.00995", "title": "Is It Still Fair? Investigating Gender Fairness in Cross-Corpus Speech\n  Emotion Recognition", "abstract": "  Speech emotion recognition (SER) is a vital component in various everyday\napplications. Cross-corpus SER models are increasingly recognized for their\nability to generalize performance. However, concerns arise regarding fairness\nacross demographics in diverse corpora. Existing fairness research often\nfocuses solely on corpus-specific fairness, neglecting its generalizability in\ncross-corpus scenarios. Our study focuses on this underexplored area, examining\nthe gender fairness generalizability in cross-corpus SER scenarios. We\nemphasize that the performance of cross-corpus SER models and their fairness\nare two distinct considerations. Moreover, we propose the approach of a\ncombined fairness adaptation mechanism to enhance gender fairness in the SER\ntransfer learning tasks by addressing both source and target genders. Our\nfindings bring one of the first insights into the generalizability of gender\nfairness in cross-corpus SER systems.\n", "date": "2025-01-02", "categories": []}
{"id": "2310.03146", "title": "Fairness-enhancing mixed effects deep learning improves fairness on in-\n  and out-of-distribution clustered (non-iid) data", "abstract": "  Traditional deep learning (DL) models have two ubiquitous limitations. First,\nthey assume training samples are independent and identically distributed\n(i.i.d), an assumption often violated in real-world datasets where samples have\nadditional correlation due to repeat measurements (e.g., on the same\nparticipants in a longitudinal study or cells from the same sequencer). This\nleads to performance degradation, limited generalization, and covariate\nconfounding, which induces Type I and Type II errors. Second, DL models\ntypically prioritize overall accuracy, favoring accuracy on the majority while\nsacrificing performance for underrepresented subpopulations, leading to unfair,\nbiased models. This is critical to remediate, particularly in models which\ninfluence decisions regarding loan approvals and healthcare. To address these\nissues, we propose the Fair Mixed Effects Deep Learning (Fair MEDL) framework.\nThis framework quantifies cluster-invariant fixed effects (FE) and\ncluster-specific random effects (RE) through: 1) a cluster adversary for\nlearning invariant FE, 2) a Bayesian neural network for RE, and 3) a mixing\nfunction combining FE and RE for final predictions. Fairness is enhanced\nthrough architectural and loss function changes introduced by an adversarial\ndebiasing network. We formally define and demonstrate improved fairness across\nthree metrics: equalized odds, demographic parity, and counterfactual fairness,\nfor both classification and regression tasks. Our method also identifies and\nde-weights confounded covariates, mitigating Type I and II errors. The\nframework is comprehensively evaluated across three datasets spanning two\nindustries, including finance and healthcare. The Fair MEDL framework improves\nfairness by 86.4% for Age, 64.9% for Race, 57.8% for Sex, and 36.2% for Marital\nstatus, while maintaining robust predictive performance.\n", "date": "2023-10-04", "categories": []}
{"id": "2412.03706", "title": "Fairness without Demographics through Learning Graph of Gradients", "abstract": "  Machine learning systems are notoriously prone to biased predictions about\ncertain demographic groups, leading to algorithmic fairness issues. Due to\nprivacy concerns and data quality problems, some demographic information may\nnot be available in the training data and the complex interaction of different\ndemographics can lead to a lot of unknown minority subpopulations, which all\nlimit the applicability of group fairness. Many existing works on fairness\nwithout demographics assume the correlation between groups and features.\nHowever, we argue that the model gradients are also valuable for fairness\nwithout demographics. In this paper, we show that the correlation between\ngradients and groups can help identify and improve group fairness. With an\nadversarial weighting architecture, we construct a graph where samples with\nsimilar gradients are connected and learn the weights of different samples from\nit. Unlike the surrogate grouping methods that cluster groups from features and\nlabels as proxy sensitive attribute, our method leverages the graph structure\nas a soft grouping mechanism, which is much more robust to noises. The results\nshow that our method is robust to noise and can improve fairness significantly\nwithout decreasing the overall accuracy too much.\n", "date": "2024-12-04", "categories": []}
{"id": "2412.20671", "title": "Two Birds with One Stone: Improving Rumor Detection by Addressing the\n  Unfairness Issue", "abstract": "  The degraded performance and group unfairness caused by confounding sensitive\nattributes in rumor detection remains relatively unexplored. To address this,\nwe propose a two-step framework. Initially, it identifies confounding sensitive\nattributes that limit rumor detection performance and cause unfairness across\ngroups. Subsequently, we aim to learn equally informative representations\nthrough invariant learning. Our method considers diverse sets of groups without\nsensitive attribute annotations. Experiments show our method easily integrates\nwith existing rumor detectors, significantly improving both their detection\nperformance and fairness.\n", "date": "2024-12-30", "categories": []}
{"id": "2412.21052", "title": "Towards Effective Discrimination Testing for Generative AI", "abstract": "  Generative AI (GenAI) models present new challenges in regulating against\ndiscriminatory behavior. In this paper, we argue that GenAI fairness research\nstill has not met these challenges; instead, a significant gap remains between\nexisting bias assessment methods and regulatory goals. This leads to\nineffective regulation that can allow deployment of reportedly fair, yet\nactually discriminatory, GenAI systems. Towards remedying this problem, we\nconnect the legal and technical literature around GenAI bias evaluation and\nidentify areas of misalignment. Through four case studies, we demonstrate how\nthis misalignment between fairness testing techniques and regulatory goals can\nresult in discriminatory outcomes in real-world deployments, especially in\nadaptive or complex environments. We offer practical recommendations for\nimproving discrimination testing to better align with regulatory goals and\nenhance the reliability of fairness assessments in future deployments.\n", "date": "2024-12-30", "categories": []}
{"id": "2407.01423", "title": "FairLay-ML: Intuitive Debugging of Fairness in Data-Driven\n  Social-Critical Software", "abstract": "  Data-driven software solutions have significantly been used in critical\ndomains with significant socio-economic, legal, and ethical implications. The\nrapid adoptions of data-driven solutions, however, pose major threats to the\ntrustworthiness of automated decision-support software. A diminished\nunderstanding of the solution by the developer and historical/current biases in\nthe data sets are primary challenges. To aid data-driven software developers\nand end-users, we present FairLay-ML, a debugging tool to test and explain the\nfairness implications of data-driven solutions. FairLay-ML visualizes the logic\nof datasets, trained models, and decisions for a given data point. In addition,\nit trains various models with varying fairness-accuracy trade-offs. Crucially,\nFairLay-ML incorporates counterfactual fairness testing that finds bugs beyond\nthe development datasets. We conducted two studies through FairLay-ML that\nallowed us to measure false positives/negatives in prevalent counterfactual\ntesting and understand the human perception of counterfactual test cases in a\nclass survey. FairLay-ML and its benchmarks are publicly available at\nhttps://github.com/Pennswood/FairLay-ML. The live version of the tool is\navailable at https://fairlayml-v2.streamlit.app/. We provide a video demo of\nthe tool at https://youtu.be/wNI9UWkywVU?t=133.\n", "date": "2024-07-01", "categories": []}
{"id": "2405.04025", "title": "A Unified Post-Processing Framework for Group Fairness in Classification", "abstract": "  We present a post-processing algorithm for fair classification that covers\ngroup fairness criteria including statistical parity, equal opportunity, and\nequalized odds under a single framework, and is applicable to multiclass\nproblems in both attribute-aware and attribute-blind settings. Our algorithm,\ncalled \"LinearPost\", achieves fairness post-hoc by linearly transforming the\npredictions of the (unfair) base predictor with a \"fairness risk\" according to\na weighted combination of the (predicted) group memberships. It yields the\nBayes optimal fair classifier if the base predictors being post-processed are\nBayes optimal, otherwise, the resulting classifier may not be optimal, but\nfairness is guaranteed as long as the group membership predictor is\nmulticalibrated. The parameters of the post-processing can be efficiently\ncomputed and estimated from solving an empirical linear program. Empirical\nevaluations demonstrate the advantage of our algorithm in the high fairness\nregime compared to existing post-processing and in-processing fair\nclassification algorithms.\n", "date": "2024-05-07", "categories": []}
{"id": "2405.18296", "title": "Bias in Motion: Theoretical Insights into the Dynamics of Bias in SGD\n  Training", "abstract": "  Machine learning systems often acquire biases by leveraging undesired\nfeatures in the data, impacting accuracy variably across different\nsub-populations. Current understanding of bias formation mostly focuses on the\ninitial and final stages of learning, leaving a gap in knowledge regarding the\ntransient dynamics. To address this gap, this paper explores the evolution of\nbias in a teacher-student setup modeling different data sub-populations with a\nGaussian-mixture model. We provide an analytical description of the stochastic\ngradient descent dynamics of a linear classifier in this setting, which we\nprove to be exact in high dimension. Notably, our analysis reveals how\ndifferent properties of sub-populations influence bias at different timescales,\nshowing a shifting preference of the classifier during training. Applying our\nfindings to fairness and robustness, we delineate how and when heterogeneous\ndata and spurious features can generate and amplify bias. We empirically\nvalidate our results in more complex scenarios by training deeper networks on\nsynthetic and real datasets, including CIFAR10, MNIST, and CelebA.\n", "date": "2024-05-28", "categories": []}
{"id": "2412.16523", "title": "Physics-Guided Fair Graph Sampling for Water Temperature Prediction in\n  River Networks", "abstract": "  This work introduces a novel graph neural networks (GNNs)-based method to\npredict stream water temperature and reduce model bias across locations of\ndifferent income and education levels. Traditional physics-based models often\nhave limited accuracy because they are necessarily approximations of reality.\nRecently, there has been an increasing interest of using GNNs in modeling\ncomplex water dynamics in stream networks. Despite their promise in improving\nthe accuracy, GNNs can bring additional model bias through the aggregation\nprocess, where node features are updated by aggregating neighboring nodes. The\nbias can be especially pronounced when nodes with similar sensitive attributes\nare frequently connected. We introduce a new method that leverages physical\nknowledge to represent the node influence in GNNs, and then utilizes\nphysics-based influence to refine the selection and weights over the neighbors.\nThe objective is to facilitate equitable treatment over different sensitive\ngroups in the graph aggregation, which helps reduce spatial bias over\nlocations, especially for those in underprivileged groups. The results on the\nDelaware River Basin demonstrate the effectiveness of the proposed method in\npreserving equitable performance across locations in different sensitive\ngroups.\n", "date": "2024-12-21", "categories": []}
{"id": "2412.16542", "title": "FairDD: Enhancing Fairness with domain-incremental learning in\n  dermatological disease diagnosis", "abstract": "  With the rapid advancement of deep learning technologies, artificial\nintelligence has become increasingly prevalent in the research and application\nof dermatological disease diagnosis. However, this data-driven approach often\nfaces issues related to decision bias. Existing fairness enhancement techniques\ntypically come at a substantial cost to accuracy. This study aims to achieve a\nbetter trade-off between accuracy and fairness in dermatological diagnostic\nmodels. To this end, we propose a novel fair dermatological diagnosis network,\nnamed FairDD, which leverages domain incremental learning to balance the\nlearning of different groups by being sensitive to changes in data\ndistribution. Additionally, we incorporate the mixup data augmentation\ntechnique and supervised contrastive learning to enhance the network's\nrobustness and generalization. Experimental validation on two dermatological\ndatasets demonstrates that our proposed method excels in both fairness criteria\nand the trade-off between fairness and performance.\n", "date": "2024-12-21", "categories": []}
{"id": "2412.17116", "title": "Fair and Accurate Regression: Strong Formulations and Algorithms", "abstract": "  This paper introduces mixed-integer optimization methods to solve regression\nproblems that incorporate fairness metrics. We propose an exact formulation for\ntraining fair regression models. To tackle this computationally hard problem,\nwe study the polynomially-solvable single-factor and single-observation\nsubproblems as building blocks and derive their closed convex hull\ndescriptions. Strong formulations obtained for the general fair regression\nproblem in this manner are utilized to solve the problem with a\nbranch-and-bound algorithm exactly or as a relaxation to produce fair and\naccurate models rapidly. Moreover, to handle large-scale instances, we develop\na coordinate descent algorithm motivated by the convex-hull representation of\nthe single-factor fair regression problem to improve a given solution\nefficiently. Numerical experiments conducted on fair least squares and fair\nlogistic regression problems show competitive statistical performance with\nstate-of-the-art methods while significantly reducing training times.\n", "date": "2024-12-22", "categories": []}
{"id": "2412.17505", "title": "More is Less? A Simulation-Based Approach to Dynamic Interactions\n  between Biases in Multimodal Models", "abstract": "  Multimodal machine learning models, such as those that combine text and image\nmodalities, are increasingly used in critical domains including public safety,\nsecurity, and healthcare. However, these systems inherit biases from their\nsingle modalities. This study proposes a systemic framework for analyzing\ndynamic multimodal bias interactions. Using the MMBias dataset, which\nencompasses categories prone to bias such as religion, nationality, and sexual\norientation, this study adopts a simulation-based heuristic approach to compute\nbias scores for text-only, image-only, and multimodal embeddings. A framework\nis developed to classify bias interactions as amplification (multimodal bias\nexceeds both unimodal biases), mitigation (multimodal bias is lower than both),\nand neutrality (multimodal bias lies between unimodal biases), with\nproportional analyzes conducted to identify the dominant mode and dynamics in\nthese interactions. The findings highlight that amplification (22\\%) occurs\nwhen text and image biases are comparable, while mitigation (11\\%) arises under\nthe dominance of text bias, highlighting the stabilizing role of image bias.\nNeutral interactions (67\\%) are related to a higher text bias without\ndivergence. Conditional probabilities highlight the text's dominance in\nmitigation and mixed contributions in neutral and amplification cases,\nunderscoring complex modality interplay. In doing so, the study encourages the\nuse of this heuristic, systemic, and interpretable framework to analyze\nmultimodal bias interactions, providing insight into how intermodal biases\ndynamically interact, with practical applications for multimodal modeling and\ntransferability to context-based datasets, all essential for developing fair\nand equitable AI models.\n", "date": "2024-12-23", "categories": []}
{"id": "2403.14633", "title": "Born With a Silver Spoon? Investigating Socioeconomic Bias in Large\n  Language Models", "abstract": "  Socioeconomic bias in society exacerbates disparities, influencing access to\nopportunities and resources based on individuals' economic and social\nbackgrounds. This pervasive issue perpetuates systemic inequalities, hindering\nthe pursuit of inclusive progress as a society. In this paper, we investigate\nthe presence of socioeconomic bias, if any, in large language models. To this\nend, we introduce a novel dataset SilverSpoon, consisting of 3000 samples that\nillustrate hypothetical scenarios that involve underprivileged people\nperforming ethically ambiguous actions due to their circumstances, and ask\nwhether the action is ethically justified. Further, this dataset has a\ndual-labeling scheme and has been annotated by people belonging to both ends of\nthe socioeconomic spectrum. Using SilverSpoon, we evaluate the degree of\nsocioeconomic bias expressed in large language models and the variation of this\ndegree as a function of model size. We also perform qualitative analysis to\nanalyze the nature of this bias. Our analysis reveals that while humans\ndisagree on which situations require empathy toward the underprivileged, most\nlarge language models are unable to empathize with the socioeconomically\nunderprivileged regardless of the situation. To foster further research in this\ndomain, we make SilverSpoon and our evaluation harness publicly available.\n", "date": "2024-02-16", "categories": []}
{"id": "2405.15443", "title": "Fairness-Accuracy Trade-Offs: A Causal Perspective", "abstract": "  Systems based on machine learning may exhibit discriminatory behavior based\non sensitive characteristics such as gender, sex, religion, or race. In light\nof this, various notions of fairness and methods to quantify discrimination\nwere proposed, leading to the development of numerous approaches for\nconstructing fair predictors. At the same time, imposing fairness constraints\nmay decrease the utility of the decision-maker, highlighting a tension between\nfairness and utility. This tension is also recognized in legal frameworks, for\ninstance in the disparate impact doctrine of Title VII of the Civil Rights Act\nof 1964 -- in which specific attention is given to considerations of business\nnecessity -- possibly allowing the usage of proxy variables associated with the\nsensitive attribute in case a high-enough utility cannot be achieved without\nthem. In this work, we analyze the tension between fairness and accuracy from a\ncausal lens for the first time. We introduce the notion of a path-specific\nexcess loss (PSEL) that captures how much the predictor's loss increases when a\ncausal fairness constraint is enforced. We then show that the total excess loss\n(TEL), defined as the difference between the loss of predictor fair along all\ncausal pathways vs. an unconstrained predictor, can be decomposed into a sum of\nmore local PSELs. At the same time, enforcing a causal constraint often reduces\nthe disparity between demographic groups. Thus, we introduce a quantity that\nsummarizes the fairness-utility trade-off, called the causal fairness/utility\nratio, defined as the ratio of the reduction in discrimination vs. the excess\nloss from constraining a causal pathway. This quantity is suitable for\ncomparing the fairness-utility trade-off across causal pathways. Finally, as\nour approach requires causally-constrained fair predictors, we introduce a new\nneural approach for causally-constrained fair learning.\n", "date": "2024-05-24", "categories": []}
{"id": "2411.01757", "title": "Mitigating Spurious Correlations via Disagreement Probability", "abstract": "  Models trained with empirical risk minimization (ERM) are prone to be biased\ntowards spurious correlations between target labels and bias attributes, which\nleads to poor performance on data groups lacking spurious correlations. It is\nparticularly challenging to address this problem when access to bias labels is\nnot permitted. To mitigate the effect of spurious correlations without bias\nlabels, we first introduce a novel training objective designed to robustly\nenhance model performance across all data samples, irrespective of the presence\nof spurious correlations. From this objective, we then derive a debiasing\nmethod, Disagreement Probability based Resampling for debiasing (DPR), which\ndoes not require bias labels. DPR leverages the disagreement between the target\nlabel and the prediction of a biased model to identify bias-conflicting\nsamples-those without spurious correlations-and upsamples them according to the\ndisagreement probability. Empirical evaluations on multiple benchmarks\ndemonstrate that DPR achieves state-of-the-art performance over existing\nbaselines that do not use bias labels. Furthermore, we provide a theoretical\nanalysis that details how DPR reduces dependency on spurious correlations.\n", "date": "2024-11-04", "categories": []}
{"id": "2412.15920", "title": "Data Preparation for Fairness-Performance Trade-Offs: A\n  Practitioner-Friendly Alternative?", "abstract": "  As machine learning (ML) systems are increasingly adopted across industries,\naddressing fairness and bias has become essential. While many solutions focus\non ethical challenges in ML, recent studies highlight that data itself is a\nmajor source of bias. Pre-processing techniques, which mitigate bias before\ntraining, are effective but may impact model performance and pose integration\ndifficulties. In contrast, fairness-aware Data Preparation practices are both\nfamiliar to practitioners and easier to implement, providing a more accessible\napproach to reducing bias. Objective. This registered report proposes an\nempirical evaluation of how optimally selected fairness-aware practices,\napplied in early ML lifecycle stages, can enhance both fairness and\nperformance, potentially outperforming standard pre-processing bias mitigation\nmethods. Method. To this end, we will introduce FATE, an optimization technique\nfor selecting 'Data Preparation' pipelines that optimize fairness and\nperformance. Using FATE, we will analyze the fairness-performance trade-off,\ncomparing pipelines selected by FATE with results by pre-processing bias\nmitigation techniques.\n", "date": "2024-12-20", "categories": []}
{"id": "2404.01349", "title": "Fairness in Large Language Models: A Taxonomic Survey", "abstract": "  Large Language Models (LLMs) have demonstrated remarkable success across\nvarious domains. However, despite their promising performance in numerous\nreal-world applications, most of these algorithms lack fairness considerations.\nConsequently, they may lead to discriminatory outcomes against certain\ncommunities, particularly marginalized populations, prompting extensive study\nin fair LLMs. On the other hand, fairness in LLMs, in contrast to fairness in\ntraditional machine learning, entails exclusive backgrounds, taxonomies, and\nfulfillment techniques. To this end, this survey presents a comprehensive\noverview of recent advances in the existing literature concerning fair LLMs.\nSpecifically, a brief introduction to LLMs is provided, followed by an analysis\nof factors contributing to bias in LLMs. Additionally, the concept of fairness\nin LLMs is discussed categorically, summarizing metrics for evaluating bias in\nLLMs and existing algorithms for promoting fairness. Furthermore, resources for\nevaluating bias in LLMs, including toolkits and datasets, are summarized.\nFinally, existing research challenges and open questions are discussed.\n", "date": "2024-03-31", "categories": []}
{"id": "2406.16606", "title": "Cherry on the Cake: Fairness is NOT an Optimization Problem", "abstract": "  In Fair AI literature, the practice of maliciously creating unfair models\nthat nevertheless satisfy fairness constraints is known as \"cherry-picking\". A\ncherry-picking model is a model that makes mistakes on purpose, selecting bad\nindividuals from a minority class instead of better candidates from the same\nminority. The model literally cherry-picks whom to select to superficially meet\nthe fairness constraints while making minimal changes to the unfair model. This\npractice has been described as \"blatantly unfair\" and has a negative impact on\nalready marginalized communities, undermining the intended purpose of fairness\nmeasures specifically designed to protect these communities. A common\nassumption is that cherry-picking arises solely from malicious intent and that\nmodels designed only to optimize fairness metrics would avoid this behavior. We\nshow that this is not the case: models optimized to minimize fairness metrics\nwhile maximizing performance are often forced to cherry-pick to some degree. In\nother words, cherry-picking might be an inevitable outcome of the optimization\nprocess itself. To demonstrate this, we use tools from fair cake-cutting, a\nmathematical subfield that studies the problem of fairly dividing a resource,\nreferred to as the \"cake,\" among a number of participants. This concept is\nconnected to supervised multi-label classification: any dataset can be thought\nof as a cake that needs to be distributed among different labels, and the model\nis the function that divides the cake. We adapt these classical results for\nmachine learning and demonstrate how this connection can be prolifically used\nfor fairness and classification in general.\n", "date": "2024-06-24", "categories": []}
{"id": "2412.14724", "title": "FROC: Building Fair ROC from a Trained Classifier", "abstract": "  This paper considers the problem of fair probabilistic binary classification\nwith binary protected groups. The classifier assigns scores, and a practitioner\npredicts labels using a certain cut-off threshold based on the desired\ntrade-off between false positives vs. false negatives. It derives these\nthresholds from the ROC of the classifier. The resultant classifier may be\nunfair to one of the two protected groups in the dataset. It is desirable that\nno matter what threshold the practitioner uses, the classifier should be fair\nto both the protected groups; that is, the $\\mathcal{L}_p$ norm between FPRs\nand TPRs of both the protected groups should be at most $\\varepsilon$. We call\nsuch fairness on ROCs of both the protected attributes\n$\\varepsilon_p$-Equalized ROC. Given a classifier not satisfying\n$\\varepsilon_1$-Equalized ROC, we aim to design a post-processing method to\ntransform the given (potentially unfair) classifier's output (score) to a\nsuitable randomized yet fair classifier. That is, the resultant classifier must\nsatisfy $\\varepsilon_1$-Equalized ROC. First, we introduce a threshold query\nmodel on the ROC curves for each protected group. The resulting classifier is\nbound to face a reduction in AUC. With the proposed query model, we provide a\nrigorous theoretical analysis of the minimal AUC loss to achieve\n$\\varepsilon_1$-Equalized ROC. To achieve this, we design a linear time\nalgorithm, namely \\texttt{FROC}, to transform a given classifier's output to a\nprobabilistic classifier that satisfies $\\varepsilon_1$-Equalized ROC. We prove\nthat under certain theoretical conditions, \\texttt{FROC}\\ achieves the\ntheoretical optimal guarantees. We also study the performance of our\n\\texttt{FROC}\\ on multiple real-world datasets with many trained classifiers.\n", "date": "2024-12-19", "categories": []}
{"id": "2401.13391", "title": "Reranking individuals: The effect of fair classification within-groups", "abstract": "  Artificial Intelligence (AI) finds widespread application across various\ndomains, but it sparks concerns about fairness in its deployment. The\nprevailing discourse in classification often emphasizes outcome-based metrics\ncomparing sensitive subgroups without a nuanced consideration of the\ndifferential impacts within subgroups. Bias mitigation techniques not only\naffect the ranking of pairs of instances across sensitive groups, but often\nalso significantly affect the ranking of instances within these groups. Such\nchanges are hard to explain and raise concerns regarding the validity of the\nintervention. Unfortunately, these effects remain under the radar in the\naccuracy-fairness evaluation framework that is usually applied. Additionally,\nwe illustrate the effect of several popular bias mitigation methods, and how\ntheir output often does not reflect real-world scenarios.\n", "date": "2024-01-24", "categories": []}
{"id": "2412.09896", "title": "Analyzing Fairness of Classification Machine Learning Model with\n  Structured Dataset", "abstract": "  Machine learning (ML) algorithms have become integral to decision making in\nvarious domains, including healthcare, finance, education, and law enforcement.\nHowever, concerns about fairness and bias in these systems pose significant\nethical and social challenges. This study investigates the fairness of ML\nmodels applied to structured datasets in classification tasks, highlighting the\npotential for biased predictions to perpetuate systemic inequalities. A\npublicly available dataset from Kaggle was selected for analysis, offering a\nrealistic scenario for evaluating fairness in machine learning workflows.\n  To assess and mitigate biases, three prominent fairness libraries; Fairlearn\nby Microsoft, AIF360 by IBM, and the What If Tool by Google were employed.\nThese libraries provide robust frameworks for analyzing fairness, offering\ntools to evaluate metrics, visualize results, and implement bias mitigation\nstrategies. The research aims to assess the extent of bias in the ML models,\ncompare the effectiveness of these libraries, and derive actionable insights\nfor practitioners.\n  The findings reveal that each library has unique strengths and limitations in\nfairness evaluation and mitigation. By systematically comparing their\ncapabilities, this study contributes to the growing field of ML fairness by\nproviding practical guidance for integrating fairness tools into real world\napplications. These insights are intended to support the development of more\nequitable machine learning systems.\n", "date": "2024-12-13", "categories": []}
{"id": "2412.12165", "title": "Multimodal Approaches to Fair Image Classification: An Ethical\n  Perspective", "abstract": "  In the rapidly advancing field of artificial intelligence, machine perception\nis becoming paramount to achieving increased performance. Image classification\nsystems are becoming increasingly integral to various applications, ranging\nfrom medical diagnostics to image generation; however, these systems often\nexhibit harmful biases that can lead to unfair and discriminatory outcomes.\nMachine Learning systems that depend on a single data modality, i.e. only\nimages or only text, can exaggerate hidden biases present in the training data,\nif the data is not carefully balanced and filtered. Even so, these models can\nstill harm underrepresented populations when used in improper contexts, such as\nwhen government agencies reinforce racial bias using predictive policing. This\nthesis explores the intersection of technology and ethics in the development of\nfair image classification models. Specifically, I focus on improving fairness\nand methods of using multiple modalities to combat harmful demographic bias.\nIntegrating multimodal approaches, which combine visual data with additional\nmodalities such as text and metadata, allows this work to enhance the fairness\nand accuracy of image classification systems. The study critically examines\nexisting biases in image datasets and classification algorithms, proposes\ninnovative methods for mitigating these biases, and evaluates the ethical\nimplications of deploying such systems in real-world scenarios. Through\ncomprehensive experimentation and analysis, the thesis demonstrates how\nmultimodal techniques can contribute to more equitable and ethical AI\nsolutions, ultimately advocating for responsible AI practices that prioritize\nfairness.\n", "date": "2024-12-11", "categories": []}
{"id": "2412.10923", "title": "Linear Programming based Approximation to Individually Fair k-Clustering\n  with Outliers", "abstract": "  Individual fairness guarantees are often desirable properties to have, but\nthey become hard to formalize when the dataset contains outliers. Here, we\ninvestigate the problem of developing an individually fair $k$-means clustering\nalgorithm for datasets that contain outliers. That is, given $n$ points and $k$\ncenters, we want that for each point which is not an outlier, there must be a\ncenter within the $\\frac{n}{k}$ nearest neighbours of the given point. While a\nfew of the recent works have looked into individually fair clustering, this is\nthe first work that explores this problem in the presence of outliers for\n$k$-means clustering.\n  For this purpose, we define and solve a linear program (LP) that helps us\nidentify the outliers. We exclude these outliers from the dataset and apply a\nrounding algorithm that computes the $k$ centers, such that the fairness\nconstraint of the remaining points is satisfied. We also provide theoretical\nguarantees that our method leads to a guaranteed approximation of the fair\nradius as well as the clustering cost. We also demonstrate our techniques\nempirically on real-world datasets.\n", "date": "2024-12-14", "categories": []}
{"id": "2412.11072", "title": "Navigating Towards Fairness with Data Selection", "abstract": "  Machine learning algorithms often struggle to eliminate inherent data biases,\nparticularly those arising from unreliable labels, which poses a significant\nchallenge in ensuring fairness. Existing fairness techniques that address label\nbias typically involve modifying models and intervening in the training\nprocess, but these lack flexibility for large-scale datasets. To address this\nlimitation, we introduce a data selection method designed to efficiently and\nflexibly mitigate label bias, tailored to more practical needs. Our approach\nutilizes a zero-shot predictor as a proxy model that simulates training on a\nclean holdout set. This strategy, supported by peer predictions, ensures the\nfairness of the proxy model and eliminates the need for an additional holdout\nset, which is a common requirement in previous methods. Without altering the\nclassifier's architecture, our modality-agnostic method effectively selects\nappropriate training data and has proven efficient and effective in handling\nlabel bias and improving fairness across diverse datasets in experimental\nevaluations.\n", "date": "2024-12-15", "categories": []}
{"id": "2412.11414", "title": "Biased or Flawed? Mitigating Stereotypes in Generative Language Models\n  by Addressing Task-Specific Flaws", "abstract": "  Recent studies have shown that generative language models often reflect and\namplify societal biases in their outputs. However, these studies frequently\nconflate observed biases with other task-specific shortcomings, such as\ncomprehension failure. For example, when a model misinterprets a text and\nproduces a response that reinforces a stereotype, it becomes difficult to\ndetermine whether the issue arises from inherent bias or from a\nmisunderstanding of the given content. In this paper, we conduct a\nmulti-faceted evaluation that distinctly disentangles bias from flaws within\nthe reading comprehension task. We propose a targeted stereotype mitigation\nframework that implicitly mitigates observed stereotypes in generative models\nthrough instruction-tuning on general-purpose datasets. We reduce stereotypical\noutputs by over 60% across multiple dimensions -- including nationality, age,\ngender, disability, and physical appearance -- by addressing\ncomprehension-based failures, and without relying on explicit debiasing\ntechniques. We evaluate several state-of-the-art generative models to\ndemonstrate the effectiveness of our approach while maintaining the overall\nutility. Our findings highlight the need to critically disentangle the concept\nof `bias' from other types of errors to build more targeted and effective\nmitigation strategies. CONTENT WARNING: Some examples contain offensive\nstereotypes.\n", "date": "2024-12-16", "categories": []}
{"id": "2412.11951", "title": "The Impact of Generalization Techniques on the Interplay Among Privacy,\n  Utility, and Fairness in Image Classification", "abstract": "  This study investigates the trade-offs between fairness, privacy, and utility\nin image classification using machine learning (ML). Recent research suggests\nthat generalization techniques can improve the balance between privacy and\nutility. One focus of this work is sharpness-aware training (SAT) and its\nintegration with differential privacy (DP-SAT) to further improve this balance.\nAdditionally, we examine fairness in both private and non-private learning\nmodels trained on datasets with synthetic and real-world biases. We also\nmeasure the privacy risks involved in these scenarios by performing membership\ninference attacks (MIAs) and explore the consequences of eliminating\nhigh-privacy risk samples, termed outliers. Moreover, we introduce a new\nmetric, named \\emph{harmonic score}, which combines accuracy, privacy, and\nfairness into a single measure.\n  Through empirical analysis using generalization techniques, we achieve an\naccuracy of 81.11\\% under $(8, 10^{-5})$-DP on CIFAR-10, surpassing the 79.5\\%\nreported by De et al. (2022). Moreover, our experiments show that memorization\nof training samples can begin before the overfitting point, and generalization\ntechniques do not guarantee the prevention of this memorization. Our analysis\nof synthetic biases shows that generalization techniques can amplify model bias\nin both private and non-private models. Additionally, our results indicate that\nincreased bias in training data leads to reduced accuracy, greater\nvulnerability to privacy attacks, and higher model bias. We validate these\nfindings with the CelebA dataset, demonstrating that similar trends persist\nwith real-world attribute imbalances. Finally, our experiments show that\nremoving outlier data decreases accuracy and further amplifies model bias.\n", "date": "2024-12-16", "categories": []}
{"id": "2412.11994", "title": "Fairness Shields: Safeguarding against Biased Decision Makers", "abstract": "  As AI-based decision-makers increasingly influence human lives, it is a\ngrowing concern that their decisions are often unfair or biased with respect to\npeople's sensitive attributes, such as gender and race. Most existing bias\nprevention measures provide probabilistic fairness guarantees in the long run,\nand it is possible that the decisions are biased on specific instances of short\ndecision sequences. We introduce fairness shielding, where a symbolic\ndecision-maker -- the fairness shield -- continuously monitors the sequence of\ndecisions of another deployed black-box decision-maker, and makes interventions\nso that a given fairness criterion is met while the total intervention costs\nare minimized. We present four different algorithms for computing fairness\nshields, among which one guarantees fairness over fixed horizons, and three\nguarantee fairness periodically after fixed intervals. Given a distribution\nover future decisions and their intervention costs, our algorithms solve\ndifferent instances of bounded-horizon optimal control problems with different\nlevels of computational costs and optimality guarantees. Our empirical\nevaluation demonstrates the effectiveness of these shields in ensuring fairness\nwhile maintaining cost efficiency across various scenarios.\n", "date": "2024-12-16", "categories": []}
{"id": "2412.09947", "title": "Towards Fair Graph Neural Networks via Graph Counterfactual without\n  Sensitive Attributes", "abstract": "  Graph-structured data is ubiquitous in today's connected world, driving\nextensive research in graph analysis. Graph Neural Networks (GNNs) have shown\ngreat success in this field, leading to growing interest in developing fair\nGNNs for critical applications. However, most existing fair GNNs focus on\nstatistical fairness notions, which may be insufficient when dealing with\nstatistical anomalies. Hence, motivated by causal theory, there has been\ngrowing attention to mitigating root causes of unfairness utilizing graph\ncounterfactuals. Unfortunately, existing methods for generating graph\ncounterfactuals invariably require the sensitive attribute. Nevertheless, in\nmany real-world applications, it is usually infeasible to obtain sensitive\nattributes due to privacy or legal issues, which challenge existing methods. In\nthis paper, we propose a framework named Fairwos (improving Fairness without\nsensitive attributes). In particular, we first propose a mechanism to generate\npseudo-sensitive attributes to remedy the problem of missing sensitive\nattributes, and then design a strategy for finding graph counterfactuals from\nthe real dataset. To train fair GNNs, we propose a method to ensure that the\nembeddings from the original data are consistent with those from the graph\ncounterfactuals, and dynamically adjust the weight of each pseudo-sensitive\nattribute to balance its contribution to fairness and utility. Furthermore, we\ntheoretically demonstrate that minimizing the relation between these\npseudo-sensitive attributes and the prediction can enable the fairness of GNNs.\nExperimental results on six real-world datasets show that our approach\noutperforms state-of-the-art methods in balancing utility and fairness.\n", "date": "2024-12-13", "categories": []}
{"id": "2412.09254", "title": "When Can Memorization Improve Fairness?", "abstract": "  We study to which extent additive fairness metrics (statistical parity, equal\nopportunity and equalized odds) can be influenced in a multi-class\nclassification problem by memorizing a subset of the population. We give\nexplicit expressions for the bias resulting from memorization in terms of the\nlabel and group membership distribution of the memorized dataset and the\nclassifier bias on the unmemorized dataset. We also characterize the memorized\ndatasets that eliminate the bias for all three metrics considered. Finally we\nprovide upper and lower bounds on the total probability mass in the memorized\ndataset that is necessary for the complete elimination of these biases.\n", "date": "2024-12-12", "categories": []}
{"id": "2412.09439", "title": "Towards Robust and Fair Vision Learning in Open-World Environments", "abstract": "  The dissertation presents four key contributions toward fairness and\nrobustness in vision learning. First, to address the problem of large-scale\ndata requirements, the dissertation presents a novel Fairness Domain Adaptation\napproach derived from two major novel research findings of Bijective Maximum\nLikelihood and Fairness Adaptation Learning. Second, to enable the capability\nof open-world modeling of vision learning, this dissertation presents a novel\nOpen-world Fairness Continual Learning Framework. The success of this research\ndirection is the result of two research lines, i.e., Fairness Continual\nLearning and Open-world Continual Learning. Third, since visual data are often\ncaptured from multiple camera views, robust vision learning methods should be\ncapable of modeling invariant features across views. To achieve this desired\ngoal, the research in this thesis will present a novel Geometry-based\nCross-view Adaptation framework to learn robust feature representations across\nviews. Finally, with the recent increase in large-scale videos and multimodal\ndata, understanding the feature representations and improving the robustness of\nlarge-scale visual foundation models is critical. Therefore, this thesis will\npresent novel Transformer-based approaches to improve the robust feature\nrepresentations against multimodal and temporal data. Then, a novel Domain\nGeneralization Approach will be presented to improve the robustness of visual\nfoundation models. The research's theoretical analysis and experimental results\nhave shown the effectiveness of the proposed approaches, demonstrating their\nsuperior performance compared to prior studies. The contributions in this\ndissertation have advanced the fairness and robustness of machine vision\nlearning.\n", "date": "2024-12-12", "categories": []}
{"id": "2412.08167", "title": "Diversity Drives Fairness: Ensemble of Higher Order Mutants for\n  Intersectional Fairness of Machine Learning Software", "abstract": "  Intersectional fairness is a critical requirement for Machine Learning (ML)\nsoftware, demanding fairness across subgroups defined by multiple protected\nattributes. This paper introduces FairHOME, a novel ensemble approach using\nhigher order mutation of inputs to enhance intersectional fairness of ML\nsoftware during the inference phase. Inspired by social science theories\nhighlighting the benefits of diversity, FairHOME generates mutants representing\ndiverse subgroups for each input instance, thus broadening the array of\nperspectives to foster a fairer decision-making process. Unlike conventional\nensemble methods that combine predictions made by different models, FairHOME\ncombines predictions for the original input and its mutants, all generated by\nthe same ML model, to reach a final decision. Notably, FairHOME is even\napplicable to deployed ML software as it bypasses the need for training new\nmodels. We extensively evaluate FairHOME against seven state-of-the-art\nfairness improvement methods across 24 decision-making tasks using widely\nadopted metrics. FairHOME consistently outperforms existing methods across all\nmetrics considered. On average, it enhances intersectional fairness by 47.5%,\nsurpassing the currently best-performing method by 9.6 percentage points.\n", "date": "2024-12-11", "categories": []}
{"id": "2412.08480", "title": "InvDiff: Invariant Guidance for Bias Mitigation in Diffusion Models", "abstract": "  As one of the most successful generative models, diffusion models have\ndemonstrated remarkable efficacy in synthesizing high-quality images. These\nmodels learn the underlying high-dimensional data distribution in an\nunsupervised manner. Despite their success, diffusion models are highly\ndata-driven and prone to inheriting the imbalances and biases present in\nreal-world data. Some studies have attempted to address these issues by\ndesigning text prompts for known biases or using bias labels to construct\nunbiased data. While these methods have shown improved results, real-world\nscenarios often contain various unknown biases, and obtaining bias labels is\nparticularly challenging. In this paper, we emphasize the necessity of\nmitigating bias in pre-trained diffusion models without relying on auxiliary\nbias annotations. To tackle this problem, we propose a framework, InvDiff,\nwhich aims to learn invariant semantic information for diffusion guidance.\nSpecifically, we propose identifying underlying biases in the training data and\ndesigning a novel debiasing training objective. Then, we employ a lightweight\ntrainable module that automatically preserves invariant semantic information\nand uses it to guide the diffusion model's sampling process toward unbiased\noutcomes simultaneously. Notably, we only need to learn a small number of\nparameters in the lightweight learnable module without altering the pre-trained\ndiffusion model. Furthermore, we provide a theoretical guarantee that the\nimplementation of InvDiff is equivalent to reducing the error upper bound of\ngeneralization. Extensive experimental results on three publicly available\nbenchmarks demonstrate that InvDiff effectively reduces biases while\nmaintaining the quality of image generation. Our code is available at\nhttps://github.com/Hundredl/InvDiff.\n", "date": "2024-12-11", "categories": []}
{"id": "2411.18122", "title": "Using Machine Bias To Measure Human Bias", "abstract": "  Biased human decisions have consequential impacts across various domains,\nyielding unfair treatment of individuals and resulting in suboptimal outcomes\nfor organizations and society. In recognition of this fact, organizations\nregularly design and deploy interventions aimed at mitigating these biases.\nHowever, measuring human decision biases remains an important but elusive task.\nOrganizations are frequently concerned with mistaken decisions\ndisproportionately affecting one group. In practice, however, this is typically\nnot possible to assess due to the scarcity of a gold standard: a label that\nindicates what the correct decision would have been. In this work, we propose a\nmachine learning-based framework to assess bias in human-generated decisions\nwhen gold standard labels are scarce. We provide theoretical guarantees and\nempirical evidence demonstrating the superiority of our method over existing\nalternatives. This proposed methodology establishes a foundation for\ntransparency in human decision-making, carrying substantial implications for\nmanagerial duties, and offering potential for alleviating algorithmic biases\nwhen human decisions are used as labels to train algorithms.\n", "date": "2024-11-27", "categories": []}
{"id": "2311.18291", "title": "TLDR: Text Based Last-layer Retraining for Debiasing Image Classifiers", "abstract": "  An image classifier may depend on incidental features stemming from a strong\ncorrelation between the feature and the classification target in the training\ndataset. Recently, Last Layer Retraining (LLR) with group-balanced datasets is\nshown to be efficient in mitigating the spurious correlation of classifiers.\nHowever, the acquisition of image-based group-balanced datasets is costly,\nwhich hinders the general applicability of the LLR method. In this work, we\npropose to perform LLR based on text datasets built with large language models\nto debias a general image classifier. To that end, we demonstrate that text can\ngenerally be a proxy for its corresponding image beyond the image-text joint\nembedding space, which is achieved with a linear projector that ensures\northogonality between its weight and the modality gap of the joint embedding\nspace. In addition, we propose a systematic validation procedure that checks\nwhether the generated words are compatible with the embedding space of CLIP and\nthe image classifier, which is shown to be effective for improving debiasing\nperformance. We dub these procedures as TLDR (Text-based Last layer retraining\nfor Debiasing image classifieRs) and show our method achieves the performance\nthat is competitive with the LLR methods that require group-balanced image\ndataset for retraining. Furthermore, TLDR outperforms other baselines that\ninvolve training the last layer without any group annotated dataset. Codes:\nhttps://github.com/beotborry/TLDR\n", "date": "2023-11-30", "categories": []}
{"id": "2403.14713", "title": "Auditing Fairness under Unobserved Confounding", "abstract": "  Many definitions of fairness or inequity involve unobservable causal\nquantities that cannot be directly estimated without strong assumptions. For\ninstance, it is particularly difficult to estimate notions of fairness that\nrely on hard-to-measure concepts such as risk (e.g., quantifying whether\npatients at the same risk level have equal probability of treatment, regardless\nof group membership). Such measurements of risk can be accurately obtained when\nno unobserved confounders have jointly influenced past decisions and outcomes.\nHowever, in the real world, this assumption rarely holds. In this paper, we\nshow that, surprisingly, one can still compute meaningful bounds on treatment\nrates for high-risk individuals (i.e., conditional on their true,\n\\textit{unobserved} negative outcome), even when entirely eliminating or\nrelaxing the assumption that we observe all relevant risk factors used by\ndecision makers. We use the fact that in many real-world settings (e.g., the\nrelease of a new treatment) we have data from prior to any allocation to derive\nunbiased estimates of risk. This result enables us to audit unfair outcomes of\nexisting decision-making systems in a principled manner. We demonstrate the\neffectiveness of our framework with a real-world study of Paxlovid allocation,\nprovably identifying that observed racial inequity cannot be explained by\nunobserved confounders of the same strength as important observed covariates.\n", "date": "2024-03-18", "categories": []}
{"id": "2405.09273", "title": "Fair Generalized Linear Mixed Models", "abstract": "  When using machine learning for automated prediction, it is important to\naccount for fairness in the prediction. Fairness in machine learning aims to\nensure that biases in the data and model inaccuracies do not lead to\ndiscriminatory decisions. E.g., predictions from fair machine learning models\nshould not discriminate against sensitive variables such as sexual orientation\nand ethnicity. The training data often in obtained from social surveys. In\nsocial surveys, oftentimes the data collection process is a strata sampling,\ne.g. due to cost restrictions. In strata samples, the assumption of\nindependence between the observation is not fulfilled. Hence, if the machine\nlearning models do not account for the strata correlations, the results may be\nbiased. Especially high is the bias in cases where the strata assignment is\ncorrelated to the variable of interest. We present in this paper an algorithm\nthat can handle both problems simultaneously, and we demonstrate the impact of\nstratified sampling on the quality of fair machine learning predictions in a\nreproducible simulation study.\n", "date": "2024-05-15", "categories": []}
{"id": "2407.17543", "title": "Dataset Distribution Impacts Model Fairness: Single vs. Multi-Task\n  Learning", "abstract": "  The influence of bias in datasets on the fairness of model predictions is a\ntopic of ongoing research in various fields. We evaluate the performance of\nskin lesion classification using ResNet-based CNNs, focusing on patient sex\nvariations in training data and three different learning strategies. We present\na linear programming method for generating datasets with varying patient sex\nand class labels, taking into account the correlations between these variables.\nWe evaluated the model performance using three different learning strategies: a\nsingle-task model, a reinforcing multi-task model, and an adversarial learning\nscheme. Our observations include: 1) sex-specific training data yields better\nresults, 2) single-task models exhibit sex bias, 3) the reinforcement approach\ndoes not remove sex bias, 4) the adversarial model eliminates sex bias in cases\ninvolving only female patients, and 5) datasets that include male patients\nenhance model performance for the male subgroup, even when female patients are\nthe majority. To generalise these findings, in future research, we will examine\nmore demographic attributes, like age, and other possibly confounding factors,\nsuch as skin colour and artefacts in the skin lesions. We make all data and\nmodels available on GitHub.\n", "date": "2024-07-24", "categories": []}
{"id": "2411.01115", "title": "Relax and Merge: A Simple Yet Effective Framework for Solving Fair\n  $k$-Means and $k$-sparse Wasserstein Barycenter Problems", "abstract": "  The fairness of clustering algorithms has gained widespread attention across\nvarious areas, including machine learning, In this paper, we study fair\n$k$-means clustering in Euclidean space. Given a dataset comprising several\ngroups, the fairness constraint requires that each cluster should contain a\nproportion of points from each group within specified lower and upper bounds.\nDue to these fairness constraints, determining the optimal locations of $k$\ncenters is a quite challenging task. We propose a novel ``Relax and Merge''\nframework that returns a $(1+4\\rho + O(\\epsilon))$-approximate solution, where\n$\\rho$ is the approximate ratio of an off-the-shelf vanilla $k$-means algorithm\nand $O(\\epsilon)$ can be an arbitrarily small positive number. If equipped with\na PTAS of $k$-means, our solution can achieve an approximation ratio of\n$(5+O(\\epsilon))$ with only a slight violation of the fairness constraints,\nwhich improves the current state-of-the-art approximation guarantee.\nFurthermore, using our framework, we can also obtain a $(1+4\\rho\n+O(\\epsilon))$-approximate solution for the $k$-sparse Wasserstein Barycenter\nproblem, which is a fundamental optimization problem in the field of optimal\ntransport, and a $(2+6\\rho)$-approximate solution for the strictly fair\n$k$-means clustering with no violation, both of which are better than the\ncurrent state-of-the-art methods. In addition, the empirical results\ndemonstrate that our proposed algorithm can significantly outperform baseline\napproaches in terms of clustering cost.\n", "date": "2024-11-02", "categories": []}
{"id": "2412.01585", "title": "FairML: A Julia Package for Fair Classification", "abstract": "  In this paper, we propose FairML.jl, a Julia package providing a framework\nfor fair classification in machine learning. In this framework, the fair\nlearning process is divided into three stages. Each stage aims to reduce\nunfairness, such as disparate impact and disparate mistreatment, in the final\nprediction. For the preprocessing stage, we present a resampling method that\naddresses unfairness coming from data imbalances. The in-processing phase\nconsist of a classification method. This can be either one coming from the\nMLJ.jl package, or a user defined one. For this phase, we incorporate fair ML\nmethods that can handle unfairness to a certain degree through their\noptimization process. In the post-processing, we discuss the choice of the\ncut-off value for fair prediction. With simulations, we show the performance of\nthe single phases and their combinations.\n", "date": "2024-12-02", "categories": []}
{"id": "2412.06063", "title": "On Socially Fair Low-Rank Approximation and Column Subset Selection", "abstract": "  Low-rank approximation and column subset selection are two fundamental and\nrelated problems that are applied across a wealth of machine learning\napplications. In this paper, we study the question of socially fair low-rank\napproximation and socially fair column subset selection, where the goal is to\nminimize the loss over all sub-populations of the data. We show that\nsurprisingly, even constant-factor approximation to fair low-rank approximation\nrequires exponential time under certain standard complexity hypotheses. On the\npositive side, we give an algorithm for fair low-rank approximation that, for a\nconstant number of groups and constant-factor accuracy, runs in\n$2^{\\text{poly}(k)}$ time rather than the na\\\"{i}ve $n^{\\text{poly}(k)}$, which\nis a substantial improvement when the dataset has a large number $n$ of\nobservations. We then show that there exist bicriteria approximation algorithms\nfor fair low-rank approximation and fair column subset selection that run in\npolynomial time.\n", "date": "2024-12-08", "categories": []}
{"id": "2407.00983", "title": "FairMedFM: Fairness Benchmarking for Medical Imaging Foundation Models", "abstract": "  The advent of foundation models (FMs) in healthcare offers unprecedented\nopportunities to enhance medical diagnostics through automated classification\nand segmentation tasks. However, these models also raise significant concerns\nabout their fairness, especially when applied to diverse and underrepresented\npopulations in healthcare applications. Currently, there is a lack of\ncomprehensive benchmarks, standardized pipelines, and easily adaptable\nlibraries to evaluate and understand the fairness performance of FMs in medical\nimaging, leading to considerable challenges in formulating and implementing\nsolutions that ensure equitable outcomes across diverse patient populations. To\nfill this gap, we introduce FairMedFM, a fairness benchmark for FM research in\nmedical imaging.FairMedFM integrates with 17 popular medical imaging datasets,\nencompassing different modalities, dimensionalities, and sensitive attributes.\nIt explores 20 widely used FMs, with various usages such as zero-shot learning,\nlinear probing, parameter-efficient fine-tuning, and prompting in various\ndownstream tasks -- classification and segmentation. Our exhaustive analysis\nevaluates the fairness performance over different evaluation metrics from\nmultiple perspectives, revealing the existence of bias, varied utility-fairness\ntrade-offs on different FMs, consistent disparities on the same datasets\nregardless FMs, and limited effectiveness of existing unfairness mitigation\nmethods. Checkout FairMedFM's project page and open-sourced codebase, which\nsupports extendible functionalities and applications as well as inclusive for\nstudies on FMs in medical imaging over the long term.\n", "date": "2024-07-01", "categories": []}
{"id": "2412.04642", "title": "Improving LLM Group Fairness on Tabular Data via In-Context Learning", "abstract": "  Large language models (LLMs) have been shown to be effective on tabular\nprediction tasks in the low-data regime, leveraging their internal knowledge\nand ability to learn from instructions and examples. However, LLMs can fail to\ngenerate predictions that satisfy group fairness, that is, produce equitable\noutcomes across groups. Critically, conventional debiasing approaches for\nnatural language tasks do not directly translate to mitigating group unfairness\nin tabular settings. In this work, we systematically investigate four empirical\napproaches to improve group fairness of LLM predictions on tabular datasets,\nincluding fair prompt optimization, soft prompt tuning, strategic selection of\nfew-shot examples, and self-refining predictions via chain-of-thought\nreasoning. Through experiments on four tabular datasets using both open-source\nand proprietary LLMs, we show the effectiveness of these methods in enhancing\ndemographic parity while maintaining high overall performance. Our analysis\nprovides actionable insights for practitioners in selecting the most suitable\napproach based on their specific requirements and constraints.\n", "date": "2024-12-05", "categories": []}
{"id": "2412.04739", "title": "Fair Diagnosis: Leveraging Causal Modeling to Mitigate Medical Bias", "abstract": "  In medical image analysis, model predictions can be affected by sensitive\nattributes, such as race and gender, leading to fairness concerns and potential\nbiases in diagnostic outcomes. To mitigate this, we present a causal modeling\nframework, which aims to reduce the impact of sensitive attributes on\ndiagnostic predictions. Our approach introduces a novel fairness criterion,\n\\textbf{Diagnosis Fairness}, and a unique fairness metric, leveraging\npath-specific fairness to control the influence of demographic attributes,\nensuring that predictions are primarily informed by clinically relevant\nfeatures rather than sensitive attributes. By incorporating adversarial\nperturbation masks, our framework directs the model to focus on critical image\nregions, suppressing bias-inducing information. Experimental results across\nmultiple datasets demonstrate that our framework effectively reduces bias\ndirectly associated with sensitive attributes while preserving diagnostic\naccuracy. Our findings suggest that causal modeling can enhance both fairness\nand interpretability in AI-powered clinical decision support systems.\n", "date": "2024-12-06", "categories": []}
{"id": "2407.17449", "title": "Looking at Model Debiasing through the Lens of Anomaly Detection", "abstract": "  It is widely recognized that deep neural networks are sensitive to bias in\nthe data. This means that during training these models are likely to learn\nspurious correlations between data and labels, resulting in limited\ngeneralization abilities and low performance. In this context, model debiasing\napproaches can be devised aiming at reducing the model's dependency on such\nunwanted correlations, either leveraging the knowledge of bias information or\nnot. In this work, we focus on the latter and more realistic scenario, showing\nthe importance of accurately predicting the bias-conflicting and bias-aligned\nsamples to obtain compelling performance in bias mitigation. On this ground, we\npropose to conceive the problem of model bias from an out-of-distribution\nperspective, introducing a new bias identification method based on anomaly\ndetection. We claim that when data is mostly biased, bias-conflicting samples\ncan be regarded as outliers with respect to the bias-aligned distribution in\nthe feature space of a biased model, thus allowing for precisely detecting them\nwith an anomaly detection method. Coupling the proposed bias identification\napproach with bias-conflicting data upsampling and augmentation in a two-step\nstrategy, we reach state-of-the-art performance on synthetic and real benchmark\ndatasets. Ultimately, our proposed approach shows that the data bias issue does\nnot necessarily require complex debiasing methods, given that an accurate bias\nidentification procedure is defined. Source code is available at\nhttps://github.com/Malga-Vision/MoDAD\n", "date": "2024-07-24", "categories": []}
{"id": "2412.03009", "title": "Data Acquisition for Improving Model Fairness using Reinforcement\n  Learning", "abstract": "  Machine learning systems are increasingly being used in critical decision\nmaking such as healthcare, finance, and criminal justice. Concerns around their\nfairness have resulted in several bias mitigation techniques that emphasize the\nneed for high-quality data to ensure fairer decisions. However, the role of\nearlier stages of machine learning pipelines in mitigating model bias has not\nbeen explored well. In this paper, we focus on the task of acquiring additional\nlabeled data points for training the downstream machine learning model to\nrapidly improve its fairness. Since not all data points in a data pool are\nequally beneficial to the task of fairness, we generate an ordering in which\ndata points should be acquired. We present DataSift, a data acquisition\nframework based on the idea of data valuation that relies on partitioning and\nmulti-armed bandits to determine the most valuable data points to acquire. Over\nseveral iterations, DataSift selects a partition and randomly samples a batch\nof data points from the selected partition, evaluates the benefit of acquiring\nthe batch on model fairness, and updates the utility of partitions depending on\nthe benefit. To further improve the effectiveness and efficiency of evaluating\nbatches, we leverage influence functions that estimate the effect of acquiring\na batch without retraining the model. We empirically evaluate DataSift on\nseveral real-world and synthetic datasets and show that the fairness of a\nmachine learning model can be significantly improved even while acquiring a few\ndata points.\n", "date": "2024-12-04", "categories": []}
{"id": "2412.03537", "title": "Evaluating Gender Bias Transfer between Pre-trained and Prompt-Adapted\n  Language Models", "abstract": "  Large language models (LLMs) are increasingly being adapted to achieve\ntask-specificity for deployment in real-world decision systems. Several\nprevious works have investigated the bias transfer hypothesis (BTH) by studying\nthe effect of the fine-tuning adaptation strategy on model fairness to find\nthat fairness in pre-trained masked language models have limited effect on the\nfairness of models when adapted using fine-tuning. In this work, we expand the\nstudy of BTH to causal models under prompt adaptations, as prompting is an\naccessible, and compute-efficient way to deploy models in real-world systems.\nIn contrast to previous works, we establish that intrinsic biases in\npre-trained Mistral, Falcon and Llama models are strongly correlated (rho >=\n0.94) with biases when the same models are zero- and few-shot prompted, using a\npronoun co-reference resolution task. Further, we find that bias transfer\nremains strongly correlated even when LLMs are specifically prompted to exhibit\nfair or biased behavior (rho >= 0.92), and few-shot length and stereotypical\ncomposition are varied (rho >= 0.97). Our findings highlight the importance of\nensuring fairness in pre-trained LLMs, especially when they are later used to\nperform downstream tasks via prompt adaptation.\n", "date": "2024-12-04", "categories": []}
{"id": "2412.00051", "title": "TransFair: Transferring Fairness from Ocular Disease Classification to\n  Progression Prediction", "abstract": "  The use of artificial intelligence (AI) in automated disease classification\nsignificantly reduces healthcare costs and improves the accessibility of\nservices. However, this transformation has given rise to concerns about the\nfairness of AI, which disproportionately affects certain groups, particularly\npatients from underprivileged populations. Recently, a number of methods and\nlarge-scale datasets have been proposed to address group performance\ndisparities. Although these methods have shown effectiveness in disease\nclassification tasks, they may fall short in ensuring fair prediction of\ndisease progression, mainly because of limited longitudinal data with diverse\ndemographics available for training a robust and equitable prediction model. In\nthis paper, we introduce TransFair to enhance demographic fairness in\nprogression prediction for ocular diseases. TransFair aims to transfer a\nfairness-enhanced disease classification model to the task of progression\nprediction with fairness preserved. Specifically, we train a fair EfficientNet,\ntermed FairEN, equipped with a fairness-aware attention mechanism using\nextensive data for ocular disease classification. Subsequently, this fair\nclassification model is adapted to a fair progression prediction model through\nknowledge distillation, which aims to minimize the latent feature distances\nbetween the classification and progression prediction models. We evaluate\nFairEN and TransFair for fairness-enhanced ocular disease classification and\nprogression prediction using both two-dimensional (2D) and 3D retinal images.\nExtensive experiments and comparisons with models with and without considering\nfairness learning show that TransFair effectively enhances demographic equity\nin predicting ocular disease progression.\n", "date": "2024-11-24", "categories": []}
{"id": "2412.02528", "title": "Bias Analysis of AI Models for Undergraduate Student Admissions", "abstract": "  Bias detection and mitigation is an active area of research in machine\nlearning. This work extends previous research done by the authors to provide a\nrigorous and more complete analysis of the bias found in AI predictive models.\nAdmissions data spanning six years was used to create an AI model to determine\nwhether a given student would be directly admitted into the School of Science\nunder various scenarios at a large urban research university. During this time,\nsubmission of standardized test scores as part of an application became\noptional which led to interesting questions about the impact of standardized\ntest scores on admission decisions. We developed and analyzed AI models to\nunderstand which variables are important in admissions decisions, and how the\ndecision to exclude test scores affects the demographics of the students who\nare admitted. We then evaluated the predictive models to detect and analyze\nbiases these models may carry with respect to three variables chosen to\nrepresent sensitive populations: gender, race, and whether a student was the\nfirst in his or her family to attend college. We also extended our analysis to\nshow that the biases detected were persistent. Finally, we included several\nfairness metrics in our analysis and discussed the uses and limitations of\nthese metrics.\n", "date": "2024-12-03", "categories": []}
{"id": "2402.01071", "title": "Chameleon: Foundation Models for Fairness-aware Multi-modal Data\n  Augmentation to Enhance Coverage of Minorities", "abstract": "  The potential harms of the under-representation of minorities in training\ndata, particularly in multi-modal settings, is a well-recognized concern. While\nthere has been extensive effort in detecting such under-representation,\nresolution has remained a challenge. With recent advancements in generative AI,\nlarge language models and foundation models have emerged as versatile tools\nacross various domains. In this paper, we propose Chameleon, a system that\nefficiently utilizes these tools to augment a data set with a minimal addition\nof synthetically generated tuples, in order to enhance the coverage of the\nunder-represented groups. Our system follows a rejection sampling approach to\nensure the generated tuples have a high quality and follow the underlying\ndistribution. In order to minimize the rejection chance of the generated\ntuples, we propose multiple strategies for providing a guide for the foundation\nmodel. Our experiment results, in addition to confirming the efficiency of our\nproposed algorithms, illustrate the effectiveness of our approach, as the\nunfairness of the model in a downstream task significantly dropped after data\nrepair using Chameleon.\n", "date": "2024-02-02", "categories": []}
{"id": "2405.06433", "title": "Fair Mixed Effects Support Vector Machine", "abstract": "  To ensure unbiased and ethical automated predictions, fairness must be a core\nprinciple in machine learning applications. Fairness in machine learning aims\nto mitigate biases present in the training data and model imperfections that\ncould lead to discriminatory outcomes. This is achieved by preventing the model\nfrom making decisions based on sensitive characteristics like ethnicity or\nsexual orientation. A fundamental assumption in machine learning is the\nindependence of observations. However, this assumption often does not hold true\nfor data describing social phenomena, where data points are often clustered\nbased. Hence, if the machine learning models do not account for the cluster\ncorrelations, the results may be biased. Especially high is the bias in cases\nwhere the cluster assignment is correlated to the variable of interest. We\npresent a fair mixed effects support vector machine algorithm that can handle\nboth problems simultaneously. With a reproducible simulation study we\ndemonstrate the impact of clustered data on the quality of fair machine\nlearning predictions.\n", "date": "2024-05-10", "categories": []}
{"id": "2406.16738", "title": "Inducing Group Fairness in Prompt-Based Language Model Decisions", "abstract": "  Classifiers are used throughout industry to enforce policies, ranging from\nthe detection of toxic content to age-appropriate content filtering. While\nthese classifiers serve important functions, it is also essential that they are\nbuilt in ways that minimize unfair biases for users.\n  One such fairness consideration is called group fairness, which desires that\ndifferent sub-population of users receive equal treatment. This is a\nwell-studied problem in the context of 'classical' classifiers. However, the\nemergence of prompt-based language model (LM) decision making has created new\nopportunities to solve text-based classification tasks, and the fairness\nproperties of these new classifiers are not yet well understood. Further, the\n`remediation toolkit' is incomplete for LM-based decision makers and little is\nunderstood about how to improve decision maker group fairness while maintaining\nclassifier performance.\n  This work sets out to add more tools to that toolbox. We introduce\nadaptations of existing effective approaches from the classical classifier\nfairness to the prompt-based classifier space. We also devise simple methods\nthat take advantage of the new structure of prompt-based decision makers and\noperate at the prompt level. We compare these approaches empirically on real\ndata. Our results suggest that adaptations of approaches that are effective for\nclassical classifiers remain effective in the LM-based classifier environment.\nHowever, there is room for further exploration of prompt-based remediation\nmethods (and other remediation methods that take advantage of LM structure).\n", "date": "2024-06-24", "categories": []}
{"id": "2411.07656", "title": "Mitigating Bias in Queer Representation within Large Language Models: A\n  Collaborative Agent Approach", "abstract": "  Large Language Models (LLMs) often perpetuate biases in pronoun usage,\nleading to misrepresentation or exclusion of queer individuals. This paper\naddresses the specific problem of biased pronoun usage in LLM outputs,\nparticularly the inappropriate use of traditionally gendered pronouns (\"he,\"\n\"she\") when inclusive language is needed to accurately represent all\nidentities. We introduce a collaborative agent pipeline designed to mitigate\nthese biases by analyzing and optimizing pronoun usage for inclusivity. Our\nmulti-agent framework includes specialized agents for both bias detection and\ncorrection. Experimental evaluations using the Tango dataset-a benchmark\nfocused on gender pronoun usage-demonstrate that our approach significantly\nimproves inclusive pronoun classification, achieving a 32.6 percentage point\nincrease over GPT-4o in correctly disagreeing with inappropriate traditionally\ngendered pronouns $(\\chi^2 = 38.57, p < 0.0001)$. These results accentuate the\npotential of agent-driven frameworks in enhancing fairness and inclusivity in\nAI-generated content, demonstrating their efficacy in reducing biases and\npromoting socially responsible AI.\n", "date": "2024-11-12", "categories": []}
{"id": "2412.00245", "title": "Integrating Social Determinants of Health into Knowledge Graphs:\n  Evaluating Prediction Bias and Fairness in Healthcare", "abstract": "  Social determinants of health (SDoH) play a crucial role in patient health\noutcomes, yet their integration into biomedical knowledge graphs remains\nunderexplored. This study addresses this gap by constructing an SDoH-enriched\nknowledge graph using the MIMIC-III dataset and PrimeKG. We introduce a novel\nfairness formulation for graph embeddings, focusing on invariance with respect\nto sensitive SDoH information. Via employing a heterogeneous-GCN model for\ndrug-disease link prediction, we detect biases related to various SDoH factors.\nTo mitigate these biases, we propose a post-processing method that\nstrategically reweights edges connected to SDoHs, balancing their influence on\ngraph representations. This approach represents one of the first comprehensive\ninvestigations into fairness issues within biomedical knowledge graphs\nincorporating SDoH. Our work not only highlights the importance of considering\nSDoH in medical informatics but also provides a concrete method for reducing\nSDoH-related biases in link prediction tasks, paving the way for more equitable\nhealthcare recommendations. Our code is available at\n\\url{https://github.com/hwq0726/SDoH-KG}.\n", "date": "2024-11-29", "categories": []}
{"id": "2412.00606", "title": "Fairness at Every Intersection: Uncovering and Mitigating Intersectional\n  Biases in Multimodal Clinical Predictions", "abstract": "  Biases in automated clinical decision-making using Electronic Healthcare\nRecords (EHR) impose significant disparities in patient care and treatment\noutcomes. Conventional approaches have primarily focused on bias mitigation\nstrategies stemming from single attributes, overlooking intersectional\nsubgroups -- groups formed across various demographic intersections (such as\nrace, gender, ethnicity, etc.). Rendering single-attribute mitigation\nstrategies to intersectional subgroups becomes statistically irrelevant due to\nthe varying distribution and bias patterns across these subgroups. The\nmultimodal nature of EHR -- data from various sources such as combinations of\ntext, time series, tabular, events, and images -- adds another layer of\ncomplexity as the influence on minority groups may fluctuate across modalities.\nIn this paper, we take the initial steps to uncover potential intersectional\nbiases in predictions by sourcing extensive multimodal datasets, MIMIC-Eye1 and\nMIMIC-IV ED, and propose mitigation at the intersectional subgroup level. We\nperform and benchmark downstream tasks and bias evaluation on the datasets by\nlearning a unified text representation from multimodal sources, harnessing the\nenormous capabilities of the pre-trained clinical Language Models (LM),\nMedBERT, Clinical BERT, and Clinical BioBERT. Our findings indicate that the\nproposed sub-group-specific bias mitigation is robust across different\ndatasets, subgroups, and embeddings, demonstrating effectiveness in addressing\nintersectional biases in multimodal settings.\n", "date": "2024-11-30", "categories": []}
{"id": "2412.01065", "title": "Lookahead Counterfactual Fairness", "abstract": "  As machine learning (ML) algorithms are used in applications that involve\nhumans, concerns have arisen that these algorithms may be biased against\ncertain social groups. \\textit{Counterfactual fairness} (CF) is a fairness\nnotion proposed in Kusner et al. (2017) that measures the unfairness of ML\npredictions; it requires that the prediction perceived by an individual in the\nreal world has the same marginal distribution as it would be in a\ncounterfactual world, in which the individual belongs to a different group.\nAlthough CF ensures fair ML predictions, it fails to consider the downstream\neffects of ML predictions on individuals. Since humans are strategic and often\nadapt their behaviors in response to the ML system, predictions that satisfy CF\nmay not lead to a fair future outcome for the individuals. In this paper, we\nintroduce \\textit{lookahead counterfactual fairness} (LCF), a fairness notion\naccounting for the downstream effects of ML models which requires the\nindividual \\textit{future status} to be counterfactually fair. We theoretically\nidentify conditions under which LCF can be satisfied and propose an algorithm\nbased on the theorems. We also extend the concept to path-dependent fairness.\nExperiments on both synthetic and real data validate the proposed method.\n", "date": "2024-12-02", "categories": []}
{"id": "2410.20739", "title": "Gender Bias in LLM-generated Interview Responses", "abstract": "  LLMs have emerged as a promising tool for assisting individuals in diverse\ntext-generation tasks, including job-related texts. However, LLM-generated\nanswers have been increasingly found to exhibit gender bias. This study\nevaluates three LLMs (GPT-3.5, GPT-4, Claude) to conduct a multifaceted audit\nof LLM-generated interview responses across models, question types, and jobs,\nand their alignment with two gender stereotypes. Our findings reveal that\ngender bias is consistent, and closely aligned with gender stereotypes and the\ndominance of jobs. Overall, this study contributes to the systematic\nexamination of gender bias in LLM-generated interview responses, highlighting\nthe need for a mindful approach to mitigate such biases in related\napplications.\n", "date": "2024-10-28", "categories": []}
{"id": "2411.18442", "title": "Metric-DST: Mitigating Selection Bias Through Diversity-Guided\n  Semi-Supervised Metric Learning", "abstract": "  Selection bias poses a critical challenge for fairness in machine learning,\nas models trained on data that is less representative of the population might\nexhibit undesirable behavior for underrepresented profiles. Semi-supervised\nlearning strategies like self-training can mitigate selection bias by\nincorporating unlabeled data into model training to gain further insight into\nthe distribution of the population. However, conventional self-training seeks\nto include high-confidence data samples, which may reinforce existing model\nbias and compromise effectiveness. We propose Metric-DST, a diversity-guided\nself-training strategy that leverages metric learning and its implicit\nembedding space to counter confidence-based bias through the inclusion of more\ndiverse samples. Metric-DST learned more robust models in the presence of\nselection bias for generated and real-world datasets with induced bias, as well\nas a molecular biology prediction task with intrinsic bias. The Metric-DST\nlearning strategy offers a flexible and widely applicable solution to mitigate\nselection bias and enhance fairness of machine learning models.\n", "date": "2024-11-27", "categories": []}
{"id": "2411.19090", "title": "ABROCA Distributions For Algorithmic Bias Assessment: Considerations\n  Around Interpretation", "abstract": "  Algorithmic bias continues to be a key concern of learning analytics. We\nstudy the statistical properties of the Absolute Between-ROC Area (ABROCA)\nmetric. This fairness measure quantifies group-level differences in classifier\nperformance through the absolute difference in ROC curves. ABROCA is\nparticularly useful for detecting nuanced performance differences even when\noverall Area Under the ROC Curve (AUC) values are similar. We sample ABROCA\nunder various conditions, including varying AUC differences and class\ndistributions. We find that ABROCA distributions exhibit high skewness\ndependent on sample sizes, AUC differences, and class imbalance. When assessing\nwhether a classifier is biased, this skewness inflates ABROCA values by chance,\neven when data is drawn (by simulation) from populations with equivalent ROC\ncurves. These findings suggest that ABROCA requires careful interpretation\ngiven its distributional properties, especially when used to assess the degree\nof bias and when classes are imbalanced.\n", "date": "2024-11-28", "categories": []}
{"id": "2404.13139", "title": "Explainable AI for Fair Sepsis Mortality Predictive Model", "abstract": "  Artificial intelligence supports healthcare professionals with predictive\nmodeling, greatly transforming clinical decision-making. This study addresses\nthe crucial need for fairness and explainability in AI applications within\nhealthcare to ensure equitable outcomes across diverse patient demographics. By\nfocusing on the predictive modeling of sepsis-related mortality, we propose a\nmethod that learns a performance-optimized predictive model and then employs\nthe transfer learning process to produce a model with better fairness. Our\nmethod also introduces a novel permutation-based feature importance algorithm\naiming at elucidating the contribution of each feature in enhancing fairness on\npredictions. Unlike existing explainability methods concentrating on explaining\nfeature contribution to predictive performance, our proposed method uniquely\nbridges the gap in understanding how each feature contributes to fairness. This\nadvancement is pivotal, given sepsis's significant mortality rate and its role\nin one-third of hospital deaths. Our method not only aids in identifying and\nmitigating biases within the predictive model but also fosters trust among\nhealthcare stakeholders by improving the transparency and fairness of model\npredictions, thereby contributing to more equitable and trustworthy healthcare\ndelivery.\n", "date": "2024-04-19", "categories": []}
{"id": "2306.01646", "title": "Auditing for Human Expertise", "abstract": "  High-stakes prediction tasks (e.g., patient diagnosis) are often handled by\ntrained human experts. A common source of concern about automation in these\nsettings is that experts may exercise intuition that is difficult to model\nand/or have access to information (e.g., conversations with a patient) that is\nsimply unavailable to a would-be algorithm. This raises a natural question\nwhether human experts add value which could not be captured by an algorithmic\npredictor. We develop a statistical framework under which we can pose this\nquestion as a natural hypothesis test. Indeed, as our framework highlights,\ndetecting human expertise is more subtle than simply comparing the accuracy of\nexpert predictions to those made by a particular learning algorithm. Instead,\nwe propose a simple procedure which tests whether expert predictions are\nstatistically independent from the outcomes of interest after conditioning on\nthe available inputs (`features'). A rejection of our test thus suggests that\nhuman experts may add value to any algorithm trained on the available data, and\nhas direct implications for whether human-AI `complementarity' is achievable in\na given prediction task. We highlight the utility of our procedure using\nadmissions data collected from the emergency department of a large academic\nhospital system, where we show that physicians' admit/discharge decisions for\npatients with acute gastrointestinal bleeding (AGIB) appear to be incorporating\ninformation that is not available to a standard algorithmic screening tool.\nThis is despite the fact that the screening tool is arguably more accurate than\nphysicians' discretionary decisions, highlighting that -- even absent normative\nconcerns about accountability or interpretability -- accuracy is insufficient\nto justify algorithmic automation.\n", "date": "2023-06-02", "categories": []}
{"id": "2305.03223", "title": "Structural Group Unfairness: Measurement and Mitigation by means of the\n  Effective Resistance", "abstract": "  Social networks contribute to the distribution of social capital, defined as\nthe relationships, norms of trust and reciprocity within a community or society\nthat facilitate cooperation and collective action. Therefore, better positioned\nmembers in a social network benefit from faster access to diverse information\nand higher influence on information dissemination. A variety of methods have\nbeen proposed in the literature to measure social capital at an individual\nlevel. However, there is a lack of methods to quantify social capital at a\ngroup level, which is particularly important when the groups are defined on the\ngrounds of protected attributes. To fill this gap, we propose to measure the\nsocial capital of a group of nodes by means of the effective resistance and\nemphasize the importance of considering the entire network topology. Grounded\nin spectral graph theory, we introduce three effective resistance-based\nmeasures of group social capital, namely group isolation, group diameter and\ngroup control, where the groups are defined according to the value of a\nprotected attribute. We denote the social capital disparity among different\ngroups in a network as structural group unfairness, and propose to mitigate it\nby means of a budgeted edge augmentation heuristic that systematically\nincreases the social capital of the most disadvantaged group. In experiments on\nreal-world networks, we uncover significant levels of structural group\nunfairness when using gender as the protected attribute, with females being the\nmost disadvantaged group in comparison to males. We also illustrate how our\nproposed edge augmentation approach is able to not only effectively mitigate\nthe structural group unfairness but also increase the social capital of all\ngroups in the network.\n", "date": "2023-05-05", "categories": []}
{"id": "2411.14469", "title": "Popular LLMs Amplify Race and Gender Disparities in Human Mobility", "abstract": "  As large language models (LLMs) are increasingly applied in areas influencing\nsocietal outcomes, it is critical to understand their tendency to perpetuate\nand amplify biases. This study investigates whether LLMs exhibit biases in\npredicting human mobility -- a fundamental human behavior -- based on race and\ngender. Using three prominent LLMs -- GPT-4, Gemini, and Claude -- we analyzed\ntheir predictions of visitations to points of interest (POIs) for individuals,\nrelying on prompts that included names with and without explicit demographic\ndetails. We find that LLMs frequently reflect and amplify existing societal\nbiases. Specifically, predictions for minority groups were disproportionately\nskewed, with these individuals being significantly less likely to be associated\nwith wealth-related points of interest (POIs). Gender biases were also evident,\nas female individuals were consistently linked to fewer career-related POIs\ncompared to their male counterparts. These biased associations suggest that\nLLMs not only mirror but also exacerbate societal stereotypes, particularly in\ncontexts involving race and gender.\n", "date": "2024-11-18", "categories": []}
{"id": "2411.14500", "title": "Exploring Accuracy-Fairness Trade-off in Large Language Models", "abstract": "  Large Language Models (LLMs) have made significant strides in the field of\nartificial intelligence, showcasing their ability to interact with humans and\ninfluence human cognition through information dissemination. However, recent\nstudies have brought to light instances of bias inherent within these LLMs,\npresenting a critical issue that demands attention. In our research, we delve\ndeeper into the intricate challenge of harmonising accuracy and fairness in the\nenhancement of LLMs. While improving accuracy can indeed enhance overall LLM\nperformance, it often occurs at the expense of fairness. Overemphasising\noptimisation of one metric invariably leads to a significant degradation of the\nother. This underscores the necessity of taking into account multiple\nconsiderations during the design and optimisation phases of LLMs. Therefore, we\nadvocate for reformulating the LLM training process as a multi-objective\nlearning task. Our investigation reveals that multi-objective evolutionary\nlearning (MOEL) methodologies offer promising avenues for tackling this\nchallenge. Our MOEL framework enables the simultaneous optimisation of both\naccuracy and fairness metrics, resulting in a Pareto-optimal set of LLMs. In\nsummary, our study sheds valuable lights on the delicate equilibrium between\naccuracy and fairness within LLMs, which is increasingly significant for their\nreal-world applications. By harnessing MOEL, we present a promising pathway\ntowards fairer and more efficacious AI technologies.\n", "date": "2024-11-21", "categories": []}
{"id": "2411.15051", "title": "Fantastic Biases (What are They) and Where to Find Them", "abstract": "  Deep Learning models tend to learn correlations of patterns on huge datasets.\nThe bigger these systems are, the more complex are the phenomena they can\ndetect, and the more data they need for this. The use of Artificial\nIntelligence (AI) is becoming increasingly ubiquitous in our society, and its\nimpact is growing everyday. The promises it holds strongly depend on their fair\nand universal use, such as access to information or education for all. In a\nworld of inequalities, they can help to reach the most disadvantaged areas.\nHowever, such a universal systems must be able to represent society, without\nbenefiting some at the expense of others. We must not reproduce the\ninequalities observed throughout the world, but educate these IAs to go beyond\nthem. We have seen cases where these systems use gender, race, or even class\ninformation in ways that are not appropriate for resolving their tasks. Instead\nof real causal reasoning, they rely on spurious correlations, which is what we\nusually call a bias. In this paper, we first attempt to define what is a bias\nin general terms. It helps us to demystify the concept of bias, to understand\nwhy we can find them everywhere and why they are sometimes useful. Second, we\nfocus over the notion of what is generally seen as negative bias, the one we\nwant to avoid in machine learning, before presenting a general zoology\ncontaining the most common of these biases. We finally conclude by looking at\nclassical methods to detect them, by means of specially crafted datasets of\ntemplates and specific algorithms, and also classical methods to mitigate them.\n", "date": "2024-11-22", "categories": []}
{"id": "2402.01002", "title": "AI-generated faces influence gender stereotypes and racial\n  homogenization", "abstract": "  Text-to-image generative AI models such as Stable Diffusion are used daily by\nmillions worldwide. However, the extent to which these models exhibit racial\nand gender stereotypes is not yet fully understood. Here, we document\nsignificant biases in Stable Diffusion across six races, two genders, 32\nprofessions, and eight attributes. Additionally, we examine the degree to which\nStable Diffusion depicts individuals of the same race as being similar to one\nanother. This analysis reveals significant racial homogenization, e.g.,\ndepicting nearly all Middle Eastern men as bearded, brown-skinned, and wearing\ntraditional attire. We then propose debiasing solutions that allow users to\nspecify the desired distributions of race and gender when generating images\nwhile minimizing racial homogenization. Finally, using a preregistered survey\nexperiment, we find evidence that being presented with inclusive AI-generated\nfaces reduces people's racial and gender biases, while being presented with\nnon-inclusive ones increases such biases, regardless of whether the images are\nlabeled as AI-generated. Taken together, our findings emphasize the need to\naddress biases and stereotypes in text-to-image models.\n", "date": "2024-02-01", "categories": []}
{"id": "2411.13738", "title": "Assessing Gender Bias in LLMs: Comparing LLM Outputs with Human\n  Perceptions and Official Statistics", "abstract": "  This study investigates gender bias in large language models (LLMs) by\ncomparing their gender perception to that of human respondents, U.S. Bureau of\nLabor Statistics data, and a 50% no-bias benchmark. We created a new evaluation\nset using occupational data and role-specific sentences. Unlike common\nbenchmarks included in LLM training data, our set is newly developed,\npreventing data leakage and test set contamination. Five LLMs were tested to\npredict the gender for each role using single-word answers. We used\nKullback-Leibler (KL) divergence to compare model outputs with human\nperceptions, statistical data, and the 50% neutrality benchmark. All LLMs\nshowed significant deviation from gender neutrality and aligned more with\nstatistical data, still reflecting inherent biases.\n", "date": "2024-11-20", "categories": []}
{"id": "2406.00599", "title": "Robust Fair Clustering with Group Membership Uncertainty Sets", "abstract": "  We study the canonical fair clustering problem where each cluster is\nconstrained to have close to population-level representation of each group.\nDespite significant attention, the salient issue of having incomplete knowledge\nabout the group membership of each point has been superficially addressed. In\nthis paper, we consider a setting where the assigned group memberships are\nnoisy. We introduce a simple noise model that requires a small number of\nparameters to be given by the decision maker. We then present an algorithm for\nfair clustering with provable \\emph{robustness} guarantees. Our framework\nenables the decision maker to trade off between the robustness and the\nclustering quality. Unlike previous work, our algorithms are backed by\nworst-case theoretical guarantees. Finally, we empirically verify the\nperformance of our algorithm on real world datasets and show its superior\nperformance over existing baselines.\n", "date": "2024-06-02", "categories": []}
{"id": "2411.10544", "title": "Debias-CLR: A Contrastive Learning Based Debiasing Method for\n  Algorithmic Fairness in Healthcare Applications", "abstract": "  Artificial intelligence based predictive models trained on the clinical notes\ncan be demographically biased. This could lead to adverse healthcare\ndisparities in predicting outcomes like length of stay of the patients. Thus,\nit is necessary to mitigate the demographic biases within these models. We\nproposed an implicit in-processing debiasing method to combat disparate\ntreatment which occurs when the machine learning model predict different\noutcomes for individuals based on the sensitive attributes like gender,\nethnicity, race, and likewise. For this purpose, we used clinical notes of\nheart failure patients and used diagnostic codes, procedure reports and\nphysiological vitals of the patients. We used Clinical BERT to obtain feature\nembeddings within the diagnostic codes and procedure reports, and LSTM\nautoencoders to obtain feature embeddings within the physiological vitals.\nThen, we trained two separate deep learning contrastive learning frameworks,\none for gender and the other for ethnicity to obtain debiased representations\nwithin those demographic traits. We called this debiasing framework Debias-CLR.\nWe leveraged clinical phenotypes of the patients identified in the diagnostic\ncodes and procedure reports in the previous study to measure fairness\nstatistically. We found that Debias-CLR was able to reduce the Single-Category\nWord Embedding Association Test (SC-WEAT) effect size score when debiasing for\ngender and ethnicity. We further found that to obtain fair representations in\nthe embedding space using Debias-CLR, the accuracy of the predictive models on\ndownstream tasks like predicting length of stay of the patients did not get\nreduced as compared to using the un-debiased counterparts for training the\npredictive models. Hence, we conclude that our proposed approach, Debias-CLR is\nfair and representative in mitigating demographic biases and can reduce health\ndisparities.\n", "date": "2024-11-15", "categories": []}
{"id": "2411.12846", "title": "Towards Fairness in AI for Melanoma Detection: Systemic Review and\n  Recommendations", "abstract": "  Early and accurate melanoma detection is crucial for improving patient\noutcomes. Recent advancements in artificial intelligence AI have shown promise\nin this area, but the technologys effectiveness across diverse skin tones\nremains a critical challenge. This study conducts a systematic review and\npreliminary analysis of AI based melanoma detection research published between\n2013 and 2024, focusing on deep learning methodologies, datasets, and skin tone\nrepresentation. Our findings indicate that while AI can enhance melanoma\ndetection, there is a significant bias towards lighter skin tones. To address\nthis, we propose including skin hue in addition to skin tone as represented by\nthe LOreal Color Chart Map for a more comprehensive skin tone assessment\ntechnique. This research highlights the need for diverse datasets and robust\nevaluation metrics to develop AI models that are equitable and effective for\nall patients. By adopting best practices outlined in a PRISMA Equity framework\ntailored for healthcare and melanoma detection, we can work towards reducing\ndisparities in melanoma outcomes.\n", "date": "2024-11-19", "categories": []}
{"id": "2305.18160", "title": "Counterpart Fairness -- Addressing Systematic between-group Differences\n  in Fairness Evaluation", "abstract": "  When using machine learning to aid decision-making, it is critical to ensure\nthat an algorithmic decision is fair and does not discriminate against specific\nindividuals/groups, particularly those from underprivileged populations.\nExisting group fairness methods aim to ensure equal outcomes (such as loan\napproval rates) across groups delineated by protected variables like race or\ngender. However, in cases where systematic differences between groups play a\nsignificant role in outcomes, these methods may overlook the influence of\nnon-protected variables that can systematically vary across groups. These\nconfounding factors can affect fairness evaluations, making it challenging to\nassess whether disparities are due to discrimination or inherent differences.\nTherefore, we recommend a more refined and comprehensive fairness index that\naccounts for both the systematic differences within groups and the\nmultifaceted, intertwined confounding effects. The proposed index evaluates\nfairness on counterparts (pairs of individuals who are similar with respect to\nthe task of interest but from different groups), whose group identities cannot\nbe distinguished algorithmically by exploring confounding factors. To identify\ncounterparts, we developed a two-step matching method inspired by propensity\nscore and metric learning. In addition, we introduced a counterpart-based\nstatistical fairness index, called Counterpart Fairness (CFair), to assess the\nfairness of machine learning models. Empirical results on the MIMIC and COMPAS\ndatasets indicate that standard group-based fairness metrics may not adequately\ninform about the degree of unfairness present in predictions, as revealed\nthrough CFair.\n", "date": "2023-05-29", "categories": []}
{"id": "2411.11101", "title": "Different Horses for Different Courses: Comparing Bias Mitigation\n  Algorithms in ML", "abstract": "  With fairness concerns gaining significant attention in Machine Learning\n(ML), several bias mitigation techniques have been proposed, often compared\nagainst each other to find the best method. These benchmarking efforts tend to\nuse a common setup for evaluation under the assumption that providing a uniform\nenvironment ensures a fair comparison. However, bias mitigation techniques are\nsensitive to hyperparameter choices, random seeds, feature selection, etc.,\nmeaning that comparison on just one setting can unfairly favour certain\nalgorithms. In this work, we show significant variance in fairness achieved by\nseveral algorithms and the influence of the learning pipeline on fairness\nscores. We highlight that most bias mitigation techniques can achieve\ncomparable performance, given the freedom to perform hyperparameter\noptimization, suggesting that the choice of the evaluation parameters-rather\nthan the mitigation technique itself-can sometimes create the perceived\nsuperiority of one method over another. We hope our work encourages future\nresearch on how various choices in the lifecycle of developing an algorithm\nimpact fairness, and trends that guide the selection of appropriate algorithms.\n", "date": "2024-11-17", "categories": []}
{"id": "2411.11939", "title": "Fair Distillation: Teaching Fairness from Biased Teachers in Medical\n  Imaging", "abstract": "  Deep learning has achieved remarkable success in image classification and\nsegmentation tasks. However, fairness concerns persist, as models often exhibit\nbiases that disproportionately affect demographic groups defined by sensitive\nattributes such as race, gender, or age. Existing bias-mitigation techniques,\nincluding Subgroup Re-balancing, Adversarial Training, and Domain\nGeneralization, aim to balance accuracy across demographic groups, but often\nfail to simultaneously improve overall accuracy, group-specific accuracy, and\nfairness due to conflicts among these interdependent objectives. We propose the\nFair Distillation (FairDi) method, a novel fairness approach that decomposes\nthese objectives by leveraging biased ``teacher'' models, each optimized for a\nspecific demographic group. These teacher models then guide the training of a\nunified ``student'' model, which distills their knowledge to maximize overall\nand group-specific accuracies, while minimizing inter-group disparities.\nExperiments on medical imaging datasets show that FairDi achieves significant\ngains in both overall and group-specific accuracy, along with improved\nfairness, compared to existing methods. FairDi is adaptable to various medical\ntasks, such as classification and segmentation, and provides an effective\nsolution for equitable model performance.\n", "date": "2024-11-18", "categories": []}
{"id": "2211.10285", "title": "A Fair Loss Function for Network Pruning", "abstract": "  Model pruning can enable the deployment of neural networks in environments\nwith resource constraints. While pruning may have a small effect on the overall\nperformance of the model, it can exacerbate existing biases into the model such\nthat subsets of samples see significantly degraded performance. In this paper,\nwe introduce the performance weighted loss function, a simple modified\ncross-entropy loss function that can be used to limit the introduction of\nbiases during pruning. Experiments using the CelebA, Fitzpatrick17k and\nCIFAR-10 datasets demonstrate that the proposed method is a simple and\neffective tool that can enable existing pruning methods to be used in fairness\nsensitive contexts. Code used to produce all experiments contained in this\npaper can be found at https://github.com/robbiemeyer/pw_loss_pruning.\n", "date": "2022-11-18", "categories": []}
{"id": "2407.01697", "title": "NLPGuard: A Framework for Mitigating the Use of Protected Attributes by\n  NLP Classifiers", "abstract": "  AI regulations are expected to prohibit machine learning models from using\nsensitive attributes during training. However, the latest Natural Language\nProcessing (NLP) classifiers, which rely on deep learning, operate as black-box\nsystems, complicating the detection and remediation of such misuse. Traditional\nbias mitigation methods in NLP aim for comparable performance across different\ngroups based on attributes like gender or race but fail to address the\nunderlying issue of reliance on protected attributes. To partly fix that, we\nintroduce NLPGuard, a framework for mitigating the reliance on protected\nattributes in NLP classifiers. NLPGuard takes an unlabeled dataset, an existing\nNLP classifier, and its training data as input, producing a modified training\ndataset that significantly reduces dependence on protected attributes without\ncompromising accuracy. NLPGuard is applied to three classification tasks:\nidentifying toxic language, sentiment analysis, and occupation classification.\nOur evaluation shows that current NLP classifiers heavily depend on protected\nattributes, with up to $23\\%$ of the most predictive words associated with\nthese attributes. However, NLPGuard effectively reduces this reliance by up to\n$79\\%$, while slightly improving accuracy.\n", "date": "2024-07-01", "categories": []}
{"id": "2410.00836", "title": "Towards Fairness and Privacy: A Novel Data Pre-processing Optimization\n  Framework for Non-binary Protected Attributes", "abstract": "  The reason behind the unfair outcomes of AI is often rooted in biased\ndatasets. Therefore, this work presents a framework for addressing fairness by\ndebiasing datasets containing a (non-)binary protected attribute. The framework\nproposes a combinatorial optimization problem where heuristics such as genetic\nalgorithms can be used to solve for the stated fairness objectives. The\nframework addresses this by finding a data subset that minimizes a certain\ndiscrimination measure. Depending on a user-defined setting, the framework\nenables different use cases, such as data removal, the addition of synthetic\ndata, or exclusive use of synthetic data. The exclusive use of synthetic data\nin particular enhances the framework's ability to preserve privacy while\noptimizing for fairness. In a comprehensive evaluation, we demonstrate that\nunder our framework, genetic algorithms can effectively yield fairer datasets\ncompared to the original data. In contrast to prior work, the framework\nexhibits a high degree of flexibility as it is metric- and task-agnostic, can\nbe applied to both binary or non-binary protected attributes, and demonstrates\nefficient runtime.\n", "date": "2024-10-01", "categories": []}
{"id": "2411.02569", "title": "The Intersectionality Problem for Algorithmic Fairness", "abstract": "  A yet unmet challenge in algorithmic fairness is the problem of\nintersectionality, that is, achieving fairness across the intersection of\nmultiple groups -- and verifying that such fairness has been attained. Because\nintersectional groups tend to be small, verifying whether a model is fair\nraises statistical as well as moral-methodological challenges. This paper (1)\nelucidates the problem of intersectionality in algorithmic fairness, (2)\ndevelops desiderata to clarify the challenges underlying the problem and guide\nthe search for potential solutions, (3) illustrates the desiderata and\npotential solutions by sketching a proposal using simple hypothesis testing,\nand (4) evaluates, partly empirically, this proposal against the proposed\ndesiderata.\n", "date": "2024-11-04", "categories": []}
{"id": "2405.04412", "title": "The Silicon Ceiling: Auditing GPT's Race and Gender Biases in Hiring", "abstract": "  Large language models (LLMs) are increasingly being introduced in workplace\nsettings, with the goals of improving efficiency and fairness. However,\nconcerns have arisen regarding these models' potential to reflect or exacerbate\nsocial biases and stereotypes. This study explores the potential impact of LLMs\non hiring practices. To do so, we conduct an AI audit of race and gender biases\nin one commonly-used LLM, OpenAI's GPT-3.5, taking inspiration from the history\nof traditional offline resume audits. We conduct two studies using names with\nvaried race and gender connotations: resume assessment (Study 1) and resume\ngeneration (Study 2). In Study 1, we ask GPT to score resumes with 32 different\nnames (4 names for each combination of the 2 gender and 4 racial groups) and\ntwo anonymous options across 10 occupations and 3 evaluation tasks (overall\nrating, willingness to interview, and hireability). We find that the model\nreflects some biases based on stereotypes. In Study 2, we prompt GPT to create\nresumes (10 for each name) for fictitious job candidates. When generating\nresumes, GPT reveals underlying biases; women's resumes had occupations with\nless experience, while Asian and Hispanic resumes had immigrant markers, such\nas non-native English and non-U.S. education and work experiences. Our findings\ncontribute to a growing body of literature on LLM biases, particularly in\nworkplace contexts.\n", "date": "2024-05-07", "categories": []}
{"id": "2411.09826", "title": "Evaluating Gender Bias in Large Language Models", "abstract": "  Gender bias in artificial intelligence has become an important issue,\nparticularly in the context of language models used in communication-oriented\napplications. This study examines the extent to which Large Language Models\n(LLMs) exhibit gender bias in pronoun selection in occupational contexts. The\nanalysis evaluates the models GPT-4, GPT-4o, PaLM 2 Text Bison and Gemini 1.0\nPro using a self-generated dataset. The jobs considered include a range of\noccupations, from those with a significant male presence to those with a\nnotable female concentration, as well as jobs with a relatively equal gender\ndistribution. Three different sentence processing methods were used to assess\npotential gender bias: masked tokens, unmasked sentences, and sentence\ncompletion. In addition, the LLMs suggested names of individuals in specific\noccupations, which were then examined for gender distribution. The results show\na positive correlation between the models' pronoun choices and the gender\ndistribution present in U.S. labor force data. Female pronouns were more often\nassociated with female-dominated occupations, while male pronouns were more\noften associated with male-dominated occupations. Sentence completion showed\nthe strongest correlation with actual gender distribution, while name\ngeneration resulted in a more balanced 'politically correct' gender\ndistribution, albeit with notable variations in predominantly male or female\noccupations. Overall, the prompting method had a greater impact on gender\ndistribution than the model selection itself, highlighting the complexity of\naddressing gender bias in LLMs. The findings highlight the importance of\nprompting in gender mapping.\n", "date": "2024-11-14", "categories": []}
{"id": "2411.09056", "title": "Optimisation Strategies for Ensuring Fairness in Machine Learning: With\n  and Without Demographics", "abstract": "  Ensuring fairness has emerged as one of the primary concerns in AI and its\nrelated algorithms. Over time, the field of machine learning fairness has\nevolved to address these issues. This paper provides an extensive overview of\nthis field and introduces two formal frameworks to tackle open questions in\nmachine learning fairness.\n  In one framework, operator-valued optimisation and min-max objectives are\nemployed to address unfairness in time-series problems. This approach showcases\nstate-of-the-art performance on the notorious COMPAS benchmark dataset,\ndemonstrating its effectiveness in real-world scenarios.\n  In the second framework, the challenge of lacking sensitive attributes, such\nas gender and race, in commonly used datasets is addressed. This issue is\nparticularly pressing because existing algorithms in this field predominantly\nrely on the availability or estimations of such attributes to assess and\nmitigate unfairness. Here, a framework for a group-blind bias-repair is\nintroduced, aiming to mitigate bias without relying on sensitive attributes.\nThe efficacy of this approach is showcased through analyses conducted on the\nAdult Census Income dataset.\n  Additionally, detailed algorithmic analyses for both frameworks are provided,\naccompanied by convergence guarantees, ensuring the robustness and reliability\nof the proposed methodologies.\n", "date": "2024-11-13", "categories": []}
{"id": "2407.17474", "title": "\"My Kind of Woman\": Analysing Gender Stereotypes in AI through The\n  Averageness Theory and EU Law", "abstract": "  This study delves into gender classification systems, shedding light on the\ninteraction between social stereotypes and algorithmic determinations. Drawing\non the \"averageness theory,\" which suggests a relationship between a face's\nattractiveness and the human ability to ascertain its gender, we explore the\npotential propagation of human bias into artificial intelligence (AI) systems.\nUtilising the AI model Stable Diffusion 2.1, we have created a dataset\ncontaining various connotations of attractiveness to test whether the\ncorrelation between attractiveness and accuracy in gender classification\nobserved in human cognition persists within AI. Our findings indicate that akin\nto human dynamics, AI systems exhibit variations in gender classification\naccuracy based on attractiveness, mirroring social prejudices and stereotypes\nin their algorithmic decisions. This discovery underscores the critical need to\nconsider the impacts of human perceptions on data collection and highlights the\nnecessity for a multidisciplinary and intersectional approach to AI development\nand AI data training. By incorporating cognitive psychology and feminist legal\ntheory, we examine how data used for AI training can foster gender diversity\nand fairness under the scope of the AI Act and GDPR, reaffirming how\npsychological and feminist legal theories can offer valuable insights for\nensuring the protection of gender equality and non-discrimination in AI\nsystems.\n", "date": "2024-06-27", "categories": []}
{"id": "2305.13862", "title": "A Trip Towards Fairness: Bias and De-Biasing in Large Language Models", "abstract": "  Cheap-to-Build Very Large-Language Models (CtB-LLMs) with affordable training\nare emerging as the next big revolution in natural language processing and\nunderstanding. These CtB-LLMs are democratizing access to trainable Very\nLarge-Language Models (VLLMs) and, thus, may represent the building blocks of\nmany NLP systems solving downstream tasks. Hence, a little or a large bias in\nCtB-LLMs may cause huge harm. In this paper, we performed a large investigation\nof the bias of three families of CtB-LLMs, and we showed that debiasing\ntechniques are effective and usable. Indeed, according to current tests, the\nLLaMA and the OPT families have an important bias in gender, race, religion,\nand profession. In contrast to the analysis for other LLMs, we discovered that\nbias depends not on the number of parameters but on the perplexity. Finally,\nthe debiasing of OPT using LoRA reduces bias up to 4.12 points in the\nnormalized stereotype score.\n", "date": "2023-05-23", "categories": []}
{"id": "2402.17106", "title": "Achievable Fairness on Your Data With Utility Guarantees", "abstract": "  In machine learning fairness, training models that minimize disparity across\ndifferent sensitive groups often leads to diminished accuracy, a phenomenon\nknown as the fairness-accuracy trade-off. The severity of this trade-off\ninherently depends on dataset characteristics such as dataset imbalances or\nbiases and therefore, using a uniform fairness requirement across diverse\ndatasets remains questionable. To address this, we present a computationally\nefficient approach to approximate the fairness-accuracy trade-off curve\ntailored to individual datasets, backed by rigorous statistical guarantees. By\nutilizing the You-Only-Train-Once (YOTO) framework, our approach mitigates the\ncomputational burden of having to train multiple models when approximating the\ntrade-off curve. Crucially, we introduce a novel methodology for quantifying\nuncertainty in our estimates, thereby providing practitioners with a robust\nframework for auditing model fairness while avoiding false conclusions due to\nestimation errors. Our experiments spanning tabular (e.g., Adult), image\n(CelebA), and language (Jigsaw) datasets underscore that our approach not only\nreliably quantifies the optimum achievable trade-offs across various data\nmodalities but also helps detect suboptimality in SOTA fairness methods.\n", "date": "2024-02-27", "categories": []}
{"id": "2411.02467", "title": "Towards Harmless Rawlsian Fairness Regardless of Demographic Prior", "abstract": "  Due to privacy and security concerns, recent advancements in group fairness\nadvocate for model training regardless of demographic information. However,\nmost methods still require prior knowledge of demographics. In this study, we\nexplore the potential for achieving fairness without compromising its utility\nwhen no prior demographics are provided to the training set, namely\n\\emph{harmless Rawlsian fairness}. We ascertain that such a fairness\nrequirement with no prior demographic information essential promotes training\nlosses to exhibit a Dirac delta distribution. To this end, we propose a simple\nbut effective method named VFair to minimize the variance of training losses\ninside the optimal set of empirical losses. This problem is then optimized by a\ntailored dynamic update approach that operates in both loss and gradient\ndimensions, directing the model towards relatively fairer solutions while\npreserving its intact utility. Our experimental findings indicate that\nregression tasks, which are relatively unexplored from literature, can achieve\nsignificant fairness improvement through VFair regardless of any prior, whereas\nclassification tasks usually do not because of their quantized utility\nmeasurements. The implementation of our method is publicly available at\n\\url{https://github.com/wxqpxw/VFair}.\n", "date": "2024-11-04", "categories": []}
{"id": "2411.06881", "title": "WassFFed: Wasserstein Fair Federated Learning", "abstract": "  Federated Learning (FL) employs a training approach to address scenarios\nwhere users' data cannot be shared across clients. Achieving fairness in FL is\nimperative since training data in FL is inherently geographically distributed\namong diverse user groups. Existing research on fairness predominantly assumes\naccess to the entire training data, making direct transfer to FL challenging.\nHowever, the limited existing research on fairness in FL does not effectively\naddress two key challenges, i.e., (CH1) Current methods fail to deal with the\ninconsistency between fair optimization results obtained with surrogate\nfunctions and fair classification results. (CH2) Directly aggregating local\nfair models does not always yield a globally fair model due to non Identical\nand Independent data Distributions (non-IID) among clients. To address these\nchallenges, we propose a Wasserstein Fair Federated Learning framework, namely\nWassFFed. To tackle CH1, we ensure that the outputs of local models, rather\nthan the loss calculated with surrogate functions or classification results\nwith a threshold, remain independent of various user groups. To resolve CH2, we\nemploy a Wasserstein barycenter calculation of all local models' outputs for\neach user group, bringing local model outputs closer to the global output\ndistribution to ensure consistency between the global model and local models.\nWe conduct extensive experiments on three real-world datasets, demonstrating\nthat WassFFed outperforms existing approaches in striking a balance between\naccuracy and fairness.\n", "date": "2024-11-11", "categories": []}
{"id": "2310.11407", "title": "Group-blind optimal transport to group parity and its constrained\n  variants", "abstract": "  Fairness holds a pivotal role in the realm of machine learning, particularly\nwhen it comes to addressing groups categorised by protected attributes, e.g.,\ngender, race. Prevailing algorithms in fair learning predominantly hinge on\naccessibility or estimations of these protected attributes, at least in the\ntraining process. We design a single group-blind projection map that aligns the\nfeature distributions of both groups in the source data, achieving\n(demographic) group parity, without requiring values of the protected attribute\nfor individual samples in the computation of the map, as well as its use.\nInstead, our approach utilises the feature distributions of the privileged and\nunprivileged groups in a boarder population and the essential assumption that\nthe source data are unbiased representation of the population. We present\nnumerical results on synthetic data and real data.\n", "date": "2023-10-17", "categories": []}
{"id": "2402.12789", "title": "Fairness Without Harm: An Influence-Guided Active Sampling Approach", "abstract": "  The pursuit of fairness in machine learning (ML), ensuring that the models do\nnot exhibit biases toward protected demographic groups, typically results in a\ncompromise scenario. This compromise can be explained by a Pareto frontier\nwhere given certain resources (e.g., data), reducing the fairness violations\noften comes at the cost of lowering the model accuracy. In this work, we aim to\ntrain models that mitigate group fairness disparity without causing harm to\nmodel accuracy. Intuitively, acquiring more data is a natural and promising\napproach to achieve this goal by reaching a better Pareto frontier of the\nfairness-accuracy tradeoff. The current data acquisition methods, such as fair\nactive learning approaches, typically require annotating sensitive attributes.\nHowever, these sensitive attribute annotations should be protected due to\nprivacy and safety concerns. In this paper, we propose a tractable active data\nsampling algorithm that does not rely on training group annotations, instead\nonly requiring group annotations on a small validation set. Specifically, the\nalgorithm first scores each new example by its influence on fairness and\naccuracy evaluated on the validation dataset, and then selects a certain number\nof examples for training. We theoretically analyze how acquiring more data can\nimprove fairness without causing harm, and validate the possibility of our\nsampling approach in the context of risk disparity. We also provide the upper\nbound of generalization error and risk disparity as well as the corresponding\nconnections. Extensive experiments on real-world data demonstrate the\neffectiveness of our proposed algorithm. Our code is available at\nhttps://github.com/UCSC-REAL/FairnessWithoutHarm.\n", "date": "2024-02-20", "categories": []}
{"id": "2408.17396", "title": "Fairness-Aware Estimation of Graphical Models", "abstract": "  This paper examines the issue of fairness in the estimation of graphical\nmodels (GMs), particularly Gaussian, Covariance, and Ising models. These models\nplay a vital role in understanding complex relationships in high-dimensional\ndata. However, standard GMs can result in biased outcomes, especially when the\nunderlying data involves sensitive characteristics or protected groups. To\naddress this, we introduce a comprehensive framework designed to reduce bias in\nthe estimation of GMs related to protected attributes. Our approach involves\nthe integration of the pairwise graph disparity error and a tailored loss\nfunction into a nonsmooth multi-objective optimization problem, striving to\nachieve fairness across different sensitive groups while maintaining the\neffectiveness of the GMs. Experimental evaluations on synthetic and real-world\ndatasets demonstrate that our framework effectively mitigates bias without\nundermining GMs' performance.\n", "date": "2024-08-30", "categories": []}
{"id": "2411.05318", "title": "Fairness in Monotone $k$-submodular Maximization: Algorithms and\n  Applications", "abstract": "  Submodular optimization has become increasingly prominent in machine learning\nand fairness has drawn much attention. In this paper, we propose to study the\nfair $k$-submodular maximization problem and develop a\n$\\frac{1}{3}$-approximation greedy algorithm with a running time of\n$\\mathcal{O}(knB)$. To the best of our knowledge, our work is the first to\nincorporate fairness in the context of $k$-submodular maximization, and our\ntheoretical guarantee matches the best-known $k$-submodular maximization\nresults without fairness constraints. In addition, we have developed a faster\nthreshold-based algorithm that achieves a $(\\frac{1}{3} - \\epsilon)$\napproximation with $\\mathcal{O}(\\frac{kn}{\\epsilon} \\log \\frac{B}{\\epsilon})$\nevaluations of the function $f$. Furthermore, for both algorithms, we provide\napproximation guarantees when the $k$-submodular function is not accessible but\nonly can be approximately accessed. We have extensively validated our\ntheoretical findings through empirical research and examined the practical\nimplications of fairness. Specifically, we have addressed the question: ``What\nis the price of fairness?\" through case studies on influence maximization with\n$k$ topics and sensor placement with $k$ types. The experimental results show\nthat the fairness constraints do not significantly undermine the quality of\nsolutions.\n", "date": "2024-11-08", "categories": []}
{"id": "2411.05648", "title": "Enhancing Model Fairness and Accuracy with Similarity Networks: A\n  Methodological Approach", "abstract": "  In this paper, we propose an innovative approach to thoroughly explore\ndataset features that introduce bias in downstream machine-learning tasks.\nDepending on the data format, we use different techniques to map instances into\na similarity feature space. Our method's ability to adjust the resolution of\npairwise similarity provides clear insights into the relationship between the\ndataset classification complexity and model fairness. Experimental results\nconfirm the promising applicability of the similarity network in promoting fair\nmodels. Moreover, leveraging our methodology not only seems promising in\nproviding a fair downstream task such as classification, it also performs well\nin imputation and augmentation of the dataset satisfying the fairness criteria\nsuch as demographic parity and imbalanced classes.\n", "date": "2024-11-08", "categories": []}
{"id": "2411.00190", "title": "Monitoring fairness in machine learning models that predict patient\n  mortality in the ICU", "abstract": "  This work proposes a fairness monitoring approach for machine learning models\nthat predict patient mortality in the ICU. We investigate how well models\nperform for patient groups with different race, sex and medical diagnoses. We\ninvestigate Documentation bias in clinical measurement, showing how fairness\nanalysis provides a more detailed and insightful comparison of model\nperformance than traditional accuracy metrics alone.\n", "date": "2024-10-31", "categories": []}
{"id": "2411.04298", "title": "A Capabilities Approach to Studying Bias and Harm in Language\n  Technologies", "abstract": "  Mainstream Natural Language Processing (NLP) research has ignored the\nmajority of the world's languages. In moving from excluding the majority of the\nworld's languages to blindly adopting what we make for English, we first risk\nimporting the same harms we have at best mitigated and at least measured for\nEnglish. However, in evaluating and mitigating harms arising from adopting new\ntechnologies into such contexts, we often disregard (1) the actual community\nneeds of Language Technologies, and (2) biases and fairness issues within the\ncontext of the communities. In this extended abstract, we consider fairness,\nbias, and inclusion in Language Technologies through the lens of the\nCapabilities Approach. The Capabilities Approach centers on what people are\ncapable of achieving, given their intersectional social, political, and\neconomic contexts instead of what resources are (theoretically) available to\nthem. We detail the Capabilities Approach, its relationship to multilingual and\nmulticultural evaluation, and how the framework affords meaningful\ncollaboration with community members in defining and measuring the harms of\nLanguage Technologies.\n", "date": "2024-11-06", "categories": []}
{"id": "2411.04453", "title": "Comparing Fairness of Generative Mobility Models", "abstract": "  This work examines the fairness of generative mobility models, addressing the\noften overlooked dimension of equity in model performance across geographic\nregions. Predictive models built on crowd flow data are instrumental in\nunderstanding urban structures and movement patterns; however, they risk\nembedding biases, particularly in spatiotemporal contexts where model\nperformance may reflect and reinforce existing inequities tied to geographic\ndistribution. We propose a novel framework for assessing fairness by measuring\nthe utility and equity of generated traces. Utility is assessed via the Common\nPart of Commuters (CPC), a similarity metric comparing generated and real\nmobility flows, while fairness is evaluated using demographic parity. By\nreformulating demographic parity to reflect the difference in CPC distribution\nbetween two groups, our analysis reveals disparities in how various models\nencode biases present in the underlying data. We utilized four models (Gravity,\nRadiation, Deep Gravity, and Non-linear Gravity) and our results indicate that\ntraditional gravity and radiation models produce fairer outcomes, although Deep\nGravity achieves higher CPC. This disparity underscores a trade-off between\nmodel accuracy and equity, with the feature-rich Deep Gravity model amplifying\npre-existing biases in community representations. Our findings emphasize the\nimportance of integrating fairness metrics in mobility modeling to avoid\nperpetuating inequities.\n", "date": "2024-11-07", "categories": []}
{"id": "2403.19863", "title": "DeNetDM: Debiasing by Network Depth Modulation", "abstract": "  Neural networks trained on biased datasets tend to inadvertently learn\nspurious correlations, hindering generalization. We formally prove that (1)\nsamples that exhibit spurious correlations lie on a lower rank manifold\nrelative to the ones that do not; and (2) the depth of a network acts as an\nimplicit regularizer on the rank of the attribute subspace that is encoded in\nits representations. Leveraging these insights, we present DeNetDM, a novel\ndebiasing method that uses network depth modulation as a way of developing\nrobustness to spurious correlations. Using a training paradigm derived from\nProduct of Experts, we create both biased and debiased branches with deep and\nshallow architectures and then distill knowledge to produce the target debiased\nmodel. Our method requires no bias annotations or explicit data augmentation\nwhile performing on par with approaches that require either or both. We\ndemonstrate that DeNetDM outperforms existing debiasing techniques on both\nsynthetic and real-world datasets by 5\\%. The project page is available at\nhttps://vssilpa.github.io/denetdm/.\n", "date": "2024-03-28", "categories": []}
{"id": "2304.02190", "title": "Globalizing Fairness Attributes in Machine Learning: A Case Study on\n  Health in Africa", "abstract": "  With growing machine learning (ML) applications in healthcare, there have\nbeen calls for fairness in ML to understand and mitigate ethical concerns these\nsystems may pose. Fairness has implications for global health in Africa, which\nalready has inequitable power imbalances between the Global North and South.\nThis paper seeks to explore fairness for global health, with Africa as a case\nstudy. We propose fairness attributes for consideration in the African context\nand delineate where they may come into play in different ML-enabled medical\nmodalities. This work serves as a basis and call for action for furthering\nresearch into fairness in global health.\n", "date": "2023-04-05", "categories": []}
{"id": "2403.03357", "title": "The Case for Globalizing Fairness: A Mixed Methods Study on Colonialism,\n  AI, and Health in Africa", "abstract": "  With growing application of machine learning (ML) technologies in healthcare,\nthere have been calls for developing techniques to understand and mitigate\nbiases these systems may exhibit. Fair-ness considerations in the development\nof ML-based solutions for health have particular implications for Africa, which\nalready faces inequitable power imbalances between the Global North and\nSouth.This paper seeks to explore fairness for global health, with Africa as a\ncase study. We conduct a scoping review to propose axes of disparities for\nfairness consideration in the African context and delineate where they may come\ninto play in different ML-enabled medical modalities. We then conduct\nqualitative research studies with 672 general population study participants and\n28 experts inML, health, and policy focused on Africa to obtain corroborative\nevidence on the proposed axes of disparities. Our analysis focuses on\ncolonialism as the attribute of interest and examines the interplay between\nartificial intelligence (AI), health, and colonialism. Among the pre-identified\nattributes, we found that colonial history, country of origin, and national\nincome level were specific axes of disparities that participants believed would\ncause an AI system to be biased.However, there was also divergence of opinion\nbetween experts and general population participants. Whereas experts generally\nexpressed a shared view about the relevance of colonial history for the\ndevelopment and implementation of AI technologies in Africa, the majority of\nthe general population participants surveyed did not think there was a direct\nlink between AI and colonialism. Based on these findings, we provide practical\nrecommendations for developing fairness-aware ML solutions for health in\nAfrica.\n", "date": "2024-03-05", "categories": []}
{"id": "2406.06487", "title": "When is Multicalibration Post-Processing Necessary?", "abstract": "  Calibration is a well-studied property of predictors which guarantees\nmeaningful uncertainty estimates. Multicalibration is a related notion --\noriginating in algorithmic fairness -- which requires predictors to be\nsimultaneously calibrated over a potentially complex and overlapping collection\nof protected subpopulations (such as groups defined by ethnicity, race, or\nincome). We conduct the first comprehensive study evaluating the usefulness of\nmulticalibration post-processing across a broad set of tabular, image, and\nlanguage datasets for models spanning from simple decision trees to 90 million\nparameter fine-tuned LLMs. Our findings can be summarized as follows: (1)\nmodels which are calibrated out of the box tend to be relatively\nmulticalibrated without any additional post-processing; (2) multicalibration\npost-processing can help inherently uncalibrated models and large vision and\nlanguage models; and (3) traditional calibration measures may sometimes provide\nmulticalibration implicitly. More generally, we also distill many independent\nobservations which may be useful for practical and effective applications of\nmulticalibration post-processing in real-world contexts. We also release a\npython package implementing multicalibration algorithms, available via `pip\ninstall multicalibration'.\n", "date": "2024-06-10", "categories": []}
{"id": "2407.13710", "title": "OxonFair: A Flexible Toolkit for Algorithmic Fairness", "abstract": "  We present OxonFair, a new open source toolkit for enforcing fairness in\nbinary classification. Compared to existing toolkits: (i) We support NLP and\nComputer Vision classification as well as standard tabular problems. (ii) We\nsupport enforcing fairness on validation data, making us robust to a wide range\nof overfitting challenges. (iii) Our approach can optimize any measure based on\nTrue Positives, False Positive, False Negatives, and True Negatives. This makes\nit easily extensible and much more expressive than existing toolkits. It\nsupports all 9 and all 10 of the decision-based group metrics of two popular\nreview articles. (iv) We jointly optimize a performance objective alongside\nfairness constraints. This minimizes degradation while enforcing fairness, and\neven improves the performance of inadequately tuned unfair baselines. OxonFair\nis compatible with standard ML toolkits, including sklearn, Autogluon, and\nPyTorch and is available at https://github.com/oxfordinternetinstitute/oxonfair\n", "date": "2024-06-30", "categories": []}
{"id": "2410.19003", "title": "Whither Bias Goes, I Will Go: An Integrative, Systematic Review of\n  Algorithmic Bias Mitigation", "abstract": "  Machine learning (ML) models are increasingly used for personnel assessment\nand selection (e.g., resume screeners, automatically scored interviews).\nHowever, concerns have been raised throughout society that ML assessments may\nbe biased and perpetuate or exacerbate inequality. Although organizational\nresearchers have begun investigating ML assessments from traditional\npsychometric and legal perspectives, there is a need to understand, clarify,\nand integrate fairness operationalizations and algorithmic bias mitigation\nmethods from the computer science, data science, and organizational research\nliteratures. We present a four-stage model of developing ML assessments and\napplying bias mitigation methods, including 1) generating the training data, 2)\ntraining the model, 3) testing the model, and 4) deploying the model. When\nintroducing the four-stage model, we describe potential sources of bias and\nunfairness at each stage. Then, we systematically review definitions and\noperationalizations of algorithmic bias, legal requirements governing personnel\nselection from the United States and Europe, and research on algorithmic bias\nmitigation across multiple domains and integrate these findings into our\nframework. Our review provides insights for both research and practice by\nelucidating possible mechanisms of algorithmic bias while identifying which\nbias mitigation methods are legal and effective. This integrative framework\nalso reveals gaps in the knowledge of algorithmic bias mitigation that should\nbe addressed by future collaborative research between organizational\nresearchers, computer scientists, and data scientists. We provide\nrecommendations for developing and deploying ML assessments, as well as\nrecommendations for future research into algorithmic bias and fairness.\n", "date": "2024-10-21", "categories": []}
{"id": "2411.02414", "title": "Fairness Evaluation with Item Response Theory", "abstract": "  Item Response Theory (IRT) has been widely used in educational psychometrics\nto assess student ability, as well as the difficulty and discrimination of test\nquestions. In this context, discrimination specifically refers to how\neffectively a question distinguishes between students of different ability\nlevels, and it does not carry any connotation related to fairness. In recent\nyears, IRT has been successfully used to evaluate the predictive performance of\nMachine Learning (ML) models, but this paper marks its first application in\nfairness evaluation. In this paper, we propose a novel Fair-IRT framework to\nevaluate a set of predictive models on a set of individuals, while\nsimultaneously eliciting specific parameters, namely, the ability to make fair\npredictions (a feature of predictive models), as well as the discrimination and\ndifficulty of individuals that affect the prediction results. Furthermore, we\nconduct a series of experiments to comprehensively understand the implications\nof these parameters for fairness evaluation. Detailed explanations for item\ncharacteristic curves (ICCs) are provided for particular individuals. We\npropose the flatness of ICCs to disentangle the unfairness between individuals\nand predictive models. The experiments demonstrate the effectiveness of this\nframework as a fairness evaluation tool. Two real-world case studies illustrate\nits potential application in evaluating fairness in both classification and\nregression tasks. Our paper aligns well with the Responsible Web track by\nproposing a Fair-IRT framework to evaluate fairness in ML models, which\ndirectly contributes to the development of a more inclusive, equitable, and\ntrustworthy AI.\n", "date": "2024-10-20", "categories": []}
{"id": "2411.02654", "title": "Fair and Welfare-Efficient Constrained Multi-matchings under Uncertainty", "abstract": "  We study fair allocation of constrained resources, where a market designer\noptimizes overall welfare while maintaining group fairness. In many large-scale\nsettings, utilities are not known in advance, but are instead observed after\nrealizing the allocation. We therefore estimate agent utilities using machine\nlearning. Optimizing over estimates requires trading-off between mean utilities\nand their predictive variances. We discuss these trade-offs under two paradigms\nfor preference modeling -- in the stochastic optimization regime, the market\ndesigner has access to a probability distribution over utilities, and in the\nrobust optimization regime they have access to an uncertainty set containing\nthe true utilities with high probability. We discuss utilitarian and\negalitarian welfare objectives, and we explore how to optimize for them under\nstochastic and robust paradigms. We demonstrate the efficacy of our approaches\non three publicly available conference reviewer assignment datasets. The\napproaches presented enable scalable constrained resource allocation under\nuncertainty for many combinations of objectives and preference models.\n", "date": "2024-11-04", "categories": []}
{"id": "2411.03068", "title": "Alpha and Prejudice: Improving $\\alpha$-sized Worst-case Fairness via\n  Intrinsic Reweighting", "abstract": "  Worst-case fairness with off-the-shelf demographics achieves group parity by\nmaximizing the model utility of the worst-off group. Nevertheless, demographic\ninformation is often unavailable in practical scenarios, which impedes the use\nof such a direct max-min formulation. Recent advances have reframed this\nlearning problem by introducing the lower bound of minimal partition ratio,\ndenoted as $\\alpha$, as side information, referred to as ``$\\alpha$-sized\nworst-case fairness'' in this paper. We first justify the practical\nsignificance of this setting by presenting noteworthy evidence from the data\nprivacy perspective, which has been overlooked by existing research. Without\nimposing specific requirements on loss functions, we propose reweighting the\ntraining samples based on their intrinsic importance to fairness. Given the\nglobal nature of the worst-case formulation, we further develop a stochastic\nlearning scheme to simplify the training process without compromising model\nperformance. Additionally, we address the issue of outliers and provide a\nrobust variant to handle potential outliers during model training. Our\ntheoretical analysis and experimental observations reveal the connections\nbetween the proposed approaches and existing ``fairness-through-reweighting''\nstudies, with extensive experimental results on fairness benchmarks\ndemonstrating the superiority of our methods.\n", "date": "2024-11-05", "categories": []}
{"id": "2304.13917", "title": "Proportionally Representative Clustering", "abstract": "  In recent years, there has been a surge in effort to formalize notions of\nfairness in machine learning. We focus on centroid clustering--one of the\nfundamental tasks in unsupervised machine learning. We propose a new axiom\n``proportionally representative fairness'' (PRF) that is designed for\nclustering problems where the selection of centroids reflects the distribution\nof data points and how tightly they are clustered together. Our fairness\nconcept is not satisfied by existing fair clustering algorithms. We design\nefficient algorithms to achieve PRF both for unconstrained and discrete\nclustering problems. Our algorithm for the unconstrained setting is also the\nfirst known polynomial-time approximation algorithm for the well-studied\nProportional Fairness (PF) axiom. Our algorithm for the discrete setting also\nmatches the best known approximation factor for PF.\n", "date": "2023-04-27", "categories": []}
{"id": "2404.13859", "title": "Unveiling and Mitigating Generalized Biases of DNNs through the\n  Intrinsic Dimensions of Perceptual Manifolds", "abstract": "  Building fair deep neural networks (DNNs) is a crucial step towards achieving\ntrustworthy artificial intelligence. Delving into deeper factors that affect\nthe fairness of DNNs is paramount and serves as the foundation for mitigating\nmodel biases. However, current methods are limited in accurately predicting DNN\nbiases, relying solely on the number of training samples and lacking more\nprecise measurement tools. Here, we establish a geometric perspective for\nanalyzing the fairness of DNNs, comprehensively exploring how DNNs internally\nshape the intrinsic geometric characteristics of datasets-the intrinsic\ndimensions (IDs) of perceptual manifolds, and the impact of IDs on the fairness\nof DNNs. Based on multiple findings, we propose Intrinsic Dimension\nRegularization (IDR), which enhances the fairness and performance of models by\npromoting the learning of concise and ID-balanced class perceptual manifolds.\nIn various image recognition benchmark tests, IDR significantly mitigates model\nbias while improving its performance.\n", "date": "2024-04-22", "categories": []}
{"id": "2405.19300", "title": "Measuring and Mitigating Bias for Tabular Datasets with Multiple\n  Protected Attributes", "abstract": "  Motivated by the recital (67) of the current corrigendum of the AI Act in the\nEuropean Union, we propose and present measures and mitigation strategies for\ndiscrimination in tabular datasets. We specifically focus on datasets that\ncontain multiple protected attributes, such as nationality, age, and sex. This\nmakes measuring and mitigating bias more challenging, as many existing methods\nare designed for a single protected attribute. This paper comes with a twofold\ncontribution: Firstly, new discrimination measures are introduced. These\nmeasures are categorized in our framework along with existing ones, guiding\nresearchers and practitioners in choosing the right measure to assess the\nfairness of the underlying dataset. Secondly, a novel application of an\nexisting bias mitigation method, FairDo, is presented. We show that this\nstrategy can mitigate any type of discrimination, including intersectional\ndiscrimination, by transforming the dataset. By conducting experiments on\nreal-world datasets (Adult, Bank, COMPAS), we demonstrate that de-biasing\ndatasets with multiple protected attributes is possible. All transformed\ndatasets show a reduction in discrimination, on average by 28%. Further, these\ndatasets do not compromise any of the tested machine learning models'\nperformances significantly compared to the original datasets. Conclusively,\nthis study demonstrates the effectiveness of the mitigation strategy used and\ncontributes to the ongoing discussion on the implementation of the European\nUnion's AI Act.\n", "date": "2024-05-29", "categories": []}
{"id": "2407.03059", "title": "FairJob: A Real-World Dataset for Fairness in Online Systems", "abstract": "  We introduce a fairness-aware dataset for job recommendations in advertising,\ndesigned to foster research in algorithmic fairness within real-world\nscenarios. It was collected and prepared to comply with privacy standards and\nbusiness confidentiality. An additional challenge is the lack of access to\nprotected user attributes such as gender, for which we propose a solution to\nobtain a proxy estimate. Despite being anonymized and including a proxy for a\nsensitive attribute, our dataset preserves predictive power and maintains a\nrealistic and challenging benchmark. This dataset addresses a significant gap\nin the availability of fairness-focused resources for high-impact domains like\nadvertising -- the actual impact being having access or not to precious\nemployment opportunities, where balancing fairness and utility is a common\nindustrial challenge. We also explore various stages in the advertising process\nwhere unfairness can occur and introduce a method to compute a fair utility\nmetric for the job recommendations in online systems case from a biased\ndataset. Experimental evaluations of bias mitigation techniques on the released\ndataset demonstrate potential improvements in fairness and the associated\ntrade-offs with utility.\n  The dataset is hosted at https://huggingface.co/datasets/criteo/FairJob.\nSource code for the experiments is hosted at\nhttps://github.com/criteo-research/FairJob-dataset/.\n", "date": "2024-07-03", "categories": []}
{"id": "2409.12677", "title": "(Un)certainty of (Un)fairness: Preference-Based Selection of Certainly\n  Fair Decision-Makers", "abstract": "  Fairness metrics are used to assess discrimination and bias in\ndecision-making processes across various domains, including machine learning\nmodels and human decision-makers in real-world applications. This involves\ncalculating the disparities between probabilistic outcomes among social groups,\nsuch as acceptance rates between male and female applicants. However,\ntraditional fairness metrics do not account for the uncertainty in these\nprocesses and lack of comparability when two decision-makers exhibit the same\ndisparity. Using Bayesian statistics, we quantify the uncertainty of the\ndisparity to enhance discrimination assessments. We represent each\ndecision-maker, whether a machine learning model or a human, by its disparity\nand the corresponding uncertainty in that disparity. We define preferences over\ndecision-makers and utilize brute-force to choose the optimal decision-maker\naccording to a utility function that ranks decision-makers based on these\npreferences. The decision-maker with the highest utility score can be\ninterpreted as the one for whom we are most certain that it is fair.\n", "date": "2024-09-19", "categories": []}
{"id": "2411.00831", "title": "Saliency-Based diversity and fairness Metric and\n  FaceKeepOriginalAugment: A Novel Approach for Enhancing Fairness and\n  Diversity", "abstract": "  Data augmentation has become a pivotal tool in enhancing the performance of\ncomputer vision tasks, with the KeepOriginalAugment method emerging as a\nstandout technique for its intelligent incorporation of salient regions within\nless prominent areas, enabling augmentation in both regions. Despite its\nsuccess in image classification, its potential in addressing biases remains\nunexplored. In this study, we introduce an extension of the KeepOriginalAugment\nmethod, termed FaceKeepOriginalAugment, which explores various debiasing\naspects-geographical, gender, and stereotypical biases-in computer vision\nmodels. By maintaining a delicate balance between data diversity and\ninformation preservation, our approach empowers models to exploit both diverse\nsalient and non-salient regions, thereby fostering increased diversity and\ndebiasing effects. We investigate multiple strategies for determining the\nplacement of the salient region and swapping perspectives to decide which part\nundergoes augmentation. Leveraging the Image Similarity Score (ISS), we\nquantify dataset diversity across a range of datasets, including Flickr Faces\nHQ (FFHQ), WIKI, IMDB, Labelled Faces in the Wild (LFW), UTK Faces, and Diverse\nDataset. We evaluate the effectiveness of FaceKeepOriginalAugment in mitigating\ngender bias across CEO, Engineer, Nurse, and School Teacher datasets, utilizing\nthe Image-Image Association Score (IIAS) in convolutional neural networks\n(CNNs) and vision transformers (ViTs). Our findings shows the efficacy of\nFaceKeepOriginalAugment in promoting fairness and inclusivity within computer\nvision models, demonstrated by reduced gender bias and enhanced overall\nfairness. Additionally, we introduce a novel metric, Saliency-Based Diversity\nand Fairness Metric, which quantifies both diversity and fairness while\nhandling data imbalance across various datasets.\n", "date": "2024-10-29", "categories": []}
{"id": "2411.00997", "title": "Identifying Implicit Social Biases in Vision-Language Models", "abstract": "  Vision-language models, like CLIP (Contrastive Language Image Pretraining),\nare becoming increasingly popular for a wide range of multimodal retrieval\ntasks. However, prior work has shown that large language and deep vision models\ncan learn historical biases contained in their training sets, leading to\nperpetuation of stereotypes and potential downstream harm. In this work, we\nconduct a systematic analysis of the social biases that are present in CLIP,\nwith a focus on the interaction between image and text modalities. We first\npropose a taxonomy of social biases called So-B-IT, which contains 374 words\ncategorized across ten types of bias. Each type can lead to societal harm if\nassociated with a particular demographic group. Using this taxonomy, we examine\nimages retrieved by CLIP from a facial image dataset using each word as part of\na prompt. We find that CLIP frequently displays undesirable associations\nbetween harmful words and specific demographic groups, such as retrieving\nmostly pictures of Middle Eastern men when asked to retrieve images of a\n\"terrorist\". Finally, we conduct an analysis of the source of such biases, by\nshowing that the same harmful stereotypes are also present in a large\nimage-text dataset used to train CLIP models for examples of biases that we\nfind. Our findings highlight the importance of evaluating and addressing bias\nin vision-language models, and suggest the need for transparency and\nfairness-aware curation of large pre-training datasets.\n", "date": "2024-11-01", "categories": []}
{"id": "2411.01316", "title": "FEED: Fairness-Enhanced Meta-Learning for Domain Generalization", "abstract": "  Generalizing to out-of-distribution data while being aware of model fairness\nis a significant and challenging problem in meta-learning. The goal of this\nproblem is to find a set of fairness-aware invariant parameters of classifier\nthat is trained using data drawn from a family of related training domains with\ndistribution shift on non-sensitive features as well as different levels of\ndependence between model predictions and sensitive features so that the\nclassifier can achieve good generalization performance on unknown but distinct\ntest domains. To tackle this challenge, existing state-of-the-art methods\neither address the domain generalization problem but completely ignore learning\nwith fairness or solely specify shifted domains with various fairness levels.\nThis paper introduces an approach to fairness-aware meta-learning that\nsignificantly enhances domain generalization capabilities. Our framework,\nFairness-Enhanced Meta-Learning for Domain Generalization (FEED), disentangles\nlatent data representations into content, style, and sensitive vectors. This\ndisentanglement facilitates the robust generalization of machine learning\nmodels across diverse domains while adhering to fairness constraints. Unlike\ntraditional methods that focus primarily on domain invariance or sensitivity to\nshifts, our model integrates a fairness-aware invariance criterion directly\ninto the meta-learning process. This integration ensures that the learned\nparameters uphold fairness consistently, even when domain characteristics vary\nwidely. We validate our approach through extensive experiments across multiple\nbenchmarks, demonstrating not only superior performance in maintaining high\naccuracy and fairness but also significant improvements over existing\nstate-of-the-art methods in domain generalization tasks.\n", "date": "2024-11-02", "categories": []}
{"id": "2406.06407", "title": "A Taxonomy of Challenges to Curating Fair Datasets", "abstract": "  Despite extensive efforts to create fairer machine learning (ML) datasets,\nthere remains a limited understanding of the practical aspects of dataset\ncuration. Drawing from interviews with 30 ML dataset curators, we present a\ncomprehensive taxonomy of the challenges and trade-offs encountered throughout\nthe dataset curation lifecycle. Our findings underscore overarching issues\nwithin the broader fairness landscape that impact data curation. We conclude\nwith recommendations aimed at fostering systemic changes to better facilitate\nfair dataset curation practices.\n", "date": "2024-06-10", "categories": []}
{"id": "2407.08571", "title": "Multi-Group Proportional Representation in Retrieval", "abstract": "  Image search and retrieval tasks can perpetuate harmful stereotypes, erase\ncultural identities, and amplify social disparities. Current approaches to\nmitigate these representational harms balance the number of retrieved items\nacross population groups defined by a small number of (often binary)\nattributes. However, most existing methods overlook intersectional groups\ndetermined by combinations of group attributes, such as gender, race, and\nethnicity. We introduce Multi-Group Proportional Representation (MPR), a novel\nmetric that measures representation across intersectional groups. We develop\npractical methods for estimating MPR, provide theoretical guarantees, and\npropose optimization algorithms to ensure MPR in retrieval. We demonstrate that\nexisting methods optimizing for equal and proportional representation metrics\nmay fail to promote MPR. Crucially, our work shows that optimizing MPR yields\nmore proportional representation across multiple intersectional groups\nspecified by a rich function class, often with minimal compromise in retrieval\naccuracy.\n", "date": "2024-07-11", "categories": []}
{"id": "2411.00360", "title": "A Simple Remedy for Dataset Bias via Self-Influence: A Mislabeled Sample\n  Perspective", "abstract": "  Learning generalized models from biased data is an important undertaking\ntoward fairness in deep learning. To address this issue, recent studies attempt\nto identify and leverage bias-conflicting samples free from spurious\ncorrelations without prior knowledge of bias or an unbiased set. However,\nspurious correlation remains an ongoing challenge, primarily due to the\ndifficulty in precisely detecting these samples. In this paper, inspired by the\nsimilarities between mislabeled samples and bias-conflicting samples, we\napproach this challenge from a novel perspective of mislabeled sample\ndetection. Specifically, we delve into Influence Function, one of the standard\nmethods for mislabeled sample detection, for identifying bias-conflicting\nsamples and propose a simple yet effective remedy for biased models by\nleveraging them. Through comprehensive analysis and experiments on diverse\ndatasets, we demonstrate that our new perspective can boost the precision of\ndetection and rectify biased models effectively. Furthermore, our approach is\ncomplementary to existing methods, showing performance improvement even when\napplied to models that have already undergone recent debiasing techniques.\n", "date": "2024-11-01", "categories": []}
{"id": "2411.00711", "title": "Debiasify: Self-Distillation for Unsupervised Bias Mitigation", "abstract": "  Simplicity bias poses a significant challenge in neural networks, often\nleading models to favor simpler solutions and inadvertently learn decision\nrules influenced by spurious correlations. This results in biased models with\ndiminished generalizability. While many current approaches depend on human\nsupervision, obtaining annotations for various bias attributes is often\nimpractical. To address this, we introduce Debiasify, a novel self-distillation\napproach that requires no prior knowledge about the nature of biases. Our\nmethod leverages a new distillation loss to transfer knowledge within the\nnetwork, from deeper layers containing complex, highly-predictive features to\nshallower layers with simpler, attribute-conditioned features in an\nunsupervised manner. This enables Debiasify to learn robust, debiased\nrepresentations that generalize effectively across diverse biases and datasets,\nimproving both worst-group performance and overall accuracy. Extensive\nexperiments on computer vision and medical imaging benchmarks demonstrate the\neffectiveness of our approach, significantly outperforming previous\nunsupervised debiasing methods (e.g., a 10.13% improvement in worst-group\naccuracy for Wavy Hair classification in CelebA) and achieving comparable or\nsuperior performance to supervised approaches. Our code is publicly available\nat the following link: Debiasify.\n", "date": "2024-11-01", "categories": []}
{"id": "2302.02910", "title": "An Empirical Analysis of Fairness Notions under Differential Privacy", "abstract": "  Recent works have shown that selecting an optimal model architecture suited\nto the differential privacy setting is necessary to achieve the best possible\nutility for a given privacy budget using differentially private stochastic\ngradient descent (DP-SGD)(Tramer and Boneh 2020; Cheng et al. 2022). In light\nof these findings, we empirically analyse how different fairness notions,\nbelonging to distinct classes of statistical fairness criteria (independence,\nseparation and sufficiency), are impacted when one selects a model architecture\nsuitable for DP-SGD, optimized for utility. Using standard datasets from ML\nfairness literature, we show using a rigorous experimental protocol, that by\nselecting the optimal model architecture for DP-SGD, the differences across\ngroups concerning the relevant fairness metrics (demographic parity, equalized\nodds and predictive parity) more often decrease or are negligibly impacted,\ncompared to the non-private baseline, for which optimal model architecture has\nalso been selected to maximize utility. These findings challenge the\nunderstanding that differential privacy will necessarily exacerbate unfairness\nin deep learning models trained on biased datasets.\n", "date": "2023-02-06", "categories": []}
{"id": "2410.22382", "title": "Debiasing Alternative Data for Credit Underwriting Using Causal\n  Inference", "abstract": "  Alternative data provides valuable insights for lenders to evaluate a\nborrower's creditworthiness, which could help expand credit access to\nunderserved groups and lower costs for borrowers. But some forms of alternative\ndata have historically been excluded from credit underwriting because it could\nact as an illegal proxy for a protected class like race or gender, causing\nredlining. We propose a method for applying causal inference to a supervised\nmachine learning model to debias alternative data so that it might be used for\ncredit underwriting. We demonstrate how our algorithm can be used against a\npublic credit dataset to improve model accuracy across different racial groups,\nwhile providing theoretically robust nondiscrimination guarantees.\n", "date": "2024-10-29", "categories": []}
{"id": "2309.01610", "title": "Fairness in Ranking under Disparate Uncertainty", "abstract": "  Ranking is a ubiquitous method for focusing the attention of human evaluators\non a manageable subset of options. Its use as part of human decision-making\nprocesses ranges from surfacing potentially relevant products on an e-commerce\nsite to prioritizing college applications for human review. While ranking can\nmake human evaluation more effective by focusing attention on the most\npromising options, we argue that it can introduce unfairness if the uncertainty\nof the underlying relevance model differs between groups of options.\nUnfortunately, such disparity in uncertainty appears widespread, often to the\ndetriment of minority groups for which relevance estimates can have higher\nuncertainty due to a lack of data or appropriate features. To address this\nfairness issue, we propose Equal-Opportunity Ranking (EOR) as a new fairness\ncriterion for ranking and show that it corresponds to a group-wise fair lottery\namong the relevant options even in the presence of disparate uncertainty. EOR\noptimizes for an even cost burden on all groups, unlike the conventional\nProbability Ranking Principle, and is fundamentally different from existing\nnotions of fairness in rankings, such as demographic parity and proportional\nRooney rule constraints that are motivated by proportional representation\nrelative to group size. To make EOR ranking practical, we present an efficient\nalgorithm for computing it in time $O(n \\log(n))$ and prove its close\napproximation guarantee to the globally optimal solution. In a comprehensive\nempirical evaluation on synthetic data, a US Census dataset, and a real-world\naudit of Amazon search queries, we find that the algorithm reliably guarantees\nEOR fairness while providing effective rankings.\n", "date": "2023-09-04", "categories": []}
{"id": "2311.05436", "title": "Fair Wasserstein Coresets", "abstract": "  Data distillation and coresets have emerged as popular approaches to generate\na smaller representative set of samples for downstream learning tasks to handle\nlarge-scale datasets. At the same time, machine learning is being increasingly\napplied to decision-making processes at a societal level, making it imperative\nfor modelers to address inherent biases towards subgroups present in the data.\nWhile current approaches focus on creating fair synthetic representative\nsamples by optimizing local properties relative to the original samples, their\nimpact on downstream learning processes has yet to be explored. In this work,\nwe present fair Wasserstein coresets (FWC), a novel coreset approach which\ngenerates fair synthetic representative samples along with sample-level weights\nto be used in downstream learning tasks. FWC uses an efficient majority\nminimization algorithm to minimize the Wasserstein distance between the\noriginal dataset and the weighted synthetic samples while enforcing demographic\nparity. We show that an unconstrained version of FWC is equivalent to Lloyd's\nalgorithm for k-medians and k-means clustering. Experiments conducted on both\nsynthetic and real datasets show that FWC: (i) achieves a competitive\nfairness-utility tradeoff in downstream models compared to existing approaches,\n(ii) improves downstream fairness when added to the existing training data and\n(iii) can be used to reduce biases in predictions from large language models\n(GPT-3.5 and GPT-4).\n", "date": "2023-11-09", "categories": []}
{"id": "2405.05809", "title": "Aequitas Flow: Streamlining Fair ML Experimentation", "abstract": "  Aequitas Flow is an open-source framework and toolkit for end-to-end Fair\nMachine Learning (ML) experimentation, and benchmarking in Python. This package\nfills integration gaps that exist in other fair ML packages. In addition to the\nexisting audit capabilities in Aequitas, the Aequitas Flow module provides a\npipeline for fairness-aware model training, hyperparameter optimization, and\nevaluation, enabling easy-to-use and rapid experiments and analysis of results.\nAimed at ML practitioners and researchers, the framework offers implementations\nof methods, datasets, metrics, and standard interfaces for these components to\nimprove extensibility. By facilitating the development of fair ML practices,\nAequitas Flow hopes to enhance the incorporation of fairness concepts in AI\nsystems making AI systems more robust and fair.\n", "date": "2024-05-09", "categories": []}
{"id": "2410.16432", "title": "Fair Bilevel Neural Network (FairBiNN): On Balancing fairness and\n  accuracy via Stackelberg Equilibrium", "abstract": "  The persistent challenge of bias in machine learning models necessitates\nrobust solutions to ensure parity and equal treatment across diverse groups,\nparticularly in classification tasks. Current methods for mitigating bias often\nresult in information loss and an inadequate balance between accuracy and\nfairness. To address this, we propose a novel methodology grounded in bilevel\noptimization principles. Our deep learning-based approach concurrently\noptimizes for both accuracy and fairness objectives, and under certain\nassumptions, achieving proven Pareto optimal solutions while mitigating bias in\nthe trained model. Theoretical analysis indicates that the upper bound on the\nloss incurred by this method is less than or equal to the loss of the\nLagrangian approach, which involves adding a regularization term to the loss\nfunction. We demonstrate the efficacy of our model primarily on tabular\ndatasets such as UCI Adult and Heritage Health. When benchmarked against\nstate-of-the-art fairness methods, our model exhibits superior performance,\nadvancing fairness-aware machine learning solutions and bridging the\naccuracy-fairness gap. The implementation of FairBiNN is available on\nhttps://github.com/yazdanimehdi/FairBiNN.\n", "date": "2024-10-21", "categories": []}
{"id": "2410.22517", "title": "Attention Speaks Volumes: Localizing and Mitigating Bias in Language\n  Models", "abstract": "  We explore the internal mechanisms of how bias emerges in large language\nmodels (LLMs) when provided with ambiguous comparative prompts: inputs that\ncompare or enforce choosing between two or more entities without providing\nclear context for preference. Most approaches for bias mitigation focus on\neither post-hoc analysis or data augmentation. However, these are transient\nsolutions, without addressing the root cause: the model itself. Numerous prior\nworks show the influence of the attention module towards steering generations.\nWe believe that analyzing attention is also crucial for understanding bias, as\nit provides insight into how the LLM distributes its focus across different\nentities and how this contributes to biased decisions. To this end, we first\nintroduce a metric to quantify the LLM's preference for one entity over\nanother. We then propose $\\texttt{ATLAS}$ (Attention-based Targeted Layer\nAnalysis and Scaling), a technique to localize bias to specific layers of the\nLLM by analyzing attention scores and then reduce bias by scaling attention in\nthese biased layers. To evaluate our method, we conduct experiments across 3\ndatasets (BBQ, Crows-Pairs, and WinoGender) using $\\texttt{GPT-2 XL}$ (1.5B),\n$\\texttt{GPT-J}$ (6B), $\\texttt{LLaMA-2}$ (7B) and $\\texttt{LLaMA-3}$ (8B). Our\nexperiments demonstrate that bias is concentrated in the later layers,\ntypically around the last third. We also show how $\\texttt{ATLAS}$ effectively\nmitigates bias through targeted interventions without compromising downstream\nperformance and an average increase of only 0.82% in perplexity when the\nintervention is applied. We see an average improvement of 0.28 points in the\nbias score across all the datasets.\n", "date": "2024-10-29", "categories": []}
{"id": "2410.23273", "title": "Proportional Fairness in Non-Centroid Clustering", "abstract": "  We revisit the recently developed framework of proportionally fair\nclustering, where the goal is to provide group fairness guarantees that become\nstronger for groups of data points (agents) that are large and cohesive. Prior\nwork applies this framework to centroid clustering, where the loss of an agent\nis its distance to the centroid assigned to its cluster. We expand the\nframework to non-centroid clustering, where the loss of an agent is a function\nof the other agents in its cluster, by adapting two proportional fairness\ncriteria -- the core and its relaxation, fully justified representation (FJR)\n-- to this setting.\n  We show that the core can be approximated only under structured loss\nfunctions, and even then, the best approximation we are able to establish,\nusing an adaptation of the GreedyCapture algorithm developed for centroid\nclustering [Chen et al., 2019; Micha and Shah, 2020], is unappealing for a\nnatural loss function. In contrast, we design a new (inefficient) algorithm,\nGreedyCohesiveClustering, which achieves the relaxation FJR exactly under\narbitrary loss functions, and show that the efficient GreedyCapture algorithm\nachieves a constant approximation of FJR. We also design an efficient auditing\nalgorithm, which estimates the FJR approximation of any given clustering\nsolution up to a constant factor. Our experiments on real data suggest that\ntraditional clustering algorithms are highly unfair, whereas GreedyCapture is\nconsiderably fairer and incurs only a modest loss in common clustering\nobjectives.\n", "date": "2024-10-30", "categories": []}
{"id": "2312.02959", "title": "Detecting algorithmic bias in medical-AI models using trees", "abstract": "  With the growing prevalence of machine learning and artificial\nintelligence-based medical decision support systems, it is equally important to\nensure that these systems provide patient outcomes in a fair and equitable\nfashion. This paper presents an innovative framework for detecting areas of\nalgorithmic bias in medical-AI decision support systems. Our approach\nefficiently identifies potential biases in medical-AI models, specifically in\nthe context of sepsis prediction, by employing the Classification and\nRegression Trees (CART) algorithm with conformity scores. We verify our\nmethodology by conducting a series of synthetic data experiments, showcasing\nits ability to estimate areas of bias in controlled settings precisely. The\neffectiveness of the concept is further validated by experiments using\nelectronic medical records from Grady Memorial Hospital in Atlanta, Georgia.\nThese tests demonstrate the practical implementation of our strategy in a\nclinical environment, where it can function as a vital instrument for\nguaranteeing fairness and equity in AI-based medical decisions.\n", "date": "2023-12-05", "categories": []}
{"id": "2410.19775", "title": "Gender Bias of LLM in Economics: An Existentialism Perspective", "abstract": "  Large Language Models (LLMs), such as GPT-4 and BERT, have rapidly gained\ntraction in natural language processing (NLP) and are now integral to financial\ndecision-making. However, their deployment introduces critical challenges,\nparticularly in perpetuating gender biases that can distort decision-making\noutcomes in high-stakes economic environments. This paper investigates gender\nbias in LLMs through both mathematical proofs and empirical experiments using\nthe Word Embedding Association Test (WEAT), demonstrating that LLMs inherently\nreinforce gender stereotypes even without explicit gender markers. By comparing\nthe decision-making processes of humans and LLMs, we reveal fundamental\ndifferences: while humans can override biases through ethical reasoning and\nindividualized understanding, LLMs maintain bias as a rational outcome of their\nmathematical optimization on biased data. Our analysis proves that bias in LLMs\nis not an unintended flaw but a systematic result of their rational processing,\nwhich tends to preserve and amplify existing societal biases encoded in\ntraining data. Drawing on existentialist theory, we argue that LLM-generated\nbias reflects entrenched societal structures and highlights the limitations of\npurely technical debiasing methods. This research underscores the need for new\ntheoretical frameworks and interdisciplinary methodologies that address the\nethical implications of integrating LLMs into economic and financial\ndecision-making. We advocate for a reconceptualization of how LLMs influence\neconomic decisions, emphasizing the importance of incorporating human-like\nethical considerations into AI governance to ensure fairness and equity in\nAI-driven financial systems.\n", "date": "2024-10-14", "categories": []}
{"id": "2410.20265", "title": "You Never Know: Quantization Induces Inconsistent Biases in\n  Vision-Language Foundation Models", "abstract": "  We study the impact of a standard practice in compressing foundation\nvision-language models - quantization - on the models' ability to produce\nsocially-fair outputs. In contrast to prior findings with unimodal models that\ncompression consistently amplifies social biases, our extensive evaluation of\nfour quantization settings across three datasets and three CLIP variants yields\na surprising result: while individual models demonstrate bias, we find no\nconsistent change in bias magnitude or direction across a population of\ncompressed models due to quantization.\n", "date": "2024-10-26", "categories": []}
{"id": "2410.21126", "title": "Current State-of-the-Art of Bias Detection and Mitigation in Machine\n  Translation for African and European Languages: a Review", "abstract": "  Studying bias detection and mitigation methods in natural language processing\nand the particular case of machine translation is highly relevant, as societal\nstereotypes might be reflected or reinforced by these systems. In this paper,\nwe analyze the state-of-the-art with a particular focus on European and African\nlanguages. We show how the majority of the work in this field concentrates on\nfew languages, and that there is potential for future research to cover also\nthe less investigated languages to contribute to more diversity in the research\nfield.\n", "date": "2024-10-28", "categories": []}
{"id": "2304.14252", "title": "FLAC: Fairness-Aware Representation Learning by Suppressing\n  Attribute-Class Associations", "abstract": "  Bias in computer vision systems can perpetuate or even amplify discrimination\nagainst certain populations. Considering that bias is often introduced by\nbiased visual datasets, many recent research efforts focus on training fair\nmodels using such data. However, most of them heavily rely on the availability\nof protected attribute labels in the dataset, which limits their applicability,\nwhile label-unaware approaches, i.e., approaches operating without such labels,\nexhibit considerably lower performance. To overcome these limitations, this\nwork introduces FLAC, a methodology that minimizes mutual information between\nthe features extracted by the model and a protected attribute, without the use\nof attribute labels. To do that, FLAC proposes a sampling strategy that\nhighlights underrepresented samples in the dataset, and casts the problem of\nlearning fair representations as a probability matching problem that leverages\nrepresentations extracted by a bias-capturing classifier. It is theoretically\nshown that FLAC can indeed lead to fair representations, that are independent\nof the protected attributes. FLAC surpasses the current state-of-the-art on\nBiased-MNIST, CelebA, and UTKFace, by 29.1%, 18.1%, and 21.9%, respectively.\nAdditionally, FLAC exhibits 2.2% increased accuracy on ImageNet-A and up to\n4.2% increased accuracy on Corrupted-Cifar10. Finally, in most experiments,\nFLAC even outperforms the bias label-aware state-of-the-art methods.\n", "date": "2023-04-27", "categories": []}
{"id": "2410.19067", "title": "Less Discriminatory Alternative and Interpretable XGBoost Framework for\n  Binary Classification", "abstract": "  Fair lending practices and model interpretability are crucial concerns in the\nfinancial industry, especially given the increasing use of complex machine\nlearning models. In response to the Consumer Financial Protection Bureau's\n(CFPB) requirement to protect consumers against unlawful discrimination, we\nintroduce LDA-XGB1, a novel less discriminatory alternative (LDA) machine\nlearning model for fair and interpretable binary classification. LDA-XGB1 is\ndeveloped through biobjective optimization that balances accuracy and fairness,\nwith both objectives formulated using binning and information value. It\nleverages the predictive power and computational efficiency of XGBoost while\nensuring inherent model interpretability, including the enforcement of\nmonotonic constraints. We evaluate LDA-XGB1 on two datasets: SimuCredit, a\nsimulated credit approval dataset, and COMPAS, a real-world recidivism\nprediction dataset. Our results demonstrate that LDA-XGB1 achieves an effective\nbalance between predictive accuracy, fairness, and interpretability, often\noutperforming traditional fair lending models. This approach equips financial\ninstitutions with a powerful tool to meet regulatory requirements for fair\nlending while maintaining the advantages of advanced machine learning\ntechniques.\n", "date": "2024-10-24", "categories": []}
{"id": "2311.00109", "title": "FairWASP: Fast and Optimal Fair Wasserstein Pre-processing", "abstract": "  Recent years have seen a surge of machine learning approaches aimed at\nreducing disparities in model outputs across different subgroups. In many\nsettings, training data may be used in multiple downstream applications by\ndifferent users, which means it may be most effective to intervene on the\ntraining data itself. In this work, we present FairWASP, a novel pre-processing\napproach designed to reduce disparities in classification datasets without\nmodifying the original data. FairWASP returns sample-level weights such that\nthe reweighted dataset minimizes the Wasserstein distance to the original\ndataset while satisfying (an empirical version of) demographic parity, a\npopular fairness criterion. We show theoretically that integer weights are\noptimal, which means our method can be equivalently understood as duplicating\nor eliminating samples. FairWASP can therefore be used to construct datasets\nwhich can be fed into any classification method, not just methods which accept\nsample weights. Our work is based on reformulating the pre-processing task as a\nlarge-scale mixed-integer program (MIP), for which we propose a highly\nefficient algorithm based on the cutting plane method. Experiments demonstrate\nthat our proposed optimization algorithm significantly outperforms\nstate-of-the-art commercial solvers in solving both the MIP and its linear\nprogram relaxation. Further experiments highlight the competitive performance\nof FairWASP in reducing disparities while preserving accuracy in downstream\nclassification settings.\n", "date": "2023-10-31", "categories": []}
{"id": "2406.04562", "title": "A Unified View of Group Fairness Tradeoffs Using Partial Information\n  Decomposition", "abstract": "  This paper introduces a novel information-theoretic perspective on the\nrelationship between prominent group fairness notions in machine learning,\nnamely statistical parity, equalized odds, and predictive parity. It is well\nknown that simultaneous satisfiability of these three fairness notions is\nusually impossible, motivating practitioners to resort to approximate fairness\nsolutions rather than stringent satisfiability of these definitions. However, a\ncomprehensive analysis of their interrelations, particularly when they are not\nexactly satisfied, remains largely unexplored. Our main contribution lies in\nelucidating an exact relationship between these three measures of (un)fairness\nby leveraging a body of work in information theory called partial information\ndecomposition (PID). In this work, we leverage PID to identify the granular\nregions where these three measures of (un)fairness overlap and where they\ndisagree with each other leading to potential tradeoffs. We also include\nnumerical simulations to complement our results.\n", "date": "2024-06-07", "categories": []}
{"id": "2407.11002", "title": "MoESD: Mixture of Experts Stable Diffusion to Mitigate Gender Bias", "abstract": "  Text-to-image models are known to propagate social biases. For example, when\nprompted to generate images of people in certain professions, these models tend\nto systematically generate specific genders or ethnicities. In this paper, we\nshow that this bias is already present in the text encoder of the model and\nintroduce a Mixture-of-Experts approach by identifying text-encoded bias in the\nlatent space and then creating a Bias-Identification Gate mechanism. More\nspecifically, we propose MoESD (Mixture of Experts Stable Diffusion) with BiAs\n(Bias Adapters) to mitigate gender bias in text-to-image models. We also\ndemonstrate that introducing an arbitrary special token to the prompt is\nessential during the mitigation process. With experiments focusing on gender\nbias, we show that our approach successfully mitigates gender bias while\nmaintaining image quality.\n", "date": "2024-06-25", "categories": []}
{"id": "2410.18749", "title": "Does Differential Privacy Impact Bias in Pretrained NLP Models?", "abstract": "  Differential privacy (DP) is applied when fine-tuning pre-trained large\nlanguage models (LLMs) to limit leakage of training examples. While most DP\nresearch has focused on improving a model's privacy-utility tradeoff, some find\nthat DP can be unfair to or biased against underrepresented groups. In this\nwork, we show the impact of DP on bias in LLMs through empirical analysis.\nDifferentially private training can increase the model bias against protected\ngroups w.r.t AUC-based bias metrics. DP makes it more difficult for the model\nto differentiate between the positive and negative examples from the protected\ngroups and other groups in the rest of the population. Our results also show\nthat the impact of DP on bias is not only affected by the privacy protection\nlevel but also the underlying distribution of the dataset.\n", "date": "2024-10-24", "categories": []}
{"id": "2410.18841", "title": "From Efficiency to Equity: Measuring Fairness in Preference Learning", "abstract": "  As AI systems, particularly generative models, increasingly influence\ndecision-making, ensuring that they are able to fairly represent diverse human\npreferences becomes crucial. This paper introduces a novel framework for\nevaluating epistemic fairness in preference learning models inspired by\neconomic theories of inequality and Rawlsian justice. We propose metrics\nadapted from the Gini Coefficient, Atkinson Index, and Kuznets Ratio to\nquantify fairness in these models. We validate our approach using two datasets:\na custom visual preference dataset (AI-EDI-Space) and the Jester Jokes dataset.\nOur analysis reveals variations in model performance across users, highlighting\npotential epistemic injustices. We explore pre-processing and in-processing\ntechniques to mitigate these inequalities, demonstrating a complex relationship\nbetween model efficiency and fairness. This work contributes to AI ethics by\nproviding a framework for evaluating and improving epistemic fairness in\npreference learning models, offering insights for developing more inclusive AI\nsystems in contexts where diverse human preferences are crucial.\n", "date": "2024-10-24", "categories": []}
{"id": "2406.16592", "title": "Toward Fairer Face Recognition Datasets", "abstract": "  Face recognition and verification are two computer vision tasks whose\nperformance has progressed with the introduction of deep representations.\nHowever, ethical, legal, and technical challenges due to the sensitive\ncharacter of face data and biases in real training datasets hinder their\ndevelopment. Generative AI addresses privacy by creating fictitious identities,\nbut fairness problems persist. We promote fairness by introducing a demographic\nattributes balancing mechanism in generated training datasets. We experiment\nwith an existing real dataset, three generated training datasets, and the\nbalanced versions of a diffusion-based dataset. We propose a comprehensive\nevaluation that considers accuracy and fairness equally and includes a rigorous\nregression-based statistical analysis of attributes. The analysis shows that\nbalancing reduces demographic unfairness. Also, a performance gap persists\ndespite generation becoming more accurate with time. The proposed balancing\nmethod and comprehensive verification evaluation promote fairer and transparent\nface recognition and verification.\n", "date": "2024-06-24", "categories": []}
{"id": "2410.17358", "title": "FairLoRA: Unpacking Bias Mitigation in Vision Models with\n  Fairness-Driven Low-Rank Adaptation", "abstract": "  Recent advances in parameter-efficient fine-tuning methods, such as Low Rank\nAdaptation (LoRA), have gained significant attention for their ability to\nefficiently adapt large foundational models to various downstream tasks. These\nmethods are appreciated for achieving performance comparable to full\nfine-tuning on aggregate-level metrics, while significantly reducing\ncomputational costs. To systematically address fairness in LLMs previous\nstudies fine-tune on fairness specific data using a larger LoRA rank than\ntypically used. In this paper, we introduce FairLoRA, a novel fairness-specific\nregularizer for LoRA aimed at reducing performance disparities across data\nsubgroups by minimizing per-class variance in loss. To the best of our\nknowledge, we are the first to introduce a fairness based finetuning through\nLoRA. Our results demonstrate that the need for higher ranks to mitigate bias\nis not universal; it depends on factors such as the pre-trained model, dataset,\nand task. More importantly, we systematically evaluate FairLoRA across various\nvision models, including ViT, DiNO, and CLIP, in scenarios involving\ndistribution shifts. We further emphasize the necessity of using multiple\nfairness metrics to obtain a holistic assessment of fairness, rather than\nrelying solely on the metric optimized during training.\n", "date": "2024-10-22", "categories": []}
{"id": "2410.17415", "title": "End-to-End Optimization and Learning of Fair Court Schedules", "abstract": "  Criminal courts across the United States handle millions of cases every year,\nand the scheduling of those cases must accommodate a diverse set of\nconstraints, including the preferences and availability of courts, prosecutors,\nand defense teams. When criminal court schedules are formed, defendants'\nscheduling preferences often take the least priority, although defendants may\nface significant consequences (including arrest or detention) for missed court\ndates. Additionally, studies indicate that defendants' nonappearances impose\ncosts on the courts and other system stakeholders. To address these issues,\ncourts and commentators have begun to recognize that pretrial outcomes for\ndefendants and for the system would be improved with greater attention to court\nprocesses, including \\emph{court scheduling practices}. There is thus a need\nfor fair criminal court pretrial scheduling systems that account for\ndefendants' preferences and availability, but the collection of such data poses\nlogistical challenges. Furthermore, optimizing schedules fairly across various\nparties' preferences is a complex optimization problem, even when such data is\navailable. In an effort to construct such a fair scheduling system under data\nuncertainty, this paper proposes a joint optimization and learning framework\nthat combines machine learning models trained end-to-end with efficient\nmatching algorithms. This framework aims to produce court scheduling schedules\nthat optimize a principled measure of fairness, balancing the availability and\npreferences of all parties.\n", "date": "2024-10-22", "categories": []}
{"id": "2410.17433", "title": "Revisiting Technical Bias Mitigation Strategies", "abstract": "  Efforts to mitigate bias and enhance fairness in the artificial intelligence\n(AI) community have predominantly focused on technical solutions. While\nnumerous reviews have addressed bias in AI, this review uniquely focuses on the\npractical limitations of technical solutions in healthcare settings, providing\na structured analysis across five key dimensions affecting their real-world\nimplementation: who defines bias and fairness; which mitigation strategy to use\nand prioritize among dozens that are inconsistent and incompatible; when in the\nAI development stages the solutions are most effective; for which populations;\nand the context in which the solutions are designed. We illustrate each\nlimitation with empirical studies focusing on healthcare and biomedical\napplications. Moreover, we discuss how value-sensitive AI, a framework derived\nfrom technology design, can engage stakeholders and ensure that their values\nare embodied in bias and fairness mitigation solutions. Finally, we discuss\nareas that require further investigation and provide practical recommendations\nto address the limitations covered in the study.\n", "date": "2024-10-22", "categories": []}
{"id": "2410.17492", "title": "BadFair: Backdoored Fairness Attacks with Group-conditioned Triggers", "abstract": "  Attacking fairness is crucial because compromised models can introduce biased\noutcomes, undermining trust and amplifying inequalities in sensitive\napplications like hiring, healthcare, and law enforcement. This highlights the\nurgent need to understand how fairness mechanisms can be exploited and to\ndevelop defenses that ensure both fairness and robustness. We introduce\nBadFair, a novel backdoored fairness attack methodology. BadFair stealthily\ncrafts a model that operates with accuracy and fairness under regular\nconditions but, when activated by certain triggers, discriminates and produces\nincorrect results for specific groups. This type of attack is particularly\nstealthy and dangerous, as it circumvents existing fairness detection methods,\nmaintaining an appearance of fairness in normal use. Our findings reveal that\nBadFair achieves a more than 85% attack success rate in attacks aimed at target\ngroups on average while only incurring a minimal accuracy loss. Moreover, it\nconsistently exhibits a significant discrimination score, distinguishing\nbetween pre-defined target and non-target attacked groups across various\ndatasets and models.\n", "date": "2024-10-23", "categories": []}
{"id": "2410.17555", "title": "FairDgcl: Fairness-aware Recommendation with Dynamic Graph Contrastive\n  Learning", "abstract": "  As trustworthy AI continues to advance, the fairness issue in recommendations\nhas received increasing attention. A recommender system is considered unfair\nwhen it produces unequal outcomes for different user groups based on\nuser-sensitive attributes (e.g., age, gender). Some researchers have proposed\ndata augmentation-based methods aiming at alleviating user-level unfairness by\naltering the skewed distribution of training data among various user groups.\nDespite yielding promising results, they often rely on fairness-related\nassumptions that may not align with reality, potentially reducing the data\nquality and negatively affecting model effectiveness. To tackle this issue, in\nthis paper, we study how to implement high-quality data augmentation to improve\nrecommendation fairness. Specifically, we propose FairDgcl, a dynamic graph\nadversarial contrastive learning framework aiming at improving fairness in\nrecommender system. First, FairDgcl develops an adversarial contrastive network\nwith a view generator and a view discriminator to learn generating fair\naugmentation strategies in an adversarial style. Then, we propose two dynamic,\nlearnable models to generate contrastive views within contrastive learning\nframework, which automatically fine-tune the augmentation strategies.\nMeanwhile, we theoretically show that FairDgcl can simultaneously generate\nenhanced representations that possess both fairness and accuracy. Lastly,\ncomprehensive experiments conducted on four real-world datasets demonstrate the\neffectiveness of the proposed FairDgcl.\n", "date": "2024-10-23", "categories": []}
{"id": "2311.18460", "title": "Causal Fairness under Unobserved Confounding: A Neural Sensitivity\n  Framework", "abstract": "  Fairness for machine learning predictions is widely required in practice for\nlegal, ethical, and societal reasons. Existing work typically focuses on\nsettings without unobserved confounding, even though unobserved confounding can\nlead to severe violations of causal fairness and, thus, unfair predictions. In\nthis work, we analyze the sensitivity of causal fairness to unobserved\nconfounding. Our contributions are three-fold. First, we derive bounds for\ncausal fairness metrics under different sources of unobserved confounding. This\nenables practitioners to examine the sensitivity of their machine learning\nmodels to unobserved confounding in fairness-critical applications. Second, we\npropose a novel neural framework for learning fair predictions, which allows us\nto offer worst-case guarantees of the extent to which causal fairness can be\nviolated due to unobserved confounding. Third, we demonstrate the effectiveness\nof our framework in a series of experiments, including a real-world case study\nabout predicting prison sentences. To the best of our knowledge, ours is the\nfirst work to study causal fairness under unobserved confounding. To this end,\nour work is of direct practical value as a refutation strategy to ensure the\nfairness of predictions in high-stakes applications.\n", "date": "2023-11-30", "categories": []}
{"id": "2409.13705", "title": "Debiasing Text Safety Classifiers through a Fairness-Aware Ensemble", "abstract": "  Increasing use of large language models (LLMs) demand performant guardrails\nto ensure the safety of inputs and outputs of LLMs. When these safeguards are\ntrained on imbalanced data, they can learn the societal biases. We present a\nlight-weight, post-processing method for mitigating counterfactual fairness in\nclosed-source text safety classifiers. Our approach involves building an\nensemble that not only outperforms the input classifiers and policy-aligns\nthem, but also acts as a debiasing regularizer. We introduce two\nthreshold-agnostic metrics to assess the counterfactual fairness of a model,\nand demonstrate how combining these metrics with Fair Data Reweighting (FDW)\nhelps mitigate biases. We create an expanded Open AI dataset, and a new\ntemplated LLM-generated dataset based on user-prompts, both of which are\ncounterfactually balanced across identity groups and cover four key areas of\nsafety; we will work towards publicly releasing these datasets. Our results\nshow that our approach improves counterfactual fairness with minimal impact on\nmodel performance.\n", "date": "2024-09-05", "categories": []}
{"id": "2409.16965", "title": "ABCFair: an Adaptable Benchmark approach for Comparing Fairness Methods", "abstract": "  Numerous methods have been implemented that pursue fairness with respect to\nsensitive features by mitigating biases in machine learning. Yet, the problem\nsettings that each method tackles vary significantly, including the stage of\nintervention, the composition of sensitive features, the fairness notion, and\nthe distribution of the output. Even in binary classification, these subtle\ndifferences make it highly complicated to benchmark fairness methods, as their\nperformance can strongly depend on exactly how the bias mitigation problem was\noriginally framed.\n  Hence, we introduce ABCFair, a benchmark approach which allows adapting to\nthe desiderata of the real-world problem setting, enabling proper comparability\nbetween methods for any use case. We apply ABCFair to a range of pre-, in-, and\npostprocessing methods on both large-scale, traditional datasets and on a dual\nlabel (biased and unbiased) dataset to sidestep the fairness-accuracy\ntrade-off.\n", "date": "2024-09-25", "categories": []}
{"id": "2410.11493", "title": "Towards Fair Graph Representation Learning in Social Networks", "abstract": "  With the widespread use of Graph Neural Networks (GNNs) for representation\nlearning from network data, the fairness of GNN models has raised great\nattention lately. Fair GNNs aim to ensure that node representations can be\naccurately classified, but not easily associated with a specific group.\nExisting advanced approaches essentially enhance the generalisation of node\nrepresentation in combination with data augmentation strategy, and do not\ndirectly impose constraints on the fairness of GNNs. In this work, we identify\nthat a fundamental reason for the unfairness of GNNs in social network learning\nis the phenomenon of social homophily, i.e., users in the same group are more\ninclined to congregate. The message-passing mechanism of GNNs can cause users\nin the same group to have similar representations due to social homophily,\nleading model predictions to establish spurious correlations with sensitive\nattributes. Inspired by this reason, we propose a method called Equity-Aware\nGNN (EAGNN) towards fair graph representation learning. Specifically, to ensure\nthat model predictions are independent of sensitive attributes while\nmaintaining prediction performance, we introduce constraints for fair\nrepresentation learning based on three principles: sufficiency, independence,\nand separation. We theoretically demonstrate that our EAGNN method can\neffectively achieve group fairness. Extensive experiments on three datasets\nwith varying levels of social homophily illustrate that our EAGNN method\nachieves the state-of-the-art performance across two fairness metrics and\noffers competitive effectiveness.\n", "date": "2024-10-15", "categories": []}
{"id": "2410.16712", "title": "DENOASR: Debiasing ASRs through Selective Denoising", "abstract": "  Automatic Speech Recognition (ASR) systems have been examined and shown to\nexhibit biases toward particular groups of individuals, influenced by factors\nsuch as demographic traits, accents, and speech styles. Noise can\ndisproportionately impact speakers with certain accents, dialects, or speaking\nstyles, leading to biased error rates. In this work, we introduce a novel\nframework DENOASR, which is a selective denoising technique to reduce the\ndisparity in the word error rates between the two gender groups, male and\nfemale. We find that a combination of two popular speech denoising techniques,\nviz. DEMUCS and LE, can be effectively used to mitigate ASR disparity without\ncompromising their overall performance. Experiments using two state-of-the-art\nopen-source ASRs - OpenAI WHISPER and NVIDIA NEMO - on multiple benchmark\ndatasets, including TIE, VOX-POPULI, TEDLIUM, and FLEURS, show that there is a\npromising reduction in the average word error rate gap across the two gender\ngroups. For a given dataset, the denoising is selectively applied on speech\nsamples having speech intelligibility below a certain threshold, estimated\nusing a small validation sample, thus ameliorating the need for large-scale\nhuman-written ground-truth transcripts. Our findings suggest that selective\ndenoising can be an elegant approach to mitigate biases in present-day ASR\nsystems.\n", "date": "2024-10-22", "categories": []}
{"id": "2410.16927", "title": "Revealing Hidden Bias in AI: Lessons from Large Language Models", "abstract": "  As large language models (LLMs) become integral to recruitment processes,\nconcerns about AI-induced bias have intensified. This study examines biases in\ncandidate interview reports generated by Claude 3.5 Sonnet, GPT-4o, Gemini 1.5,\nand Llama 3.1 405B, focusing on characteristics such as gender, race, and age.\nWe evaluate the effectiveness of LLM-based anonymization in reducing these\nbiases. Findings indicate that while anonymization reduces certain biases,\nparticularly gender bias, the degree of effectiveness varies across models and\nbias types. Notably, Llama 3.1 405B exhibited the lowest overall bias.\nMoreover, our methodology of comparing anonymized and non-anonymized data\nreveals a novel approach to assessing inherent biases in LLMs beyond\nrecruitment applications. This study underscores the importance of careful LLM\nselection and suggests best practices for minimizing bias in AI applications,\npromoting fairness and inclusivity.\n", "date": "2024-10-22", "categories": []}
{"id": "2410.13286", "title": "A Human-in-the-Loop Fairness-Aware Model Selection Framework for Complex\n  Fairness Objective Landscapes", "abstract": "  Fairness-aware Machine Learning (FairML) applications are often characterized\nby complex social objectives and legal requirements, frequently involving\nmultiple, potentially conflicting notions of fairness. Despite the well-known\nImpossibility Theorem of Fairness and extensive theoretical research on the\nstatistical and socio-technical trade-offs between fairness metrics, many\nFairML tools still optimize or constrain for a single fairness objective.\nHowever, this one-sided optimization can inadvertently lead to violations of\nother relevant notions of fairness. In this socio-technical and empirical\nstudy, we frame fairness as a many-objective (MaO) problem by treating fairness\nmetrics as conflicting objectives. We introduce ManyFairHPO, a\nhuman-in-the-loop, fairness-aware model selection framework that enables\npractitioners to effectively navigate complex and nuanced fairness objective\nlandscapes. ManyFairHPO aids in the identification, evaluation, and balancing\nof fairness metric conflicts and their related social consequences, leading to\nmore informed and socially responsible model-selection decisions. Through a\ncomprehensive empirical evaluation and a case study on the Law School\nAdmissions problem, we demonstrate the effectiveness of ManyFairHPO in\nbalancing multiple fairness objectives, mitigating risks such as\nself-fulfilling prophecies, and providing interpretable insights to guide\nstakeholders in making fairness-aware modeling decisions.\n", "date": "2024-10-17", "categories": []}
{"id": "2410.15094", "title": "CosFairNet:A Parameter-Space based Approach for Bias Free Learning", "abstract": "  Deep neural networks trained on biased data often inadvertently learn\nunintended inference rules, particularly when labels are strongly correlated\nwith biased features. Existing bias mitigation methods typically involve either\na) predefining bias types and enforcing them as prior knowledge or b)\nreweighting training samples to emphasize bias-conflicting samples over\nbias-aligned samples. However, both strategies address bias indirectly in the\nfeature or sample space, with no control over learned weights, making it\ndifficult to control the bias propagation across different layers. Based on\nthis observation, we introduce a novel approach to address bias directly in the\nmodel's parameter space, preventing its propagation across layers. Our method\ninvolves training two models: a bias model for biased features and a debias\nmodel for unbiased details, guided by the bias model. We enforce dissimilarity\nin the debias model's later layers and similarity in its initial layers with\nthe bias model, ensuring it learns unbiased low-level features without adopting\nbiased high-level abstractions. By incorporating this explicit constraint\nduring training, our approach shows enhanced classification accuracy and\ndebiasing effectiveness across various synthetic and real-world datasets of\ndifferent sizes. Moreover, the proposed method demonstrates robustness across\ndifferent bias types and percentages of biased samples in the training data.\nThe code is available at: https://visdomlab.github.io/CosFairNet/\n", "date": "2024-10-19", "categories": []}
{"id": "2410.15233", "title": "A Semidefinite Relaxation Approach for Fair Graph Clustering", "abstract": "  Fair graph clustering is crucial for ensuring equitable representation and\ntreatment of diverse communities in network analysis. Traditional methods often\nignore disparities among social, economic, and demographic groups, perpetuating\nbiased outcomes and reinforcing inequalities. This study introduces fair graph\nclustering within the framework of the disparate impact doctrine, treating it\nas a joint optimization problem integrating clustering quality and fairness\nconstraints. Given the NP-hard nature of this problem, we employ a semidefinite\nrelaxation approach to approximate the underlying optimization problem. For up\nto medium-sized graphs, we utilize a singular value decomposition-based\nalgorithm, while for larger graphs, we propose a novel algorithm based on the\nalternative direction method of multipliers. Unlike existing methods, our\nformulation allows for tuning the trade-off between clustering quality and\nfairness. Experimental results on graphs generated from the standard stochastic\nblock model demonstrate the superiority of our approach in achieving an optimal\naccuracy-fairness trade-off compared to state-of-the-art methods.\n", "date": "2024-10-19", "categories": []}
{"id": "2410.15470", "title": "Data Augmentation via Diffusion Model to Enhance AI Fairness", "abstract": "  AI fairness seeks to improve the transparency and explainability of AI\nsystems by ensuring that their outcomes genuinely reflect the best interests of\nusers. Data augmentation, which involves generating synthetic data from\nexisting datasets, has gained significant attention as a solution to data\nscarcity. In particular, diffusion models have become a powerful technique for\ngenerating synthetic data, especially in fields like computer vision. This\npaper explores the potential of diffusion models to generate synthetic tabular\ndata to improve AI fairness. The Tabular Denoising Diffusion Probabilistic\nModel (Tab-DDPM), a diffusion model adaptable to any tabular dataset and\ncapable of handling various feature types, was utilized with different amounts\nof generated data for data augmentation. Additionally, reweighting samples from\nAIF360 was employed to further enhance AI fairness. Five traditional machine\nlearning models-Decision Tree (DT), Gaussian Naive Bayes (GNB), K-Nearest\nNeighbors (KNN), Logistic Regression (LR), and Random Forest (RF)-were used to\nvalidate the proposed approach. Experimental results demonstrate that the\nsynthetic data generated by Tab-DDPM improves fairness in binary\nclassification.\n", "date": "2024-10-20", "categories": []}
{"id": "2410.14029", "title": "Auditing and Enforcing Conditional Fairness via Optimal Transport", "abstract": "  Conditional demographic parity (CDP) is a measure of the demographic parity\nof a predictive model or decision process when conditioning on an additional\nfeature or set of features. Many algorithmic fairness techniques exist to\ntarget demographic parity, but CDP is much harder to achieve, particularly when\nthe conditioning variable has many levels and/or when the model outputs are\ncontinuous. The problem of auditing and enforcing CDP is understudied in the\nliterature. In light of this, we propose novel measures of {conditional\ndemographic disparity (CDD)} which rely on statistical distances borrowed from\nthe optimal transport literature. We further design and evaluate\nregularization-based approaches based on these CDD measures. Our methods,\n\\fairbit{} and \\fairlp{}, allow us to target CDP even when the conditioning\nvariable has many levels. When model outputs are continuous, our methods target\nfull equality of the conditional distributions, unlike other methods that only\nconsider first moments or related proxy quantities. We validate the efficacy of\nour approaches on real-world datasets.\n", "date": "2024-10-17", "categories": []}
{"id": "2410.14070", "title": "FaceSaliencyAug: Mitigating Geographic, Gender and Stereotypical Biases\n  via Saliency-Based Data Augmentation", "abstract": "  Geographical, gender and stereotypical biases in computer vision models pose\nsignificant challenges to their performance and fairness. {In this study, we\npresent an approach named FaceSaliencyAug aimed at addressing the gender bias\nin} {Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs).\nLeveraging the salient regions} { of faces detected by saliency, the propose\napproach mitigates geographical and stereotypical biases } {in the datasets.\nFaceSaliencyAug} randomly selects masks from a predefined search space and\napplies them to the salient region of face images, subsequently restoring the\noriginal image with masked salient region. {The proposed} augmentation strategy\nenhances data diversity, thereby improving model performance and debiasing\neffects. We quantify dataset diversity using Image Similarity Score (ISS)\nacross five datasets, including Flickr Faces HQ (FFHQ), WIKI, IMDB, Labelled\nFaces in the Wild (LFW), UTK Faces, and Diverse Dataset. The proposed approach\ndemonstrates superior diversity metrics, as evaluated by ISS-intra and\nISS-inter algorithms. Furthermore, we evaluate the effectiveness of our\napproach in mitigating gender bias on CEO, Engineer, Nurse, and School Teacher\ndatasets. We use the Image-Image Association Score (IIAS) to measure gender\nbias in these occupations. Our experiments reveal a reduction in gender bias\nfor both CNNs and ViTs, indicating the efficacy of our method in promoting\nfairness and inclusivity in computer vision models.\n", "date": "2024-10-17", "categories": []}
{"id": "2410.14201", "title": "Text-to-Image Representativity Fairness Evaluation Framework", "abstract": "  Text-to-Image generative systems are progressing rapidly to be a source of\nadvertisement and media and could soon serve as image searches or artists.\nHowever, there is a significant concern about the representativity bias these\nmodels embody and how these biases can propagate in the social fabric after\nfine-tuning them. Therefore, continuously monitoring and evaluating these\nmodels for fairness is important. To address this issue, we propose\nText-to-Image (TTI) Representativity Fairness Evaluation Framework. In this\nframework, we evaluate three aspects of a TTI system; diversity, inclusion, and\nquality. For each aspect, human-based and model-based approaches are proposed\nand evaluated for their ability to capture the bias and whether they can\nsubstitute each other. The framework starts by suggesting the prompts for\ngenerating the images for the evaluation based on the context and the sensitive\nattributes under study. Then the three aspects are evaluated using the proposed\napproaches. Based on the evaluation, a decision is made regarding the\nrepresentativity bias within the TTI system. The evaluation of our framework on\nStable Diffusion shows that the framework can effectively capture the bias in\nTTI systems. The results also confirm that our proposed model based-approaches\ncan substitute human-based approaches in three out of four components with high\ncorrelation, which could potentially reduce costs and automate the process. The\nstudy suggests that continual learning of the model on more inclusive data\nacross disadvantaged minorities such as Indians and Middle Easterners is\nessential to mitigate current stereotyping and lack of inclusiveness.\n", "date": "2024-10-18", "categories": []}
{"id": "2308.02935", "title": "Bias Behind the Wheel: Fairness Testing of Autonomous Driving Systems", "abstract": "  This paper conducts fairness testing of automated pedestrian detection, a\ncrucial but under-explored issue in autonomous driving systems. We evaluate\neight state-of-the-art deep learning-based pedestrian detectors across\ndemographic groups on large-scale real-world datasets. To enable thorough\nfairness testing, we provide extensive annotations for the datasets, resulting\nin 8,311 images with 16,070 gender labels, 20,115 age labels, and 3,513 skin\ntone labels. Our findings reveal significant fairness issues, particularly\nrelated to age. The proportion of undetected children is 20.14% higher compared\nto adults. Furthermore, we explore how various driving scenarios affect the\nfairness of pedestrian detectors. We find that pedestrian detectors demonstrate\nsignificant gender biases during night time, potentially exacerbating the\nprevalent societal issue of female safety concerns during nighttime out.\nMoreover, we observe that pedestrian detectors can demonstrate both enhanced\nfairness and superior performance under specific driving conditions, which\nchallenges the fairness-performance trade-off theory widely acknowledged in the\nfairness literature. We publicly release the code, data, and results to support\nfuture research on fairness in autonomous driving.\n", "date": "2023-08-05", "categories": []}
{"id": "2410.12864", "title": "Investigating Implicit Bias in Large Language Models: A Large-Scale\n  Study of Over 50 LLMs", "abstract": "  Large Language Models (LLMs) are being adopted across a wide range of tasks,\nincluding decision-making processes in industries where bias in AI systems is a\nsignificant concern. Recent research indicates that LLMs can harbor implicit\nbiases even when they pass explicit bias evaluations. Building upon the\nframeworks of the LLM Implicit Association Test (IAT) Bias and LLM Decision\nBias, this study highlights that newer or larger language models do not\nautomatically exhibit reduced bias; in some cases, they displayed higher bias\nscores than their predecessors, such as in Meta's Llama series and OpenAI's GPT\nmodels. This suggests that increasing model complexity without deliberate bias\nmitigation strategies can unintentionally amplify existing biases. The\nvariability in bias scores within and across providers underscores the need for\nstandardized evaluation metrics and benchmarks for bias assessment. The lack of\nconsistency indicates that bias mitigation is not yet a universally prioritized\ngoal in model development, which can lead to unfair or discriminatory outcomes.\nBy broadening the detection of implicit bias, this research provides a more\ncomprehensive understanding of the biases present in advanced models and\nunderscores the critical importance of addressing these issues to ensure the\ndevelopment of fair and responsible AI systems.\n", "date": "2024-10-13", "categories": []}
{"id": "2410.12913", "title": "Fair Clustering for Data Summarization: Improved Approximation\n  Algorithms and Complexity Insights", "abstract": "  Data summarization tasks are often modeled as $k$-clustering problems, where\nthe goal is to choose $k$ data points, called cluster centers, that best\nrepresent the dataset by minimizing a clustering objective. A popular objective\nis to minimize the maximum distance between any data point and its nearest\ncenter, which is formalized as the $k$-center problem. While in some\napplications all data points can be chosen as centers, in the general setting,\ncenters must be chosen from a predefined subset of points, referred as\nfacilities or suppliers; this is known as the $k$-supplier problem. In this\nwork, we focus on fair data summarization modeled as the fair $k$-supplier\nproblem, where data consists of several groups, and a minimum number of centers\nmust be selected from each group while minimizing the $k$-supplier objective.\nThe groups can be disjoint or overlapping, leading to two distinct problem\nvariants each with different computational complexity.\n  We present $3$-approximation algorithms for both variants, improving the\npreviously known factor of $5$. For disjoint groups, our algorithm runs in\npolynomial time, while for overlapping groups, we present a fixed-parameter\ntractable algorithm, where the exponential runtime depends only on the number\nof groups and centers. We show that these approximation factors match the\ntheoretical lower bounds, assuming standard complexity theory conjectures.\nFinally, using an open-source implementation, we demonstrate the scalability of\nour algorithms on large synthetic datasets and assess the price of fairness on\nreal-world data, comparing solution quality with and without fairness\nconstraints.\n", "date": "2024-10-16", "categories": []}
{"id": "2410.13296", "title": "Fairness-Enhancing Ensemble Classification in Water Distribution\n  Networks", "abstract": "  As relevant examples such as the future criminal detection software [1] show,\nfairness of AI-based and social domain affecting decision support tools\nconstitutes an important area of research. In this contribution, we investigate\nthe applications of AI to socioeconomically relevant infrastructures such as\nthose of water distribution networks (WDNs), where fairness issues have yet to\ngain a foothold. To establish the notion of fairness in this domain, we propose\nan appropriate definition of protected groups and group fairness in WDNs as an\nextension of existing definitions. We demonstrate that typical methods for the\ndetection of leakages in WDNs are unfair in this sense. Further, we thus\npropose a remedy to increase the fairness which can be applied even to\nnon-differentiable ensemble classification methods as used in this context.\n", "date": "2024-10-17", "categories": []}
{"id": "2312.06499", "title": "TaCo: Targeted Concept Erasure Prevents Non-Linear Classifiers From\n  Detecting Protected Attributes", "abstract": "  Ensuring fairness in NLP models is crucial, as they often encode sensitive\nattributes like gender and ethnicity, leading to biased outcomes. Current\nconcept erasure methods attempt to mitigate this by modifying final latent\nrepresentations to remove sensitive information without retraining the entire\nmodel. However, these methods typically rely on linear classifiers, which leave\nmodels vulnerable to non-linear adversaries capable of recovering sensitive\ninformation.\n  We introduce Targeted Concept Erasure (TaCo), a novel approach that removes\nsensitive information from final latent representations, ensuring fairness even\nagainst non-linear classifiers. Our experiments show that TaCo outperforms\nstate-of-the-art methods, achieving greater reductions in the prediction\naccuracy of sensitive attributes by non-linear classifier while preserving\noverall task performance. Code is available on\nhttps://github.com/fanny-jourdan/TaCo.\n", "date": "2023-12-11", "categories": []}
{"id": "2410.12452", "title": "FairGLVQ: Fairness in Partition-Based Classification", "abstract": "  Fairness is an important objective throughout society. From the distribution\nof limited goods such as education, over hiring and payment, to taxes,\nlegislation, and jurisprudence. Due to the increasing importance of machine\nlearning approaches in all areas of daily life including those related to\nhealth, security, and equity, an increasing amount of research focuses on fair\nmachine learning. In this work, we focus on the fairness of partition- and\nprototype-based models. The contribution of this work is twofold: 1) we develop\na general framework for fair machine learning of partition-based models that\ndoes not depend on a specific fairness definition, and 2) we derive a fair\nversion of learning vector quantization (LVQ) as a specific instantiation. We\ncompare the resulting algorithm against other algorithms from the literature on\ntheoretical and real-world data showing its practical relevance.\n", "date": "2024-10-16", "categories": []}
{"id": "2410.12511", "title": "Advancing Fairness in Natural Language Processing: From Traditional\n  Methods to Explainability", "abstract": "  The burgeoning field of Natural Language Processing (NLP) stands at a\ncritical juncture where the integration of fairness within its frameworks has\nbecome an imperative. This PhD thesis addresses the need for equity and\ntransparency in NLP systems, recognizing that fairness in NLP is not merely a\ntechnical challenge but a moral and ethical necessity, requiring a rigorous\nexamination of how these technologies interact with and impact diverse human\npopulations. Through this lens, this thesis undertakes a thorough investigation\ninto the development of equitable NLP methodologies and the evaluation of\nbiases that prevail in current systems.\n  First, it introduces an innovative algorithm to mitigate biases in\nmulti-class classifiers, tailored for high-risk NLP applications, surpassing\ntraditional methods in both bias mitigation and prediction accuracy. Then, an\nanalysis of the Bios dataset reveals the impact of dataset size on\ndiscriminatory biases and the limitations of standard fairness metrics. This\nawareness has led to explorations in the field of explainable AI, aiming for a\nmore complete understanding of biases where traditional metrics are limited.\nConsequently, the thesis presents COCKATIEL, a model-agnostic explainability\nmethod that identifies and ranks concepts in Transformer models, outperforming\nprevious approaches in sentiment analysis tasks. Finally, the thesis\ncontributes to bridging the gap between fairness and explainability by\nintroducing TaCo, a novel method to neutralize bias in Transformer model\nembeddings.\n  In conclusion, this thesis constitutes a significant interdisciplinary\nendeavor that intertwines explicability and fairness to challenge and reshape\ncurrent NLP paradigms. The methodologies and critiques presented contribute to\nthe ongoing discourse on fairness in machine learning, offering actionable\nsolutions for more equitable and responsible AI systems.\n", "date": "2024-10-16", "categories": []}
{"id": "2410.09600", "title": "The Fragility of Fairness: Causal Sensitivity Analysis for Fair Machine\n  Learning", "abstract": "  Fairness metrics are a core tool in the fair machine learning literature\n(FairML), used to determine that ML models are, in some sense, \"fair\".\nReal-world data, however, are typically plagued by various measurement biases\nand other violated assumptions, which can render fairness assessments\nmeaningless. We adapt tools from causal sensitivity analysis to the FairML\ncontext, providing a general framework which (1) accommodates effectively any\ncombination of fairness metric and bias that can be posed in the \"oblivious\nsetting\"; (2) allows researchers to investigate combinations of biases,\nresulting in non-linear sensitivity; and (3) enables flexible encoding of\ndomain-specific constraints and assumptions. Employing this framework, we\nanalyze the sensitivity of the most common parity metrics under 3 varieties of\nclassifier across 14 canonical fairness datasets. Our analysis reveals the\nstriking fragility of fairness assessments to even minor dataset biases. We\nshow that causal sensitivity analysis provides a powerful and necessary toolkit\nfor gauging the informativeness of parity metric evaluations. Our repository is\navailable here: https://github.com/Jakefawkes/fragile_fair.\n", "date": "2024-10-12", "categories": []}
{"id": "2410.11084", "title": "Gender Bias in Decision-Making with Large Language Models: A Study of\n  Relationship Conflicts", "abstract": "  Large language models (LLMs) acquire beliefs about gender from training data\nand can therefore generate text with stereotypical gender attitudes. Prior\nstudies have demonstrated model generations favor one gender or exhibit\nstereotypes about gender, but have not investigated the complex dynamics that\ncan influence model reasoning and decision-making involving gender. We study\ngender equity within LLMs through a decision-making lens with a new dataset,\nDeMET Prompts, containing scenarios related to intimate, romantic\nrelationships. We explore nine relationship configurations through name pairs\nacross three name lists (men, women, neutral). We investigate equity in the\ncontext of gender roles through numerous lenses: typical and gender-neutral\nnames, with and without model safety enhancements, same and mixed-gender\nrelationships, and egalitarian versus traditional scenarios across various\ntopics. While all models exhibit the same biases (women favored, then those\nwith gender-neutral names, and lastly men), safety guardrails reduce bias. In\naddition, models tend to circumvent traditional male dominance stereotypes and\nside with 'traditionally female' individuals more often, suggesting\nrelationships are viewed as a female domain by the models.\n", "date": "2024-10-14", "categories": []}
{"id": "2307.08486", "title": "Fairness in KI-Systemen", "abstract": "  The more AI-assisted decisions affect people's lives, the more important the\nfairness of such decisions becomes. In this chapter, we provide an introduction\nto research on fairness in machine learning. We explain the main fairness\ndefinitions and strategies for achieving fairness using concrete examples and\nplace fairness research in the European context. Our contribution is aimed at\nan interdisciplinary audience and therefore avoids mathematical formulation but\nemphasizes visualizations and examples.\n  --\n  Je mehr KI-gest\\\"utzte Entscheidungen das Leben von Menschen betreffen, desto\nwichtiger ist die Fairness solcher Entscheidungen. In diesem Kapitel geben wir\neine Einf\\\"uhrung in die Forschung zu Fairness im maschinellen Lernen. Wir\nerkl\\\"aren die wesentlichen Fairness-Definitionen und Strategien zur Erreichung\nvon Fairness anhand konkreter Beispiele und ordnen die Fairness-Forschung in\nden europ\\\"aischen Kontext ein. Unser Beitrag richtet sich dabei an ein\ninterdisziplin\\\"ares Publikum und verzichtet daher auf die mathematische\nFormulierung sondern betont Visualisierungen und Beispiele.\n", "date": "2023-07-17", "categories": []}
{"id": "2405.13805", "title": "Perceptual Fairness in Image Restoration", "abstract": "  Fairness in image restoration tasks is the desire to treat different\nsub-groups of images equally well. Existing definitions of fairness in image\nrestoration are highly restrictive. They consider a reconstruction to be a\ncorrect outcome for a group (e.g., women) only if it falls within the group's\nset of ground truth images (e.g., natural images of women); otherwise, it is\nconsidered entirely incorrect. Consequently, such definitions are prone to\ncontroversy, as errors in image restoration can manifest in various ways. In\nthis work we offer an alternative approach towards fairness in image\nrestoration, by considering the Group Perceptual Index (GPI), which we define\nas the statistical distance between the distribution of the group's ground\ntruth images and the distribution of their reconstructions. We assess the\nfairness of an algorithm by comparing the GPI of different groups, and say that\nit achieves perfect Perceptual Fairness (PF) if the GPIs of all groups are\nidentical. We motivate and theoretically study our new notion of fairness, draw\nits connection to previous ones, and demonstrate its utility on\nstate-of-the-art face image restoration algorithms.\n", "date": "2024-05-22", "categories": []}
{"id": "2410.05180", "title": "Mitigating the Risk of Health Inequity Exacerbated by Large Language\n  Models", "abstract": "  Recent advancements in large language models have demonstrated their\npotential in numerous medical applications, particularly in automating clinical\ntrial matching for translational research and enhancing medical question\nanswering for clinical decision support. However, our study shows that\nincorporating non decisive sociodemographic factors such as race, sex, income\nlevel, LGBT+ status, homelessness, illiteracy, disability, and unemployment\ninto the input of LLMs can lead to incorrect and harmful outputs for these\npopulations. These discrepancies risk exacerbating existing health disparities\nif LLMs are widely adopted in healthcare. To address this issue, we introduce\nEquityGuard, a novel framework designed to detect and mitigate the risk of\nhealth inequities in LLM based medical applications. Our evaluation\ndemonstrates its efficacy in promoting equitable outcomes across diverse\npopulations.\n", "date": "2024-10-07", "categories": []}
{"id": "2410.06214", "title": "Fair-OBNC: Correcting Label Noise for Fairer Datasets", "abstract": "  Data used by automated decision-making systems, such as Machine Learning\nmodels, often reflects discriminatory behavior that occurred in the past. These\nbiases in the training data are sometimes related to label noise, such as in\nCOMPAS, where more African-American offenders are wrongly labeled as having a\nhigher risk of recidivism when compared to their White counterparts. Models\ntrained on such biased data may perpetuate or even aggravate the biases with\nrespect to sensitive information, such as gender, race, or age. However, while\nmultiple label noise correction approaches are available in the literature,\nthese focus on model performance exclusively. In this work, we propose\nFair-OBNC, a label noise correction method with fairness considerations, to\nproduce training datasets with measurable demographic parity. The presented\nmethod adapts Ordering-Based Noise Correction, with an adjusted criterion of\nordering, based both on the margin of error of an ensemble, and the potential\nincrease in the observed demographic parity of the dataset. We evaluate\nFair-OBNC against other different pre-processing techniques, under different\nscenarios of controlled label noise. Our results show that the proposed method\nis the overall better alternative within the pool of label correction methods,\nbeing capable of attaining better reconstructions of the original labels.\nModels trained in the corrected data have an increase, on average, of 150% in\ndemographic parity, when compared to models trained in data with noisy labels,\nacross the considered levels of label noise.\n", "date": "2024-10-08", "categories": []}
{"id": "2410.10160", "title": "Will the Inclusion of Generated Data Amplify Bias Across Generations in\n  Future Image Classification Models?", "abstract": "  As the demand for high-quality training data escalates, researchers have\nincreasingly turned to generative models to create synthetic data, addressing\ndata scarcity and enabling continuous model improvement. However, reliance on\nself-generated data introduces a critical question: Will this practice amplify\nbias in future models? While most research has focused on overall performance,\nthe impact on model bias, particularly subgroup bias, remains underexplored. In\nthis work, we investigate the effects of the generated data on image\nclassification tasks, with a specific focus on bias. We develop a practical\nsimulation environment that integrates a self-consuming loop, where the\ngenerative model and classification model are trained synergistically. Hundreds\nof experiments are conducted on Colorized MNIST, CIFAR-20/100, and Hard\nImageNet datasets to reveal changes in fairness metrics across generations. In\naddition, we provide a conjecture to explain the bias dynamics when training\nmodels on continuously augmented datasets across generations. Our findings\ncontribute to the ongoing debate on the implications of synthetic data for\nfairness in real-world applications.\n", "date": "2024-10-14", "categories": []}
{"id": "2409.03220", "title": "FairQuant: Certifying and Quantifying Fairness of Deep Neural Networks", "abstract": "  We propose a method for formally certifying and quantifying individual\nfairness of deep neural networks (DNN). Individual fairness guarantees that any\ntwo individuals who are identical except for a legally protected attribute\n(e.g., gender or race) receive the same treatment. While there are existing\ntechniques that provide such a guarantee, they tend to suffer from lack of\nscalability or accuracy as the size and input dimension of the DNN increase.\nOur method overcomes this limitation by applying abstraction to a symbolic\ninterval based analysis of the DNN followed by iterative refinement guided by\nthe fairness property. Furthermore, our method lifts the symbolic interval\nbased analysis from conventional qualitative certification to quantitative\ncertification, by computing the percentage of individuals whose classification\noutputs are provably fair, instead of merely deciding if the DNN is fair. We\nhave implemented our method and evaluated it on deep neural networks trained on\nfour popular fairness research datasets. The experimental results show that our\nmethod is not only more accurate than state-of-the-art techniques but also\nseveral orders-of-magnitude faster.\n", "date": "2024-09-05", "categories": []}
{"id": "2410.07589", "title": "No Free Lunch: Retrieval-Augmented Generation Undermines Fairness in\n  LLMs, Even for Vigilant Users", "abstract": "  Retrieval-Augmented Generation (RAG) is widely adopted for its effectiveness\nand cost-efficiency in mitigating hallucinations and enhancing the\ndomain-specific generation capabilities of large language models (LLMs).\nHowever, is this effectiveness and cost-efficiency truly a free lunch? In this\nstudy, we comprehensively investigate the fairness costs associated with RAG by\nproposing a practical three-level threat model from the perspective of user\nawareness of fairness. Specifically, varying levels of user fairness awareness\nresult in different degrees of fairness censorship on the external dataset. We\nexamine the fairness implications of RAG using uncensored, partially censored,\nand fully censored datasets. Our experiments demonstrate that fairness\nalignment can be easily undermined through RAG without the need for fine-tuning\nor retraining. Even with fully censored and supposedly unbiased external\ndatasets, RAG can lead to biased outputs. Our findings underscore the\nlimitations of current alignment methods in the context of RAG-based LLMs and\nhighlight the urgent need for new strategies to ensure fairness. We propose\npotential mitigations and call for further research to develop robust fairness\nsafeguards in RAG-based LLMs.\n", "date": "2024-10-10", "categories": []}
{"id": "2410.07820", "title": "Mitigating Gender Bias in Code Large Language Models via Model Editing", "abstract": "  In recent years, with the maturation of large language model (LLM) technology\nand the emergence of high-quality programming code datasets, researchers have\nbecome increasingly confident in addressing the challenges of program synthesis\nautomatically. However, since most of the training samples for LLMs are\nunscreened, it is inevitable that LLMs' performance may not align with\nreal-world scenarios, leading to the presence of social bias. To evaluate and\nquantify the gender bias in code LLMs, we propose a dataset named CodeGenBias\n(Gender Bias in the Code Generation) and an evaluation metric called FB-Score\n(Factual Bias Score) based on the actual gender distribution of correlative\nprofessions. With the help of CodeGenBias and FB-Score, we evaluate and analyze\nthe gender bias in eight mainstream Code LLMs. Previous work has demonstrated\nthat model editing methods that perform well in knowledge editing have the\npotential to mitigate social bias in LLMs. Therefore, we develop a model\nediting approach named MG-Editing (Multi-Granularity model Editing), which\nincludes the locating and editing phases. Our model editing method MG-Editing\ncan be applied at five different levels of model parameter granularity: full\nparameters level, layer level, module level, row level, and neuron level.\nExtensive experiments not only demonstrate that our MG-Editing can effectively\nmitigate the gender bias in code LLMs while maintaining their general code\ngeneration capabilities, but also showcase its excellent generalization. At the\nsame time, the experimental results show that, considering both the gender bias\nof the model and its general code generation capability, MG-Editing is most\neffective when applied at the row and neuron levels of granularity.\n", "date": "2024-10-10", "categories": []}
{"id": "2410.07884", "title": "Generated Bias: Auditing Internal Bias Dynamics of Text-To-Image\n  Generative Models", "abstract": "  Text-To-Image (TTI) Diffusion Models such as DALL-E and Stable Diffusion are\ncapable of generating images from text prompts. However, they have been shown\nto perpetuate gender stereotypes. These models process data internally in\nmultiple stages and employ several constituent models, often trained\nseparately. In this paper, we propose two novel metrics to measure bias\ninternally in these multistage multimodal models. Diffusion Bias was developed\nto detect and measures bias introduced by the diffusion stage of the models.\nBias Amplification measures amplification of bias during the text-to-image\nconversion process. Our experiments reveal that TTI models amplify gender bias,\nthe diffusion process itself contributes to bias and that Stable Diffusion v2\nis more prone to gender bias than DALL-E 2.\n", "date": "2024-10-10", "categories": []}
{"id": "2409.15567", "title": "Asking an AI for salary negotiation advice is a matter of concern:\n  Controlled experimental perturbation of ChatGPT for protected and\n  non-protected group discrimination on a contextual task with no clear ground\n  truth answers", "abstract": "  We conducted controlled experimental bias audits for four versions of\nChatGPT, which we asked to recommend an opening offer in salary negotiations\nfor a new hire. We submitted 98,800 prompts to each version, systematically\nvarying the employee's gender, university, and major, and tested prompts in\nvoice of each side of the negotiation: the employee versus employer. We find\nChatGPT as a multi-model platform is not robust and consistent enough to be\ntrusted for such a task. We observed statistically significant salary offers\nwhen varying gender for all four models, although with smaller gaps than for\nother attributes tested. The largest gaps were different model versions and\nbetween the employee- vs employer-voiced prompts. We also observed substantial\ngaps when varying university and major, but many of the biases were not\nconsistent across model versions. We tested for fictional and fraudulent\nuniversities and found wildly inconsistent results across cases and model\nversions. We make broader contributions to the AI/ML fairness literature. Our\nscenario and our experimental design differ from mainstream AI/ML auditing\nefforts in key ways. Bias audits typically test discrimination for protected\nclasses like gender, which we contrast with testing non-protected classes of\nuniversity and major. Asking for negotiation advice includes how aggressive one\nought to be in a negotiation relative to known empirical salary distributions\nand scales, which is a deeply contextual and personalized task that has no\nobjective ground truth to validate. These results raise concerns for the\nspecific model versions we tested and ChatGPT as a multi-model platform in\ncontinuous development. Our epistemology does not permit us to definitively\ncertify these models as either generally biased or unbiased on the attributes\nwe test, but our study raises matters of concern for stakeholders to further\ninvestigate.\n", "date": "2024-09-23", "categories": []}
{"id": "2409.20126", "title": "DCAST: Diverse Class-Aware Self-Training Mitigates Selection Bias for\n  Fairer Learning", "abstract": "  Fairness in machine learning seeks to mitigate model bias against individuals\nbased on sensitive features such as sex or age, often caused by an uneven\nrepresentation of the population in the training data due to selection bias.\nNotably, bias unascribed to sensitive features is challenging to identify and\ntypically goes undiagnosed, despite its prominence in complex high-dimensional\ndata from fields like computer vision and molecular biomedicine. Strategies to\nmitigate unidentified bias and evaluate mitigation methods are crucially\nneeded, yet remain underexplored. We introduce: (i) Diverse Class-Aware\nSelf-Training (DCAST), model-agnostic mitigation aware of class-specific bias,\nwhich promotes sample diversity to counter confirmation bias of conventional\nself-training while leveraging unlabeled samples for an improved representation\nof the underlying population; (ii) hierarchy bias, multivariate and class-aware\nbias induction without prior knowledge. Models learned with DCAST showed\nimproved robustness to hierarchy and other biases across eleven datasets,\nagainst conventional self-training and six prominent domain adaptation\ntechniques. Advantage was largest on multi-class classification, emphasizing\nDCAST as a promising strategy for fairer learning in different contexts.\n", "date": "2024-09-30", "categories": []}
{"id": "2410.05810", "title": "Uncertainty-Aware Fairness-Adaptive Classification Trees", "abstract": "  In an era where artificial intelligence and machine learning algorithms\nincreasingly impact human life, it is crucial to develop models that account\nfor potential discrimination in their predictions. This paper tackles this\nproblem by introducing a new classification tree algorithm using a novel\nsplitting criterion that incorporates fairness adjustments into the\ntree-building process. The proposed method integrates a fairness-aware impurity\nmeasure that balances predictive accuracy with fairness across protected\ngroups. By ensuring that each splitting node considers both the gain in\nclassification error and the fairness, our algorithm encourages splits that\nmitigate discrimination. Importantly, in penalizing unfair splits, we account\nfor the uncertainty in the fairness metric by utilizing its confidence interval\ninstead of relying on its point estimate. Experimental results on benchmark and\nsynthetic datasets illustrate that our method effectively reduces\ndiscriminatory predictions compared to traditional classification trees,\nwithout significant loss in overall accuracy.\n", "date": "2024-10-08", "categories": []}
{"id": "2410.06423", "title": "FAIREDU: A Multiple Regression-Based Method for Enhancing Fairness in\n  Machine Learning Models for Educational Applications", "abstract": "  Fairness in artificial intelligence and machine learning (AI/ML) models is\nbecoming critically important, especially as decisions made by these systems\nimpact diverse groups. In education, a vital sector for all countries, the\nwidespread application of AI/ML systems raises specific concerns regarding\nfairness. Current research predominantly focuses on fairness for individual\nsensitive features, which limits the comprehensiveness of fairness assessments.\nThis paper introduces FAIREDU, a novel and effective method designed to improve\nfairness across multiple sensitive features. Through extensive experiments, we\nevaluate FAIREDU effectiveness in enhancing fairness without compromising model\nperformance. The results demonstrate that FAIREDU addresses intersectionality\nacross features such as gender, race, age, and other sensitive features,\noutperforming state-of-the-art methods with minimal effect on model accuracy.\nThe paper also explores potential future research directions to enhance further\nthe method robustness and applicability to various machine-learning models and\ndatasets.\n", "date": "2024-10-08", "categories": []}
{"id": "2404.08230", "title": "Enhancing Fairness and Performance in Machine Learning Models: A\n  Multi-Task Learning Approach with Monte-Carlo Dropout and Pareto Optimality", "abstract": "  Bias originates from both data and algorithmic design, often exacerbated by\ntraditional fairness methods that fail to address the subtle impacts of\nprotected attributes. This study introduces an approach to mitigate bias in\nmachine learning by leveraging model uncertainty. Our approach utilizes a\nmulti-task learning (MTL) framework combined with Monte Carlo (MC) Dropout to\nassess and mitigate uncertainty in predictions related to protected labels. By\nincorporating MC Dropout, our framework quantifies prediction uncertainty,\nwhich is crucial in areas with vague decision boundaries, thereby enhancing\nmodel fairness. Our methodology integrates multi-objective learning through\npareto-optimality to balance fairness and performance across various\napplications. We demonstrate the effectiveness and transferability of our\napproach across multiple datasets and enhance model explainability through\nsaliency maps to interpret how input features influence predictions, thereby\nenhancing the interpretability of machine learning models in practical\napplications.\n", "date": "2024-04-12", "categories": []}
{"id": "2406.03142", "title": "On the Power of Randomization in Fair Classification and Representation", "abstract": "  Fair classification and fair representation learning are two important\nproblems in supervised and unsupervised fair machine learning, respectively.\nFair classification asks for a classifier that maximizes accuracy on a given\ndata distribution subject to fairness constraints. Fair representation maps a\ngiven data distribution over the original feature space to a distribution over\na new representation space such that all classifiers over the representation\nsatisfy fairness. In this paper, we examine the power of randomization in both\nthese problems to minimize the loss of accuracy that results when we impose\nfairness constraints. Previous work on fair classification has characterized\nthe optimal fair classifiers on a given data distribution that maximize\naccuracy subject to fairness constraints, e.g., Demographic Parity (DP), Equal\nOpportunity (EO), and Predictive Equality (PE). We refine these\ncharacterizations to demonstrate when the optimal randomized fair classifiers\ncan surpass their deterministic counterparts in accuracy. We also show how the\noptimal randomized fair classifier that we characterize can be obtained as a\nsolution to a convex optimization problem. Recent work has provided techniques\nto construct fair representations for a given data distribution such that any\nclassifier over this representation satisfies DP. However, the classifiers on\nthese fair representations either come with no or weak accuracy guarantees when\ncompared to the optimal fair classifier on the original data distribution.\nExtending our ideas for randomized fair classification, we improve on these\nworks, and construct DP-fair, EO-fair, and PE-fair representations that have\nprovably optimal accuracy and suffer no accuracy loss compared to the optimal\nDP-fair, EO-fair, and PE-fair classifiers respectively on the original data\ndistribution.\n", "date": "2024-06-05", "categories": []}
{"id": "2406.07685", "title": "Test-Time Fairness and Robustness in Large Language Models", "abstract": "  Frontier Large Language Models (LLMs) can be socially discriminatory or\nsensitive to spurious features of their inputs. Because only well-resourced\ncorporations can train frontier LLMs, we need robust test-time strategies to\ncontrol such biases. Existing solutions, which instruct the LLM to be fair or\nrobust, rely on the model's implicit understanding of bias. Causality provides\na rich formalism through which we can be explicit about our debiasing\nrequirements. Yet, as we show, a naive application of the standard causal\ndebiasing strategy, counterfactual data augmentation, fails under standard\nassumptions to debias predictions at an individual level at test time. To\naddress this, we develop a stratified notion of debiasing called stratified\ninvariance, which can capture a range of debiasing requirements from population\nlevel to individual level through an additional measurement that stratifies the\npredictions. We present a complete observational test for stratified\ninvariance. Finally, we introduce a data augmentation strategy that guarantees\nstratified invariance at test time under suitable assumptions, together with a\nprompting strategy that encourages stratified invariance in LLMs. We show that\nour prompting strategy, unlike implicit instructions, consistently reduces the\nbias of frontier LLMs across a suite of synthetic and real-world benchmarks\nwithout requiring additional data, finetuning or pre-training.\n", "date": "2024-06-11", "categories": []}
{"id": "2406.12232", "title": "\"You Gotta be a Doctor, Lin\": An Investigation of Name-Based Bias of\n  Large Language Models in Employment Recommendations", "abstract": "  Social science research has shown that candidates with names indicative of\ncertain races or genders often face discrimination in employment practices.\nSimilarly, Large Language Models (LLMs) have demonstrated racial and gender\nbiases in various applications. In this study, we utilize GPT-3.5-Turbo and\nLlama 3-70B-Instruct to simulate hiring decisions and salary recommendations\nfor candidates with 320 first names that strongly signal their race and gender,\nacross over 750,000 prompts. Our empirical results indicate a preference among\nthese models for hiring candidates with White female-sounding names over other\ndemographic groups across 40 occupations. Additionally, even among candidates\nwith identical qualifications, salary recommendations vary by as much as 5%\nbetween different subgroups. A comparison with real-world labor data reveals\ninconsistent alignment with U.S. labor market characteristics, underscoring the\nnecessity of risk investigation of LLM-powered systems.\n", "date": "2024-06-18", "categories": []}
{"id": "2410.03996", "title": "On the Influence of Gender and Race in Romantic Relationship Prediction\n  from Large Language Models", "abstract": "  We study the presence of heteronormative biases and prejudice against\ninterracial romantic relationships in large language models by performing\ncontrolled name-replacement experiments for the task of relationship\nprediction. We show that models are less likely to predict romantic\nrelationships for (a) same-gender character pairs than different-gender pairs;\nand (b) intra/inter-racial character pairs involving Asian names as compared to\nBlack, Hispanic, or White names. We examine the contextualized embeddings of\nfirst names and find that gender for Asian names is less discernible than\nnon-Asian names. We discuss the social implications of our findings,\nunderlining the need to prioritize the development of inclusive and equitable\ntechnology.\n", "date": "2024-10-05", "categories": []}
{"id": "2410.04287", "title": "Unveiling the Impact of Local Homophily on GNN Fairness: In-Depth\n  Analysis and New Benchmarks", "abstract": "  Graph Neural Networks (GNNs) often struggle to generalize when graphs exhibit\nboth homophily (same-class connections) and heterophily (different-class\nconnections). Specifically, GNNs tend to underperform for nodes with local\nhomophily levels that differ significantly from the global homophily level.\nThis issue poses a risk in user-centric applications where underrepresented\nhomophily levels are present. Concurrently, fairness within GNNs has received\nsubstantial attention due to the potential amplification of biases via message\npassing. However, the connection between local homophily and fairness in GNNs\nremains underexplored. In this work, we move beyond global homophily and\nexplore how local homophily levels can lead to unfair predictions. We begin by\nformalizing the challenge of fair predictions for underrepresented homophily\nlevels as an out-of-distribution (OOD) problem. We then conduct a theoretical\nanalysis that demonstrates how local homophily levels can alter predictions for\ndiffering sensitive attributes. We additionally introduce three new GNN\nfairness benchmarks, as well as a novel semi-synthetic graph generator, to\nempirically study the OOD problem. Across extensive analysis we find that two\nfactors can promote unfairness: (a) OOD distance, and (b) heterophilous nodes\nsituated in homophilous graphs. In cases where these two conditions are met,\nfairness drops by up to 24% on real world datasets, and 30% in semi-synthetic\ndatasets. Together, our theoretical insights, empirical analysis, and\nalgorithmic contributions unveil a previously overlooked source of unfairness\nrooted in the graph's homophily information.\n", "date": "2024-10-05", "categories": []}
{"id": "2410.04551", "title": "Social Choice for Heterogeneous Fairness in Recommendation", "abstract": "  Algorithmic fairness in recommender systems requires close attention to the\nneeds of a diverse set of stakeholders that may have competing interests.\nPrevious work in this area has often been limited by fixed, single-objective\ndefinitions of fairness, built into algorithms or optimization criteria that\nare applied to a single fairness dimension or, at most, applied identically\nacross dimensions. These narrow conceptualizations limit the ability to adapt\nfairness-aware solutions to the wide range of stakeholder needs and fairness\ndefinitions that arise in practice. Our work approaches recommendation fairness\nfrom the standpoint of computational social choice, using a multi-agent\nframework. In this paper, we explore the properties of different social choice\nmechanisms and demonstrate the successful integration of multiple,\nheterogeneous fairness definitions across multiple data sets.\n", "date": "2024-10-06", "categories": []}
{"id": "2312.08064", "title": "Human-in-the-loop Fairness: Integrating Stakeholder Feedback to\n  Incorporate Fairness Perspectives in Responsible AI", "abstract": "  Fairness is a growing concern for high-risk decision-making using Artificial\nIntelligence (AI) but ensuring it through purely technical means is\nchallenging: there is no universally accepted fairness measure, fairness is\ncontext-dependent, and there might be conflicting perspectives on what is\nconsidered fair. Thus, involving stakeholders, often without a background in AI\nor fairness, is a promising avenue. Research to directly involve stakeholders\nis in its infancy, and many questions remain on how to support stakeholders to\nfeedback on fairness, and how this feedback can be integrated into AI models.\nOur work follows an approach where stakeholders can give feedback on specific\ndecision instances and their outcomes with respect to their fairness, and then\nto retrain an AI model. In order to investigate this approach, we conducted two\nstudies of a complex AI model for credit rating used in loan applications. In\nstudy 1, we collected feedback from 58 lay users on loan application decisions,\nand conducted offline experiments to investigate the effects on accuracy and\nfairness metrics. In study 2, we deepened this investigation by showing 66\nparticipants the results of their feedback with respect to fairness, and then\nconducted further offline analyses. Our work contributes two datasets and\nassociated code frameworks to bootstrap further research, highlights the\nopportunities and challenges of employing lay user feedback for improving AI\nfairness, and discusses practical implications for developing AI applications\nthat more closely reflect stakeholder views about fairness.\n", "date": "2023-12-13", "categories": []}
{"id": "2405.13977", "title": "Improving Fairness and Mitigating MADness in Generative Models", "abstract": "  Generative models unfairly penalize data belonging to minority classes,\nsuffer from model autophagy disorder (MADness), and learn biased estimates of\nthe underlying distribution parameters. Our theoretical and empirical results\nshow that training generative models with intentionally designed hypernetworks\nleads to models that 1) are more fair when generating datapoints belonging to\nminority classes 2) are more stable in a self-consumed (i.e., MAD) setting, and\n3) learn parameters that are less statistically biased. To further mitigate\nunfairness, MADness, and bias, we introduce a regularization term that\npenalizes discrepancies between a generative model's estimated weights when\ntrained on real data versus its own synthetic data. To facilitate training\nexisting deep generative models within our framework, we offer a scalable\nimplementation of hypernetworks that automatically generates a hypernetwork\narchitecture for any given generative model.\n", "date": "2024-05-22", "categories": []}
{"id": "2407.02066", "title": "BiasDora: Exploring Hidden Biased Associations in Vision-Language Models", "abstract": "  Existing works examining Vision-Language Models (VLMs) for social biases\npredominantly focus on a limited set of documented bias associations, such as\ngender:profession or race:crime. This narrow scope often overlooks a vast range\nof unexamined implicit associations, restricting the identification and, hence,\nmitigation of such biases. We address this gap by probing VLMs to (1) uncover\nhidden, implicit associations across 9 bias dimensions. We systematically\nexplore diverse input and output modalities and (2) demonstrate how biased\nassociations vary in their negativity, toxicity, and extremity. Our work (3)\nidentifies subtle and extreme biases that are typically not recognized by\nexisting methodologies. We make the Dataset of retrieved associations, (Dora),\npublicly available here https://github.com/chahatraj/BiasDora.\n", "date": "2024-07-02", "categories": []}
{"id": "2410.03474", "title": "Group Fairness in Peer Review", "abstract": "  Large conferences such as NeurIPS and AAAI serve as crossroads of various AI\nfields, since they attract submissions from a vast number of communities.\nHowever, in some cases, this has resulted in a poor reviewing experience for\nsome communities, whose submissions get assigned to less qualified reviewers\noutside of their communities. An often-advocated solution is to break up any\nsuch large conference into smaller conferences, but this can lead to isolation\nof communities and harm interdisciplinary research. We tackle this challenge by\nintroducing a notion of group fairness, called the core, which requires that\nevery possible community (subset of researchers) to be treated in a way that\nprevents them from unilaterally benefiting by withdrawing from a large\nconference.\n  We study a simple peer review model, prove that it always admits a reviewing\nassignment in the core, and design an efficient algorithm to find one such\nassignment. We use real data from CVPR and ICLR conferences to compare our\nalgorithm to existing reviewing assignment algorithms on a number of metrics.\n", "date": "2024-10-04", "categories": []}
{"id": "2401.03140", "title": "Fair Sampling in Diffusion Models through Switching Mechanism", "abstract": "  Diffusion models have shown their effectiveness in generation tasks by\nwell-approximating the underlying probability distribution. However, diffusion\nmodels are known to suffer from an amplified inherent bias from the training\ndata in terms of fairness. While the sampling process of diffusion models can\nbe controlled by conditional guidance, previous works have attempted to find\nempirical guidance to achieve quantitative fairness. To address this\nlimitation, we propose a fairness-aware sampling method called\n\\textit{attribute switching} mechanism for diffusion models. Without additional\ntraining, the proposed sampling can obfuscate sensitive attributes in generated\ndata without relying on classifiers. We mathematically prove and experimentally\ndemonstrate the effectiveness of the proposed method on two key aspects: (i)\nthe generation of fair data and (ii) the preservation of the utility of the\ngenerated data.\n", "date": "2024-01-06", "categories": []}
{"id": "2401.03562", "title": "GLOCALFAIR: Jointly Improving Global and Local Group Fairness in\n  Federated Learning", "abstract": "  Federated learning (FL) has emerged as a prospective solution for\ncollaboratively learning a shared model across clients without sacrificing\ntheir data privacy. However, the federated learned model tends to be biased\nagainst certain demographic groups (e.g., racial and gender groups) due to the\ninherent FL properties, such as data heterogeneity and party selection. Unlike\ncentralized learning, mitigating bias in FL is particularly challenging as\nprivate training datasets and their sensitive attributes are typically not\ndirectly accessible. Most prior research in this field only focuses on global\nfairness while overlooking the local fairness of individual clients. Moreover,\nexisting methods often require sensitive information about the client's local\ndatasets to be shared, which is not desirable. To address these issues, we\npropose GLOCALFAIR, a client-server co-design fairness framework that can\njointly improve global and local group fairness in FL without the need for\nsensitive statistics about the client's private datasets. Specifically, we\nutilize constrained optimization to enforce local fairness on the client side\nand adopt a fairness-aware clustering-based aggregation on the server to\nfurther ensure the global model fairness across different sensitive groups\nwhile maintaining high utility. Experiments on two image datasets and one\ntabular dataset with various state-of-the-art fairness baselines show that\nGLOCALFAIR can achieve enhanced fairness under both global and local data\ndistributions while maintaining a good level of utility and client fairness.\n", "date": "2024-01-07", "categories": []}
{"id": "2410.02005", "title": "FairlyUncertain: A Comprehensive Benchmark of Uncertainty in Algorithmic\n  Fairness", "abstract": "  Fair predictive algorithms hinge on both equality and trust, yet inherent\nuncertainty in real-world data challenges our ability to make consistent, fair,\nand calibrated decisions. While fairly managing predictive error has been\nextensively explored, some recent work has begun to address the challenge of\nfairly accounting for irreducible prediction uncertainty. However, a clear\ntaxonomy and well-specified objectives for integrating uncertainty into\nfairness remains undefined. We address this gap by introducing FairlyUncertain,\nan axiomatic benchmark for evaluating uncertainty estimates in fairness. Our\nbenchmark posits that fair predictive uncertainty estimates should be\nconsistent across learning pipelines and calibrated to observed randomness.\nThrough extensive experiments on ten popular fairness datasets, our evaluation\nreveals: (1) A theoretically justified and simple method for estimating\nuncertainty in binary settings is more consistent and calibrated than prior\nwork; (2) Abstaining from binary predictions, even with improved uncertainty\nestimates, reduces error but does not alleviate outcome imbalances between\ndemographic groups; (3) Incorporating consistent and calibrated uncertainty\nestimates in regression tasks improves fairness without any explicit fairness\ninterventions. Additionally, our benchmark package is designed to be extensible\nand open-source, to grow with the field. By providing a standardized framework\nfor assessing the interplay between uncertainty and fairness, FairlyUncertain\npaves the way for more equitable and trustworthy machine learning practices.\n", "date": "2024-10-02", "categories": []}
{"id": "2410.02618", "title": "Achieving Fairness in Predictive Process Analytics via Adversarial\n  Learning (Extended Version)", "abstract": "  Predictive business process analytics has become important for organizations,\noffering real-time operational support for their processes. However, these\nalgorithms often perform unfair predictions because they are based on biased\nvariables (e.g., gender or nationality), namely variables embodying\ndiscrimination. This paper addresses the challenge of integrating a debiasing\nphase into predictive business process analytics to ensure that predictions are\nnot influenced by biased variables. Our framework leverages on adversial\ndebiasing is evaluated on four case studies, showing a significant reduction in\nthe contribution of biased variables to the predicted value. The proposed\ntechnique is also compared with the state of the art in fairness in process\nmining, illustrating that our framework allows for a more enhanced level of\nfairness, while retaining a better prediction quality.\n", "date": "2024-10-03", "categories": []}
{"id": "2409.19541", "title": "Unlabeled Debiasing in Downstream Tasks via Class-wise Low Variance\n  Regularization", "abstract": "  Language models frequently inherit societal biases from their training data.\nNumerous techniques have been proposed to mitigate these biases during both the\npre-training and fine-tuning stages. However, fine-tuning a pre-trained\ndebiased language model on a downstream task can reintroduce biases into the\nmodel. Additionally, existing debiasing methods for downstream tasks either (i)\nrequire labels of protected attributes (e.g., age, race, or political views)\nthat are often not available or (ii) rely on indicators of bias, which\nrestricts their applicability to gender debiasing since they rely on\ngender-specific words. To address this, we introduce a novel debiasing\nregularization technique based on the class-wise variance of embeddings.\nCrucially, our method does not require attribute labels and targets any\nattribute, thus addressing the shortcomings of existing debiasing methods. Our\nexperiments on encoder language models and three datasets demonstrate that our\nmethod outperforms existing strong debiasing baselines that rely on target\nattribute labels while maintaining performance on the target task.\n", "date": "2024-09-29", "categories": []}
{"id": "2410.01089", "title": "FMBench: Benchmarking Fairness in Multimodal Large Language Models on\n  Medical Tasks", "abstract": "  Advancements in Multimodal Large Language Models (MLLMs) have significantly\nimproved medical task performance, such as Visual Question Answering (VQA) and\nReport Generation (RG). However, the fairness of these models across diverse\ndemographic groups remains underexplored, despite its importance in healthcare.\nThis oversight is partly due to the lack of demographic diversity in existing\nmedical multimodal datasets, which complicates the evaluation of fairness. In\nresponse, we propose FMBench, the first benchmark designed to evaluate the\nfairness of MLLMs performance across diverse demographic attributes. FMBench\nhas the following key features: 1: It includes four demographic attributes:\nrace, ethnicity, language, and gender, across two tasks, VQA and RG, under\nzero-shot settings. 2: Our VQA task is free-form, enhancing real-world\napplicability and mitigating the biases associated with predefined choices. 3:\nWe utilize both lexical metrics and LLM-based metrics, aligned with clinical\nevaluations, to assess models not only for linguistic accuracy but also from a\nclinical perspective. Furthermore, we introduce a new metric, Fairness-Aware\nPerformance (FAP), to evaluate how fairly MLLMs perform across various\ndemographic attributes. We thoroughly evaluate the performance and fairness of\neight state-of-the-art open-source MLLMs, including both general and medical\nMLLMs, ranging from 7B to 26B parameters on the proposed benchmark. We aim for\nFMBench to assist the research community in refining model evaluation and\ndriving future advancements in the field. All data and code will be released\nupon acceptance.\n", "date": "2024-10-01", "categories": []}
{"id": "2410.01145", "title": "ProxiMix: Enhancing Fairness with Proximity Samples in Subgroups", "abstract": "  Many bias mitigation methods have been developed for addressing fairness\nissues in machine learning. We found that using linear mixup alone, a data\naugmentation technique, for bias mitigation, can still retain biases present in\ndataset labels. Research presented in this paper aims to address this issue by\nproposing a novel pre-processing strategy in which both an existing mixup\nmethod and our new bias mitigation algorithm can be utilized to improve the\ngeneration of labels of augmented samples, which are proximity aware.\nSpecifically, we proposed ProxiMix which keeps both pairwise and proximity\nrelationships for fairer data augmentation. We conducted thorough experiments\nwith three datasets, three ML models, and different hyperparameters settings.\nOur experimental results showed the effectiveness of ProxiMix from both\nfairness of predictions and fairness of recourse perspectives.\n", "date": "2024-10-02", "categories": []}
{"id": "2304.09362", "title": "Long-Term Fairness with Unknown Dynamics", "abstract": "  While machine learning can myopically reinforce social inequalities, it may\nalso be used to dynamically seek equitable outcomes. In this paper, we\nformalize long-term fairness in the context of online reinforcement learning.\nThis formulation can accommodate dynamical control objectives, such as driving\nequity inherent in the state of a population, that cannot be incorporated into\nstatic formulations of fairness. We demonstrate that this framing allows an\nalgorithm to adapt to unknown dynamics by sacrificing short-term incentives to\ndrive a classifier-population system towards more desirable equilibria. For the\nproposed setting, we develop an algorithm that adapts recent work in online\nlearning. We prove that this algorithm achieves simultaneous probabilistic\nbounds on cumulative loss and cumulative violations of fairness (as statistical\nregularities between demographic groups). We compare our proposed algorithm to\nthe repeated retraining of myopic classifiers, as a baseline, and to a deep\nreinforcement learning algorithm that lacks safety guarantees. Our experiments\nmodel human populations according to evolutionary game theory and integrate\nreal-world datasets.\n", "date": "2023-04-19", "categories": []}
{"id": "2408.06890", "title": "BMFT: Achieving Fairness via Bias-based Weight Masking Fine-tuning", "abstract": "  Developing models with robust group fairness properties is paramount,\nparticularly in ethically sensitive domains such as medical diagnosis. Recent\napproaches to achieving fairness in machine learning require a substantial\namount of training data and depend on model retraining, which may not be\npractical in real-world scenarios. To mitigate these challenges, we propose\nBias-based Weight Masking Fine-Tuning (BMFT), a novel post-processing method\nthat enhances the fairness of a trained model in significantly fewer epochs\nwithout requiring access to the original training data. BMFT produces a mask\nover model parameters, which efficiently identifies the weights contributing\nthe most towards biased predictions. Furthermore, we propose a two-step\ndebiasing strategy, wherein the feature extractor undergoes initial fine-tuning\non the identified bias-influenced weights, succeeded by a fine-tuning phase on\na reinitialised classification layer to uphold discriminative performance.\nExtensive experiments across four dermatological datasets and two sensitive\nattributes demonstrate that BMFT outperforms existing state-of-the-art (SOTA)\ntechniques in both diagnostic accuracy and fairness metrics. Our findings\nunderscore the efficacy and robustness of BMFT in advancing fairness across\nvarious out-of-distribution (OOD) settings. Our code is available at:\nhttps://github.com/vios-s/BMFT\n", "date": "2024-08-13", "categories": []}
{"id": "2408.14126", "title": "Enhancing Fairness through Reweighting: A Path to Attain the Sufficiency\n  Rule", "abstract": "  We introduce an innovative approach to enhancing the empirical risk\nminimization (ERM) process in model training through a refined reweighting\nscheme of the training data to enhance fairness. This scheme aims to uphold the\nsufficiency rule in fairness by ensuring that optimal predictors maintain\nconsistency across diverse sub-groups. We employ a bilevel formulation to\naddress this challenge, wherein we explore sample reweighting strategies.\nUnlike conventional methods that hinge on model size, our formulation bases\ngeneralization complexity on the space of sample weights. We discretize the\nweights to improve training speed. Empirical validation of our method showcases\nits effectiveness and robustness, revealing a consistent improvement in the\nbalance between prediction performance and fairness metrics across various\nexperiments.\n", "date": "2024-08-26", "categories": []}
{"id": "2401.02150", "title": "Marginal Debiased Network for Fair Visual Recognition", "abstract": "  Deep neural networks (DNNs) are often prone to learn the spurious\ncorrelations between target classes and bias attributes, like gender and race,\ninherent in a major portion of training data (bias-aligned samples), thus\nshowing unfair behavior and arising controversy in the modern pluralistic and\negalitarian society. In this paper, we propose a novel marginal debiased\nnetwork (MDN) to learn debiased representations. More specifically, a marginal\nsoftmax loss (MSL) is designed by introducing the idea of margin penalty into\nthe fairness problem, which assigns a larger margin for bias-conflicting\nsamples (data without spurious correlations) than for bias-aligned ones, so as\nto deemphasize the spurious correlations and improve generalization on unbiased\ntest criteria. To determine the margins, our MDN is optimized through a meta\nlearning framework. We propose a meta equalized loss (MEL) to perceive the\nmodel fairness, and adaptively update the margin parameters by\nmeta-optimization which requires the trained model guided by the optimal\nmargins should minimize MEL computed on an unbiased meta-validation set.\nExtensive experiments on BiasedMNIST, Corrupted CIFAR-10, CelebA and UTK-Face\ndatasets demonstrate that our MDN can achieve a remarkable performance on\nunder-represented samples and obtain superior debiased results against the\nprevious approaches.\n", "date": "2024-01-04", "categories": []}
{"id": "2409.19318", "title": "Fairness Analysis with Shapley-Owen Effects", "abstract": "  We argue that relative importance and its equitable attribution in terms of\nShapley-Owen effects is an appropriate one, and, if we accept a small number of\nreasonable imperatives for equitable attribution, the only way to measure\nfairness. On the other hand, the computation of Shapley-Owen effects can be\nvery demanding. Our main technical result is a spectral decomposition of the\nShapley-Owen effects, which decomposes the computation of these indices into a\nmodel-specific and a model-independent part. The model-independent part is\nprecomputed once and for all, and the model-specific computation of\nShapley-Owen effects is expressed analytically in terms of the coefficients of\nthe model's \\emph{polynomial chaos expansion} (PCE), which can now be reused to\ncompute different Shapley-Owen effects. We also propose an algorithm for\ncomputing precise and sparse truncations of the PCE of the model and the\nspectral decomposition of the Shapley-Owen effects, together with upper bounds\non the accumulated approximation errors. The approximations of both the PCE and\nthe Shapley-Owen effects converge to their true values.\n", "date": "2024-09-28", "categories": []}
{"id": "2409.19940", "title": "Positive-Sum Fairness: Leveraging Demographic Attributes to Achieve Fair\n  AI Outcomes Without Sacrificing Group Gains", "abstract": "  Fairness in medical AI is increasingly recognized as a crucial aspect of\nhealthcare delivery. While most of the prior work done on fairness emphasizes\nthe importance of equal performance, we argue that decreases in fairness can be\neither harmful or non-harmful, depending on the type of change and how\nsensitive attributes are used. To this end, we introduce the notion of\npositive-sum fairness, which states that an increase in performance that\nresults in a larger group disparity is acceptable as long as it does not come\nat the cost of individual subgroup performance. This allows sensitive\nattributes correlated with the disease to be used to increase performance\nwithout compromising on fairness.\n  We illustrate this idea by comparing four CNN models that make different use\nof the race attribute in the training phase. The results show that removing all\ndemographic encodings from the images helps close the gap in performance\nbetween the different subgroups, whereas leveraging the race attribute as a\nmodel's input increases the overall performance while widening the disparities\nbetween subgroups. These larger gaps are then put in perspective of the\ncollective benefit through our notion of positive-sum fairness to distinguish\nharmful from non harmful disparities.\n", "date": "2024-09-30", "categories": []}
{"id": "2409.20536", "title": "Best Practices for Responsible Machine Learning in Credit Scoring", "abstract": "  The widespread use of machine learning in credit scoring has brought\nsignificant advancements in risk assessment and decision-making. However, it\nhas also raised concerns about potential biases, discrimination, and lack of\ntransparency in these automated systems. This tutorial paper performed a\nnon-systematic literature review to guide best practices for developing\nresponsible machine learning models in credit scoring, focusing on fairness,\nreject inference, and explainability. We discuss definitions, metrics, and\ntechniques for mitigating biases and ensuring equitable outcomes across\ndifferent groups. Additionally, we address the issue of limited data\nrepresentativeness by exploring reject inference methods that incorporate\ninformation from rejected loan applications. Finally, we emphasize the\nimportance of transparency and explainability in credit models, discussing\ntechniques that provide insights into the decision-making process and enable\nindividuals to understand and potentially improve their creditworthiness. By\nadopting these best practices, financial institutions can harness the power of\nmachine learning while upholding ethical and responsible lending practices.\n", "date": "2024-09-30", "categories": []}
{"id": "2409.18470", "title": "Fairness without Sensitive Attributes via Knowledge Sharing", "abstract": "  While model fairness improvement has been explored previously, existing\nmethods invariably rely on adjusting explicit sensitive attribute values in\norder to improve model fairness in downstream tasks. However, we observe a\ntrend in which sensitive demographic information becomes inaccessible as public\nconcerns around data privacy grow. In this paper, we propose a confidence-based\nhierarchical classifier structure called \"Reckoner\" for reliable fair model\nlearning under the assumption of missing sensitive attributes. We first present\nresults showing that if the dataset contains biased labels or other hidden\nbiases, classifiers significantly increase the bias gap across different\ndemographic groups in the subset with higher prediction confidence. Inspired by\nthese findings, we devised a dual-model system in which a version of the model\ninitialised with a high-confidence data subset learns from a version of the\nmodel initialised with a low-confidence data subset, enabling it to avoid\nbiased predictions. Our experimental results show that Reckoner consistently\noutperforms state-of-the-art baselines in COMPAS dataset and New Adult dataset,\nconsidering both accuracy and fairness metrics.\n", "date": "2024-09-27", "categories": []}
{"id": "2409.18499", "title": "Fairness-aware Multiobjective Evolutionary Learning", "abstract": "  Multiobjective evolutionary learning (MOEL) has demonstrated its advantages\nof training fairer machine learning models considering a predefined set of\nconflicting objectives, including accuracy and different fairness measures.\nRecent works propose to construct a representative subset of fairness measures\nas optimisation objectives of MOEL throughout model training. However, the\ndetermination of a representative measure set relies on dataset, prior\nknowledge and requires substantial computational costs. What's more, those\nrepresentative measures may differ across different model training processes.\nInstead of using a static predefined set determined before model training, this\npaper proposes to dynamically and adaptively determine a representative measure\nset online during model training. The dynamically determined representative set\nis then used as optimising objectives of the MOEL framework and can vary with\ntime. Extensive experimental results on 12 well-known benchmark datasets\ndemonstrate that our proposed framework achieves outstanding performance\ncompared to state-of-the-art approaches for mitigating unfairness in terms of\naccuracy as well as 25 fairness measures although only a few of them were\ndynamically selected and used as optimisation objectives. The results indicate\nthe importance of setting optimisation objectives dynamically during training.\n", "date": "2024-09-27", "categories": []}
{"id": "2210.16754", "title": "Mitigating Unfairness via Evolutionary Multi-objective Ensemble Learning", "abstract": "  In the literature of mitigating unfairness in machine learning, many fairness\nmeasures are designed to evaluate predictions of learning models and also\nutilised to guide the training of fair models. It has been theoretically and\nempirically shown that there exist conflicts and inconsistencies among accuracy\nand multiple fairness measures. Optimising one or several fairness measures may\nsacrifice or deteriorate other measures. Two key questions should be\nconsidered, how to simultaneously optimise accuracy and multiple fairness\nmeasures, and how to optimise all the considered fairness measures more\neffectively. In this paper, we view the mitigating unfairness problem as a\nmulti-objective learning problem considering the conflicts among fairness\nmeasures. A multi-objective evolutionary learning framework is used to\nsimultaneously optimise several metrics (including accuracy and multiple\nfairness measures) of machine learning models. Then, ensembles are constructed\nbased on the learning models in order to automatically balance different\nmetrics. Empirical results on eight well-known datasets demonstrate that\ncompared with the state-of-the-art approaches for mitigating unfairness, our\nproposed algorithm can provide decision-makers with better tradeoffs among\naccuracy and multiple fairness metrics. Furthermore, the high-quality models\ngenerated by the framework can be used to construct an ensemble to\nautomatically achieve a better tradeoff among all the considered fairness\nmetrics than other ensemble methods. Our code is publicly available at\nhttps://github.com/qingquan63/FairEMOL\n", "date": "2022-10-30", "categories": []}
{"id": "2409.17336", "title": "The Technology of Outrage: Bias in Artificial Intelligence", "abstract": "  Artificial intelligence and machine learning are increasingly used to offload\ndecision making from people. In the past, one of the rationales for this\nreplacement was that machines, unlike people, can be fair and unbiased.\nEvidence suggests otherwise. We begin by entertaining the ideas that algorithms\ncan replace people and that algorithms cannot be biased. Taken as axioms, these\nstatements quickly lead to absurdity. Spurred on by this result, we investigate\nthe slogans more closely and identify equivocation surrounding the word 'bias.'\nWe diagnose three forms of outrage-intellectual, moral, and political-that are\nat play when people react emotionally to algorithmic bias. Then we suggest\nthree practical approaches to addressing bias that the AI community could take,\nwhich include clarifying the language around bias, developing new auditing\nmethods for intelligent systems, and building certain capabilities into these\nsystems. We conclude by offering a moral regarding the conversations about\nalgorithmic bias that may transfer to other areas of artificial intelligence.\n", "date": "2024-09-25", "categories": []}
{"id": "2409.17691", "title": "Efficient Bias Mitigation Without Privileged Information", "abstract": "  Deep neural networks trained via empirical risk minimisation often exhibit\nsignificant performance disparities across groups, particularly when group and\ntask labels are spuriously correlated (e.g., \"grassy background\" and \"cows\").\nExisting bias mitigation methods that aim to address this issue often either\nrely on group labels for training or validation, or require an extensive\nhyperparameter search. Such data and computational requirements hinder the\npractical deployment of these methods, especially when datasets are too large\nto be group-annotated, computational resources are limited, and models are\ntrained through already complex pipelines. In this paper, we propose Targeted\nAugmentations for Bias Mitigation (TAB), a simple hyperparameter-free framework\nthat leverages the entire training history of a helper model to identify\nspurious samples, and generate a group-balanced training set from which a\nrobust model can be trained. We show that TAB improves worst-group performance\nwithout any group information or model selection, outperforming existing\nmethods while maintaining overall accuracy.\n", "date": "2024-09-26", "categories": []}
{"id": "2305.12620", "title": "Keeping Up with the Language Models: Systematic Benchmark Extension for\n  Bias Auditing", "abstract": "  Bias auditing of language models (LMs) has received considerable attention as\nLMs are becoming widespread. As such, several benchmarks for bias auditing have\nbeen proposed. At the same time, the rapid evolution of LMs can make these\nbenchmarks irrelevant in no time. Bias auditing is further complicated by LM\nbrittleness: when a presumably biased outcome is observed, is it due to model\nbias or model brittleness? We propose enlisting the models themselves to help\nconstruct bias auditing datasets that remain challenging, and introduce bias\nmeasures that distinguish between different types of model errors. First, we\nextend an existing bias benchmark for NLI (BBNLI) using a combination of\nLM-generated lexical variations, adversarial filtering, and human validation.\nWe demonstrate that the newly created dataset BBNLI-next is more challenging\nthan BBNLI: on average, BBNLI-next reduces the accuracy of state-of-the-art NLI\nmodels from 95.3%, as observed by BBNLI, to a strikingly low 57.5%. Second, we\nemploy BBNLI-next to showcase the interplay between robustness and bias: we\npoint out shortcomings in current bias scores and propose bias measures that\ntake into account both bias and model brittleness. Third, despite the fact that\nBBNLI-next was designed with non-generative models in mind, we show that the\nnew dataset is also able to uncover bias in state-of-the-art open-source\ngenerative LMs.\n  Note: All datasets included in this work are in English and they address\nUS-centered social biases. In the spirit of efficient NLP research, no model\ntraining or fine-tuning was performed to conduct this research.\n  Warning: This paper contains offensive text examples.\n", "date": "2023-05-22", "categories": []}
{"id": "2409.16371", "title": "Do the Right Thing, Just Debias! Multi-Category Bias Mitigation Using\n  LLMs", "abstract": "  This paper tackles the challenge of building robust and generalizable bias\nmitigation models for language. Recognizing the limitations of existing\ndatasets, we introduce ANUBIS, a novel dataset with 1507 carefully curated\nsentence pairs encompassing nine social bias categories. We evaluate\nstate-of-the-art models like T5, utilizing Supervised Fine-Tuning (SFT),\nReinforcement Learning (PPO, DPO), and In-Context Learning (ICL) for effective\nbias mitigation. Our analysis focuses on multi-class social bias reduction,\ncross-dataset generalizability, and environmental impact of the trained models.\nANUBIS and our findings offer valuable resources for building more equitable AI\nsystems and contribute to the development of responsible and unbiased\ntechnologies with broad societal impact.\n", "date": "2024-09-24", "categories": []}
{"id": "2409.16430", "title": "A Comprehensive Survey of Bias in LLMs: Current Landscape and Future\n  Directions", "abstract": "  Large Language Models(LLMs) have revolutionized various applications in\nnatural language processing (NLP) by providing unprecedented text generation,\ntranslation, and comprehension capabilities. However, their widespread\ndeployment has brought to light significant concerns regarding biases embedded\nwithin these models. This paper presents a comprehensive survey of biases in\nLLMs, aiming to provide an extensive review of the types, sources, impacts, and\nmitigation strategies related to these biases. We systematically categorize\nbiases into several dimensions. Our survey synthesizes current research\nfindings and discusses the implications of biases in real-world applications.\nAdditionally, we critically assess existing bias mitigation techniques and\npropose future research directions to enhance fairness and equity in LLMs. This\nsurvey serves as a foundational resource for researchers, practitioners, and\npolicymakers concerned with addressing and understanding biases in LLMs.\n", "date": "2024-09-24", "categories": []}
{"id": "2310.13746", "title": "FairBranch: Mitigating Bias Transfer in Fair Multi-task Learning", "abstract": "  The generalisation capacity of Multi-Task Learning (MTL) suffers when\nunrelated tasks negatively impact each other by updating shared parameters with\nconflicting gradients. This is known as negative transfer and leads to a drop\nin MTL accuracy compared to single-task learning (STL). Lately, there has been\na growing focus on the fairness of MTL models, requiring the optimization of\nboth accuracy and fairness for individual tasks. Analogously to negative\ntransfer for accuracy, task-specific fairness considerations might adversely\naffect the fairness of other tasks when there is a conflict of fairness loss\ngradients between the jointly learned tasks - we refer to this as Bias\nTransfer. To address both negative- and bias-transfer in MTL, we propose a\nnovel method called FairBranch, which branches the MTL model by assessing the\nsimilarity of learned parameters, thereby grouping related tasks to alleviate\nnegative transfer. Moreover, it incorporates fairness loss gradient conflict\ncorrection between adjoining task-group branches to address bias transfer\nwithin these task groups. Our experiments on tabular and visual MTL problems\nshow that FairBranch outperforms state-of-the-art MTLs on both fairness and\naccuracy.\n", "date": "2023-10-20", "categories": []}
{"id": "2311.12435", "title": "Fair Enough? A map of the current limitations of the requirements to\n  have fair algorithms", "abstract": "  In recent years, the increase in the usage and efficiency of Artificial\nIntelligence and, more in general, of Automated Decision-Making systems has\nbrought with it an increasing and welcome awareness of the risks associated\nwith such systems. One of such risks is that of perpetuating or even amplifying\nbias and unjust disparities present in the data from which many of these\nsystems learn to adjust and optimise their decisions. This awareness has on the\none hand encouraged several scientific communities to come up with more and\nmore appropriate ways and methods to assess, quantify, and possibly mitigate\nsuch biases and disparities. On the other hand, it has prompted more and more\nlayers of society, including policy makers, to call for fair algorithms. We\nbelieve that while many excellent and multidisciplinary research is currently\nbeing conducted, what is still fundamentally missing is the awareness that\nhaving fair algorithms is per se a nearly meaningless requirement that needs to\nbe complemented with many additional social choices to become actionable.\nNamely, there is a hiatus between what the society is demanding from Automated\nDecision-Making systems, and what this demand actually means in real-world\nscenarios. In this work, we outline the key features of such a hiatus and\npinpoint a set of crucial open points that we as a society must address in\norder to give a concrete meaning to the increasing demand of fairness in\nAutomated Decision-Making systems.\n", "date": "2023-11-21", "categories": []}
{"id": "2404.04838", "title": "Data Bias According to Bipol: Men are Naturally Right and It is the Role\n  of Women to Follow Their Lead", "abstract": "  We introduce new large labeled datasets on bias in 3 languages and show in\nexperiments that bias exists in all 10 datasets of 5 languages evaluated,\nincluding benchmark datasets on the English GLUE/SuperGLUE leaderboards. The 3\nnew languages give a total of almost 6 million labeled samples and we benchmark\non these datasets using SotA multilingual pretrained models: mT5 and mBERT. The\nchallenge of social bias, based on prejudice, is ubiquitous, as recent events\nwith AI and large language models (LLMs) have shown. Motivated by this\nchallenge, we set out to estimate bias in multiple datasets. We compare some\nrecent bias metrics and use bipol, which has explainability in the metric. We\nalso confirm the unverified assumption that bias exists in toxic comments by\nrandomly sampling 200 samples from a toxic dataset population using the\nconfidence level of 95% and error margin of 7%. Thirty gold samples were\nrandomly distributed in the 200 samples to secure the quality of the\nannotation. Our findings confirm that many of the datasets have male bias\n(prejudice against women), besides other types of bias. We publicly release our\nnew datasets, lexica, models, and codes.\n", "date": "2024-04-07", "categories": []}
{"id": "2405.12926", "title": "Trusting Fair Data: Leveraging Quality in Fairness-Driven Data Removal\n  Techniques", "abstract": "  In this paper, we deal with bias mitigation techniques that remove specific\ndata points from the training set to aim for a fair representation of the\npopulation in that set. Machine learning models are trained on these\npre-processed datasets, and their predictions are expected to be fair. However,\nsuch approaches may exclude relevant data, making the attained subsets less\ntrustworthy for further usage. To enhance the trustworthiness of prior methods,\nwe propose additional requirements and objectives that the subsets must fulfill\nin addition to fairness: (1) group coverage, and (2) minimal data loss. While\nremoving entire groups may improve the measured fairness, this practice is very\nproblematic as failing to represent every group cannot be considered fair. In\nour second concern, we advocate for the retention of data while minimizing\ndiscrimination. By introducing a multi-objective optimization problem that\nconsiders fairness and data loss, we propose a methodology to find\nPareto-optimal solutions that balance these objectives. By identifying such\nsolutions, users can make informed decisions about the trade-off between\nfairness and data quality and select the most suitable subset for their\napplication. Our method is distributed as a Python package via PyPI under the\nname FairDo (https://github.com/mkduong-ai/fairdo).\n", "date": "2024-05-21", "categories": []}
{"id": "2409.13484", "title": "'Since Lawyers are Males..': Examining Implicit Gender Bias in Hindi\n  Language Generation by LLMs", "abstract": "  Large Language Models (LLMs) are increasingly being used to generate text\nacross various languages, for tasks such as translation, customer support, and\neducation. Despite these advancements, LLMs show notable gender biases in\nEnglish, which become even more pronounced when generating content in\nrelatively underrepresented languages like Hindi. This study explores implicit\ngender biases in Hindi text generation and compares them to those in English.\nWe developed Hindi datasets inspired by WinoBias to examine stereotypical\npatterns in responses from models like GPT-4o and Claude-3 sonnet. Our results\nreveal a significant gender bias of 87.8% in Hindi, compared to 33.4% in\nEnglish GPT-4o generation, with Hindi responses frequently relying on gender\nstereotypes related to occupations, power hierarchies, and social class. This\nresearch underscores the variation in gender biases across languages and\nprovides considerations for navigating these biases in generative AI systems.\n", "date": "2024-09-20", "categories": []}
{"id": "2409.12318", "title": "A large-scale study of performance and equity of commercial remote\n  identity verification technologies across demographics", "abstract": "  As more types of transactions move online, there is an increasing need to\nverify someone's identity remotely. Remote identity verification (RIdV)\ntechnologies have emerged to fill this need. RIdV solutions typically use a\nsmart device to validate an identity document like a driver's license by\ncomparing a face selfie to the face photo on the document. Recent research has\nbeen focused on ensuring that biometric systems work fairly across demographic\ngroups. This study assesses five commercial RIdV solutions for equity across\nage, gender, race/ethnicity, and skin tone across 3,991 test subjects. This\npaper employs statistical methods to discern whether the RIdV result across\ndemographic groups is statistically distinguishable. Two of the RIdV solutions\nwere equitable across all demographics, while two RIdV solutions had at least\none demographic that was inequitable. For example, the results for one\ntechnology had a false negative rate of 10.5% +/- 4.5% and its performance for\neach demographic category was within the error bounds, and, hence, were\nequitable. The other technologies saw either poor overall performance or\ninequitable performance. For one of these, participants of the race\nBlack/African American (B/AA) as well as those with darker skin tones (Monk\nscale 7/8/9/10) experienced higher false rejections. Finally, one technology\ndemonstrated more favorable but inequitable performance for the Asian American\nand Pacific Islander (AAPI) demographic. This study confirms that it is\nnecessary to evaluate products across demographic groups to fully understand\nthe performance of remote identity verification technologies.\n", "date": "2024-09-18", "categories": []}
{"id": "2409.12428", "title": "Is it Still Fair? A Comparative Evaluation of Fairness Algorithms\n  through the Lens of Covariate Drift", "abstract": "  Over the last few decades, machine learning (ML) applications have grown\nexponentially, yielding several benefits to society. However, these benefits\nare tempered with concerns of discriminatory behaviours exhibited by ML models.\nIn this regard, fairness in machine learning has emerged as a priority research\narea. Consequently, several fairness metrics and algorithms have been developed\nto mitigate against discriminatory behaviours that ML models may possess. Yet\nstill, very little attention has been paid to the problem of naturally\noccurring changes in data patterns (\\textit{aka} data distributional drift),\nand its impact on fairness algorithms and metrics. In this work, we study this\nproblem comprehensively by analyzing 4 fairness-unaware baseline algorithms and\n7 fairness-aware algorithms, carefully curated to cover the breadth of its\ntypology, across 5 datasets including public and proprietary data, and\nevaluated them using 3 predictive performance and 10 fairness metrics. In doing\nso, we show that (1) data distributional drift is not a trivial occurrence, and\nin several cases can lead to serious deterioration of fairness in so-called\nfair models; (2) contrary to some existing literature, the size and direction\nof data distributional drift is not correlated to the resulting size and\ndirection of unfairness; and (3) choice of, and training of fairness algorithms\nis impacted by the effect of data distributional drift which is largely ignored\nin the literature. Emanating from our findings, we synthesize several policy\nimplications of data distributional drift on fairness algorithms that can be\nvery relevant to stakeholders and practitioners.\n", "date": "2024-09-19", "categories": []}
{"id": "2307.13081", "title": "Fairness Under Demographic Scarce Regime", "abstract": "  Most existing works on fairness assume the model has full access to\ndemographic information. However, there exist scenarios where demographic\ninformation is partially available because a record was not maintained\nthroughout data collection or for privacy reasons. This setting is known as\ndemographic scarce regime. Prior research has shown that training an attribute\nclassifier to replace the missing sensitive attributes (proxy) can still\nimprove fairness. However, using proxy-sensitive attributes worsens\nfairness-accuracy tradeoffs compared to true sensitive attributes. To address\nthis limitation, we propose a framework to build attribute classifiers that\nachieve better fairness-accuracy tradeoffs. Our method introduces uncertainty\nawareness in the attribute classifier and enforces fairness on samples with\ndemographic information inferred with the lowest uncertainty. We show\nempirically that enforcing fairness constraints on samples with uncertain\nsensitive attributes can negatively impact the fairness-accuracy tradeoff. Our\nexperiments on five datasets showed that the proposed framework yields models\nwith significantly better fairness-accuracy tradeoffs than classic attribute\nclassifiers. Surprisingly, our framework can outperform models trained with\nfairness constraints on the true sensitive attributes in most benchmarks. We\nalso show that these findings are consistent with other uncertainty measures\nsuch as conformal prediction.\n", "date": "2023-07-24", "categories": []}
{"id": "2405.17512", "title": "On Fairness of Low-Rank Adaptation of Large Models", "abstract": "  Low-rank adaptation of large models, particularly LoRA, has gained traction\ndue to its computational efficiency. This efficiency, contrasted with the\nprohibitive costs of full-model fine-tuning, means that practitioners often\nturn to LoRA and sometimes without a complete understanding of its\nramifications. In this study, we focus on fairness and ask whether LoRA has an\nunexamined impact on utility, calibration, and resistance to membership\ninference across different subgroups (e.g., genders, races, religions) compared\nto a full-model fine-tuning baseline. We present extensive experiments across\nvision and language domains and across classification and generation tasks\nusing ViT-Base, Swin-v2-Large, Llama-2 7B, and Mistral 7B. Intriguingly,\nexperiments suggest that while one can isolate cases where LoRA exacerbates\nmodel bias across subgroups, the pattern is inconsistent -- in many cases, LoRA\nhas equivalent or even improved fairness compared to the base model or its full\nfine-tuning baseline. We also examine the complications of evaluating\nfine-tuning fairness relating to task design and model token bias, calling for\nmore careful fairness evaluations in future work.\n", "date": "2024-05-27", "categories": []}
{"id": "2401.13867", "title": "Unmasking and Quantifying Racial Bias of Large Language Models in\n  Medical Report Generation", "abstract": "  Large language models like GPT-3.5-turbo and GPT-4 hold promise for\nhealthcare professionals, but they may inadvertently inherit biases during\ntheir training, potentially affecting their utility in medical applications.\nDespite few attempts in the past, the precise impact and extent of these biases\nremain uncertain. Through both qualitative and quantitative analyses, we find\nthat these models tend to project higher costs and longer hospitalizations for\nWhite populations and exhibit optimistic views in challenging medical scenarios\nwith much higher survival rates. These biases, which mirror real-world\nhealthcare disparities, are evident in the generation of patient backgrounds,\nthe association of specific diseases with certain races, and disparities in\ntreatment recommendations, etc. Our findings underscore the critical need for\nfuture research to address and mitigate biases in language models, especially\nin critical healthcare applications, to ensure fair and accurate outcomes for\nall patients.\n", "date": "2024-01-25", "categories": []}
{"id": "2402.08250", "title": "A survey of recent methods for addressing AI fairness and bias in\n  biomedicine", "abstract": "  Artificial intelligence (AI) systems have the potential to revolutionize\nclinical practices, including improving diagnostic accuracy and surgical\ndecision-making, while also reducing costs and manpower. However, it is\nimportant to recognize that these systems may perpetuate social inequities or\ndemonstrate biases, such as those based on race or gender. Such biases can\noccur before, during, or after the development of AI models, making it critical\nto understand and address potential biases to enable the accurate and reliable\napplication of AI models in clinical settings. To mitigate bias concerns during\nmodel development, we surveyed recent publications on different debiasing\nmethods in the fields of biomedical natural language processing (NLP) or\ncomputer vision (CV). Then we discussed the methods that have been applied in\nthe biomedical domain to address bias. We performed our literature search on\nPubMed, ACM digital library, and IEEE Xplore of relevant articles published\nbetween January 2018 and December 2023 using multiple combinations of keywords.\nWe then filtered the result of 10,041 articles automatically with loose\nconstraints, and manually inspected the abstracts of the remaining 890 articles\nto identify the 55 articles included in this review. Additional articles in the\nreferences are also included in this review. We discuss each method and compare\nits strengths and weaknesses. Finally, we review other potential methods from\nthe general domain that could be applied to biomedicine to address bias and\nimprove fairness.The bias of AIs in biomedicine can originate from multiple\nsources. Existing debiasing methods that focus on algorithms can be categorized\ninto distributional or algorithmic.\n", "date": "2024-02-13", "categories": []}
{"id": "2406.08818", "title": "Linguistic Bias in ChatGPT: Language Models Reinforce Dialect\n  Discrimination", "abstract": "  We present a large-scale study of linguistic bias exhibited by ChatGPT\ncovering ten dialects of English (Standard American English, Standard British\nEnglish, and eight widely spoken non-\"standard\" varieties from around the\nworld). We prompted GPT-3.5 Turbo and GPT-4 with text by native speakers of\neach variety and analyzed the responses via detailed linguistic feature\nannotation and native speaker evaluation. We find that the models default to\n\"standard\" varieties of English; based on evaluation by native speakers, we\nalso find that model responses to non-\"standard\" varieties consistently exhibit\na range of issues: stereotyping (19% worse than for \"standard\" varieties),\ndemeaning content (25% worse), lack of comprehension (9% worse), and\ncondescending responses (15% worse). We also find that if these models are\nasked to imitate the writing style of prompts in non-\"standard\" varieties, they\nproduce text that exhibits lower comprehension of the input and is especially\nprone to stereotyping. GPT-4 improves on GPT-3.5 in terms of comprehension,\nwarmth, and friendliness, but also exhibits a marked increase in stereotyping\n(+18%). The results indicate that GPT-3.5 Turbo and GPT-4 can perpetuate\nlinguistic discrimination toward speakers of non-\"standard\" varieties.\n", "date": "2024-06-13", "categories": []}
{"id": "2409.04652", "title": "Privacy-Preserving Race/Ethnicity Estimation for Algorithmic Bias\n  Measurement in the U.S", "abstract": "  AI fairness measurements, including tests for equal treatment, often take the\nform of disaggregated evaluations of AI systems. Such measurements are an\nimportant part of Responsible AI operations. These measurements compare system\nperformance across demographic groups or sub-populations and typically require\nmember-level demographic signals such as gender, race, ethnicity, and location.\nHowever, sensitive member-level demographic attributes like race and ethnicity\ncan be challenging to obtain and use due to platform choices, legal\nconstraints, and cultural norms. In this paper, we focus on the task of\nenabling AI fairness measurements on race/ethnicity for \\emph{U.S. LinkedIn\nmembers} in a privacy-preserving manner. We present the Privacy-Preserving\nProbabilistic Race/Ethnicity Estimation (PPRE) method for performing this task.\nPPRE combines the Bayesian Improved Surname Geocoding (BISG) model, a sparse\nLinkedIn survey sample of self-reported demographics, and privacy-enhancing\ntechnologies like secure two-party computation and differential privacy to\nenable meaningful fairness measurements while preserving member privacy. We\nprovide details of the PPRE method and its privacy guarantees. We then\nillustrate sample measurement operations. We conclude with a review of open\nresearch and engineering challenges for expanding our privacy-preserving\nfairness measurement capabilities.\n", "date": "2024-09-06", "categories": []}
{"id": "2409.10538", "title": "Fairness in Survival Analysis with Distributionally Robust Optimization", "abstract": "  We propose a general approach for encouraging fairness in survival analysis\nmodels based on minimizing a worst-case error across all subpopulations that\noccur with at least a user-specified probability. This approach can be used to\nconvert many existing survival analysis models into ones that simultaneously\nencourage fairness, without requiring the user to specify which attributes or\nfeatures to treat as sensitive in the training loss function. From a technical\nstandpoint, our approach applies recent developments of distributionally robust\noptimization (DRO) to survival analysis. The complication is that existing DRO\ntheory uses a training loss function that decomposes across contributions of\nindividual data points, i.e., any term that shows up in the loss function\ndepends only on a single training point. This decomposition does not hold for\ncommonly used survival loss functions, including for the Cox proportional\nhazards model, its deep neural network variants, and many other recently\ndeveloped models that use loss functions involving ranking or similarity score\ncalculations. We address this technical hurdle using a sample splitting\nstrategy. We demonstrate our sample splitting DRO approach by using it to\ncreate fair versions of a diverse set of existing survival analysis models\nincluding the Cox model (and its deep variant DeepSurv), the discrete-time\nmodel DeepHit, and the neural ODE model SODEN. We also establish a\nfinite-sample theoretical guarantee to show what our sample splitting DRO loss\nconverges to. For the Cox model, we further derive an exact DRO approach that\ndoes not use sample splitting. For all the models that we convert into DRO\nvariants, we show that the DRO variants often score better on recently\nestablished fairness metrics (without incurring a significant drop in accuracy)\ncompared to existing survival analysis fairness regularization techniques.\n", "date": "2024-08-31", "categories": []}
{"id": "2409.10951", "title": "Fair Anomaly Detection For Imbalanced Groups", "abstract": "  Anomaly detection (AD) has been widely studied for decades in many real-world\napplications, including fraud detection in finance, and intrusion detection for\ncybersecurity, etc. Due to the imbalanced nature between protected and\nunprotected groups and the imbalanced distributions of normal examples and\nanomalies, the learning objectives of most existing anomaly detection methods\ntend to solely concentrate on the dominating unprotected group. Thus, it has\nbeen recognized by many researchers about the significance of ensuring model\nfairness in anomaly detection. However, the existing fair anomaly detection\nmethods tend to erroneously label most normal examples from the protected group\nas anomalies in the imbalanced scenario where the unprotected group is more\nabundant than the protected group. This phenomenon is caused by the improper\ndesign of learning objectives, which statistically focus on learning the\nfrequent patterns (i.e., the unprotected group) while overlooking the\nunder-represented patterns (i.e., the protected group). To address these\nissues, we propose FairAD, a fairness-aware anomaly detection method targeting\nthe imbalanced scenario. It consists of a fairness-aware contrastive learning\nmodule and a rebalancing autoencoder module to ensure fairness and handle the\nimbalanced data issue, respectively. Moreover, we provide the theoretical\nanalysis that shows our proposed contrastive learning regularization guarantees\ngroup fairness. Empirical studies demonstrate the effectiveness and efficiency\nof FairAD across multiple real-world datasets.\n", "date": "2024-09-17", "categories": []}
{"id": "2409.09569", "title": "Bias Begets Bias: The Impact of Biased Embeddings on Diffusion Models", "abstract": "  With the growing adoption of Text-to-Image (TTI) systems, the social biases\nof these models have come under increased scrutiny. Herein we conduct a\nsystematic investigation of one such source of bias for diffusion models:\nembedding spaces. First, because traditional classifier-based fairness\ndefinitions require true labels not present in generative modeling, we propose\nstatistical group fairness criteria based on a model's internal representation\nof the world. Using these definitions, we demonstrate theoretically and\nempirically that an unbiased text embedding space for input prompts is a\nnecessary condition for representationally balanced diffusion models, meaning\nthe distribution of generated images satisfy diversity requirements with\nrespect to protected attributes. Next, we investigate the impact of biased\nembeddings on evaluating the alignment between generated images and prompts, a\nprocess which is commonly used to assess diffusion models. We find that biased\nmultimodal embeddings like CLIP can result in lower alignment scores for\nrepresentationally balanced TTI models, thus rewarding unfair behavior.\nFinally, we develop a theoretical framework through which biases in alignment\nevaluation can be studied and propose bias mitigation methods. By specifically\nadapting the perspective of embedding spaces, we establish new fairness\nconditions for diffusion model development and evaluation.\n", "date": "2024-09-15", "categories": []}
{"id": "2409.09652", "title": "Unveiling Gender Bias in Large Language Models: Using Teacher's\n  Evaluation in Higher Education As an Example", "abstract": "  This paper investigates gender bias in Large Language Model (LLM)-generated\nteacher evaluations in higher education setting, focusing on evaluations\nproduced by GPT-4 across six academic subjects. By applying a comprehensive\nanalytical framework that includes Odds Ratio (OR) analysis, Word Embedding\nAssociation Test (WEAT), sentiment analysis, and contextual analysis, this\npaper identified patterns of gender-associated language reflecting societal\nstereotypes. Specifically, words related to approachability and support were\nused more frequently for female instructors, while words related to\nentertainment were predominantly used for male instructors, aligning with the\nconcepts of communal and agentic behaviors. The study also found moderate to\nstrong associations between male salient adjectives and male names, though\ncareer and family words did not distinctly capture gender biases. These\nfindings align with prior research on societal norms and stereotypes,\nreinforcing the notion that LLM-generated text reflects existing biases.\n", "date": "2024-09-15", "categories": []}
{"id": "2402.16158", "title": "Distribution-Free Fair Federated Learning with Small Samples", "abstract": "  As federated learning gains increasing importance in real-world applications\ndue to its capacity for decentralized data training, addressing fairness\nconcerns across demographic groups becomes critically important. However, most\nexisting machine learning algorithms for ensuring fairness are designed for\ncentralized data environments and generally require large-sample and\ndistributional assumptions, underscoring the urgent need for fairness\ntechniques adapted for decentralized and heterogeneous systems with\nfinite-sample and distribution-free guarantees. To address this issue, this\npaper introduces FedFaiREE, a post-processing algorithm developed specifically\nfor distribution-free fair learning in decentralized settings with small\nsamples. Our approach accounts for unique challenges in decentralized\nenvironments, such as client heterogeneity, communication costs, and small\nsample sizes. We provide rigorous theoretical guarantees for both fairness and\naccuracy, and our experimental results further provide robust empirical\nvalidation for our proposed method.\n", "date": "2024-02-25", "categories": []}
{"id": "2407.13070", "title": "The Cost of Arbitrariness for Individuals: Examining the Legal and\n  Technical Challenges of Model Multiplicity", "abstract": "  Model multiplicity, the phenomenon where multiple models achieve similar\nperformance despite different underlying learned functions, introduces\narbitrariness in model selection. While this arbitrariness may seem\ninconsequential in expectation, its impact on individuals can be severe. This\npaper explores various individual concerns stemming from multiplicity,\nincluding the effects of arbitrariness beyond final predictions, disparate\narbitrariness for individuals belonging to protected groups, and the challenges\nassociated with the arbitrariness of a single algorithmic system creating a\nmonopoly across various contexts. It provides both an empirical examination of\nthese concerns and a comprehensive analysis from the legal standpoint,\naddressing how these issues are perceived in the anti-discrimination law in\nCanada. We conclude the discussion with technical challenges in the current\nlandscape of model multiplicity to meet legal requirements and the legal gap\nbetween current law and the implications of arbitrariness in model selection,\nhighlighting relevant future research directions for both disciplines.\n", "date": "2024-05-28", "categories": []}
{"id": "2409.08658", "title": "Promoting Fairness in Link Prediction with Graph Enhancement", "abstract": "  Link prediction is a crucial task in network analysis, but it has been shown\nto be prone to biased predictions, particularly when links are unfairly\npredicted between nodes from different sensitive groups. In this paper, we\nstudy the fair link prediction problem, which aims to ensure that the predicted\nlink probability is independent of the sensitive attributes of the connected\nnodes. Existing methods typically incorporate debiasing techniques within graph\nembeddings to mitigate this issue. However, training on large real-world graphs\nis already challenging, and adding fairness constraints can further complicate\nthe process. To overcome this challenge, we propose FairLink, a method that\nlearns a fairness-enhanced graph to bypass the need for debiasing during the\nlink predictor's training. FairLink maintains link prediction accuracy by\nensuring that the enhanced graph follows a training trajectory similar to that\nof the original input graph. Meanwhile, it enhances fairness by minimizing the\nabsolute difference in link probabilities between node pairs within the same\nsensitive group and those between node pairs from different sensitive groups.\nOur extensive experiments on multiple large-scale graphs demonstrate that\nFairLink not only promotes fairness but also often achieves link prediction\naccuracy comparable to baseline methods. Most importantly, the enhanced graph\nexhibits strong generalizability across different GNN architectures.\n", "date": "2024-09-13", "categories": []}
{"id": "2305.10566", "title": "Smiling Women Pitching Down: Auditing Representational and\n  Presentational Gender Biases in Image Generative AI", "abstract": "  Generative AI models like DALL-E 2 can interpret textual prompts and generate\nhigh-quality images exhibiting human creativity. Though public enthusiasm is\nbooming, systematic auditing of potential gender biases in AI-generated images\nremains scarce. We addressed this gap by examining the prevalence of two\noccupational gender biases (representational and presentational biases) in\n15,300 DALL-E 2 images spanning 153 occupations, and assessed potential bias\namplification by benchmarking against 2021 census labor statistics and Google\nImages. Our findings reveal that DALL-E 2 underrepresents women in\nmale-dominated fields while overrepresenting them in female-dominated\noccupations. Additionally, DALL-E 2 images tend to depict more women than men\nwith smiling faces and downward-pitching heads, particularly in\nfemale-dominated (vs. male-dominated) occupations. Our computational algorithm\nauditing study demonstrates more pronounced representational and presentational\nbiases in DALL-E 2 compared to Google Images and calls for feminist\ninterventions to prevent such bias-laden AI-generated images to feedback into\nthe media ecology.\n", "date": "2023-05-17", "categories": []}
{"id": "2409.05990", "title": "FairHome: A Fair Housing and Fair Lending Dataset", "abstract": "  We present a Fair Housing and Fair Lending dataset (FairHome): A dataset with\naround 75,000 examples across 9 protected categories. To the best of our\nknowledge, FairHome is the first publicly available dataset labeled with binary\nlabels for compliance risk in the housing domain. We demonstrate the usefulness\nand effectiveness of such a dataset by training a classifier and using it to\ndetect potential violations when using a large language model (LLM) in the\ncontext of real-estate transactions. We benchmark the trained classifier\nagainst state-of-the-art LLMs including GPT-3.5, GPT-4, LLaMA-3, and Mistral\nLarge in both zero-shot and few-shot contexts. Our classifier outperformed with\nan F1-score of 0.91, underscoring the effectiveness of our dataset.\n", "date": "2024-09-09", "categories": []}
{"id": "2409.06916", "title": "Interactive Counterfactual Exploration of Algorithmic Harms in\n  Recommender Systems", "abstract": "  Recommender systems have become integral to digital experiences, shaping user\ninteractions and preferences across various platforms. Despite their widespread\nuse, these systems often suffer from algorithmic biases that can lead to unfair\nand unsatisfactory user experiences. This study introduces an interactive tool\ndesigned to help users comprehend and explore the impacts of algorithmic harms\nin recommender systems. By leveraging visualizations, counterfactual\nexplanations, and interactive modules, the tool allows users to investigate how\nbiases such as miscalibration, stereotypes, and filter bubbles affect their\nrecommendations. Informed by in-depth user interviews, this tool benefits both\ngeneral users and researchers by increasing transparency and offering\npersonalized impact assessments, ultimately fostering a better understanding of\nalgorithmic biases and contributing to more equitable recommendation outcomes.\nThis work provides valuable insights for future research and practical\napplications in mitigating bias and enhancing fairness in machine learning\nalgorithms.\n", "date": "2024-09-10", "categories": []}
{"id": "2409.06043", "title": "Identifying the sources of ideological bias in GPT models through\n  linguistic variation in output", "abstract": "  Extant work shows that generative AI models such as GPT-3.5 and 4 perpetuate\nsocial stereotypes and biases. One concerning but less explored source of bias\nis ideology. Do GPT models take ideological stances on politically sensitive\ntopics? In this article, we provide an original approach to identifying\nideological bias in generative models, showing that bias can stem from both the\ntraining data and the filtering algorithm. We leverage linguistic variation in\ncountries with contrasting political attitudes to evaluate bias in average GPT\nresponses to sensitive political topics in those languages. First, we find that\nGPT output is more conservative in languages that map well onto conservative\nsocieties (i.e., Polish), and more liberal in languages used uniquely in\nliberal societies (i.e., Swedish). This result provides strong evidence of\ntraining data bias in GPT models. Second, differences across languages observed\nin GPT-3.5 persist in GPT-4, even though GPT-4 is significantly more liberal\ndue to OpenAI's filtering policy. Our main takeaway is that generative model\ntraining must focus on high-quality, curated datasets to reduce bias, even if\nit entails a compromise in training data size. Filtering responses after\ntraining only introduces new biases and does not remove the underlying training\nbiases.\n", "date": "2024-09-09", "categories": []}
{"id": "2406.19097", "title": "Fairness and Bias in Multimodal AI: A Survey", "abstract": "  The importance of addressing fairness and bias in artificial intelligence\n(AI) systems cannot be over-emphasized. Mainstream media has been awashed with\nnews of incidents around stereotypes and other types of bias in many of these\nsystems in recent years. In this survey, we fill a gap with regards to the\nrelatively minimal study of fairness and bias in Large Multimodal Models (LMMs)\ncompared to Large Language Models (LLMs), providing 50 examples of datasets and\nmodels related to both types of AI along with the challenges of bias affecting\nthem. We discuss the less-mentioned category of mitigating bias, preprocessing\n(with particular attention on the first part of it, which we call preuse). The\nmethod is less-mentioned compared to the two well-known ones in the literature:\nintrinsic and extrinsic mitigation methods. We critically discuss the various\nways researchers are addressing these challenges. Our method involved two\nslightly different search queries on two reputable search engines, Google\nScholar and Web of Science (WoS), which revealed that for the queries 'Fairness\nand bias in Large Multimodal Models' and 'Fairness and bias in Large Language\nModels', 33,400 and 538,000 links are the initial results, respectively, for\nScholar while 4 and 50 links are the initial results, respectively, for WoS.\nFor reproducibility and verification, we provide links to the search results\nand the citations to all the final reviewed papers. We believe this work\ncontributes to filling this gap and providing insight to researchers and other\nstakeholders on ways to address the challenges of fairness and bias in\nmultimodal and language AI.\n", "date": "2024-06-27", "categories": []}
{"id": "2409.03893", "title": "Understanding Fairness in Recommender Systems: A Healthcare Perspective", "abstract": "  Fairness in AI-driven decision-making systems has become a critical concern,\nespecially when these systems directly affect human lives. This paper explores\nthe public's comprehension of fairness in healthcare recommendations. We\nconducted a survey where participants selected from four fairness metrics --\nDemographic Parity, Equal Accuracy, Equalized Odds, and Positive Predictive\nValue -- across different healthcare scenarios to assess their understanding of\nthese concepts. Our findings reveal that fairness is a complex and often\nmisunderstood concept, with a generally low level of public understanding\nregarding fairness metrics in recommender systems. This study highlights the\nneed for enhanced information and education on algorithmic fairness to support\ninformed decision-making in using these systems. Furthermore, the results\nsuggest that a one-size-fits-all approach to fairness may be insufficient,\npointing to the importance of context-sensitive designs in developing equitable\nAI systems.\n", "date": "2024-09-05", "categories": []}
{"id": "2409.04897", "title": "Centralized Selection with Preferences in the Presence of Biases", "abstract": "  This paper considers the scenario in which there are multiple institutions,\neach with a limited capacity for candidates, and candidates, each with\npreferences over the institutions. A central entity evaluates the utility of\neach candidate to the institutions, and the goal is to select candidates for\neach institution in a way that maximizes utility while also considering the\ncandidates' preferences. The paper focuses on the setting in which candidates\nare divided into multiple groups and the observed utilities of candidates in\nsome groups are biased--systematically lower than their true utilities. The\nfirst result is that, in these biased settings, prior algorithms can lead to\nselections with sub-optimal true utility and significant discrepancies in the\nfraction of candidates from each group that get their preferred choices.\nSubsequently, an algorithm is presented along with proof that it produces\nselections that achieve near-optimal group fairness with respect to preferences\nwhile also nearly maximizing the true utility under distributional assumptions.\nFurther, extensive empirical validation of these results in real-world and\nsynthetic settings, in which the distributional assumptions may not hold, are\npresented.\n", "date": "2024-09-07", "categories": []}
{"id": "2409.05215", "title": "Synthetic Tabular Data Generation for Class Imbalance and Fairness: A\n  Comparative Study", "abstract": "  Due to their data-driven nature, Machine Learning (ML) models are susceptible\nto bias inherited from data, especially in classification problems where class\nand group imbalances are prevalent. Class imbalance (in the classification\ntarget) and group imbalance (in protected attributes like sex or race) can\nundermine both ML utility and fairness. Although class and group imbalances\ncommonly coincide in real-world tabular datasets, limited methods address this\nscenario. While most methods use oversampling techniques, like interpolation,\nto mitigate imbalances, recent advancements in synthetic tabular data\ngeneration offer promise but have not been adequately explored for this\npurpose. To this end, this paper conducts a comparative analysis to address\nclass and group imbalances using state-of-the-art models for synthetic tabular\ndata generation and various sampling strategies. Experimental results on four\ndatasets, demonstrate the effectiveness of generative models for bias\nmitigation, creating opportunities for further exploration in this direction.\n", "date": "2024-09-08", "categories": []}
{"id": "2409.04340", "title": "AGR: Age Group fairness Reward for Bias Mitigation in LLMs", "abstract": "  LLMs can exhibit age biases, resulting in unequal treatment of individuals\nacross age groups. While much research has addressed racial and gender biases,\nage bias remains little explored. The scarcity of instruction-tuning and\npreference datasets for age bias hampers its detection and measurement, and\nexisting fine-tuning methods seldom address age-related fairness. In this\npaper, we construct age bias preference datasets and instruction-tuning\ndatasets for RLHF. We introduce ARG, an age fairness reward to reduce\ndifferences in the response quality of LLMs across different age groups.\nExtensive experiments demonstrate that this reward significantly improves\nresponse accuracy and reduces performance disparities across age groups. Our\nsource code and datasets are available at the anonymous\n\\href{https://anonymous.4open.science/r/FairRLHF-D445/readme.md}{link}.\n", "date": "2024-09-06", "categories": []}
{"id": "2409.04373", "title": "Evaluating Fairness in Transaction Fraud Models: Fairness Metrics, Bias\n  Audits, and Challenges", "abstract": "  Ensuring fairness in transaction fraud detection models is vital due to the\npotential harms and legal implications of biased decision-making. Despite\nextensive research on algorithmic fairness, there is a notable gap in the study\nof bias in fraud detection models, mainly due to the field's unique challenges.\nThese challenges include the need for fairness metrics that account for fraud\ndata's imbalanced nature and the tradeoff between fraud protection and service\nquality. To address this gap, we present a comprehensive fairness evaluation of\ntransaction fraud models using public synthetic datasets, marking the first\nalgorithmic bias audit in this domain. Our findings reveal three critical\ninsights: (1) Certain fairness metrics expose significant bias only after\nnormalization, highlighting the impact of class imbalance. (2) Bias is\nsignificant in both service quality-related parity metrics and fraud\nprotection-related parity metrics. (3) The fairness through unawareness\napproach, which involved removing sensitive attributes such as gender, does not\nimprove bias mitigation within these datasets, likely due to the presence of\ncorrelated proxies. We also discuss socio-technical fairness-related challenges\nin transaction fraud models. These insights underscore the need for a nuanced\napproach to fairness in fraud detection, balancing protection and service\nquality, and moving beyond simple bias mitigation strategies. Future work must\nfocus on refining fairness metrics and developing methods tailored to the\nunique complexities of the transaction fraud domain.\n", "date": "2024-09-06", "categories": []}
{"id": "2409.03303", "title": "Improving Robustness to Multiple Spurious Correlations by\n  Multi-Objective Optimization", "abstract": "  We study the problem of training an unbiased and accurate model given a\ndataset with multiple biases. This problem is challenging since the multiple\nbiases cause multiple undesirable shortcuts during training, and even worse,\nmitigating one may exacerbate the other. We propose a novel training method to\ntackle this challenge. Our method first groups training data so that different\ngroups induce different shortcuts, and then optimizes a linear combination of\ngroup-wise losses while adjusting their weights dynamically to alleviate\nconflicts between the groups in performance; this approach, rooted in the\nmulti-objective optimization theory, encourages to achieve the minimax Pareto\nsolution. We also present a new benchmark with multiple biases, dubbed\nMultiCelebA, for evaluating debiased training methods under realistic and\nchallenging scenarios. Our method achieved the best on three datasets with\nmultiple biases, and also showed superior performance on conventional\nsingle-bias datasets.\n", "date": "2024-09-05", "categories": []}
{"id": "2409.02152", "title": "Fair Railway Network Design", "abstract": "  When designing a public transportation network in a country, one may want to\nminimise the sum of travel duration of all inhabitants. This corresponds to a\npurely utilitarian view and does not involve any fairness consideration, as the\nresulting network will typically benefit the capital city and/or large central\ncities while leaving some peripheral cities behind. On the other hand, a more\negalitarian view will allow some people to travel between peripheral cities\nwithout having to go through a central city. We define a model, propose\nalgorithms for computing solution networks, and report on experiments based on\nreal data.\n", "date": "2024-09-03", "categories": []}
{"id": "2409.02471", "title": "Demographic parity in regression and classification within the\n  unawareness framework", "abstract": "  This paper explores the theoretical foundations of fair regression under the\nconstraint of demographic parity within the unawareness framework, where\ndisparate treatment is prohibited, extending existing results where such\ntreatment is permitted. Specifically, we aim to characterize the optimal fair\nregression function when minimizing the quadratic loss. Our results reveal that\nthis function is given by the solution to a barycenter problem with optimal\ntransport costs. Additionally, we study the connection between optimal fair\ncost-sensitive classification, and optimal fair regression. We demonstrate that\nnestedness of the decision sets of the classifiers is both necessary and\nsufficient to establish a form of equivalence between classification and\nregression. Under this nestedness assumption, the optimal classifiers can be\nderived by applying thresholds to the optimal fair regression function;\nconversely, the optimal fair regression function is characterized by the family\nof cost-sensitive classifiers.\n", "date": "2024-09-04", "categories": []}
{"id": "2312.03886", "title": "On The Fairness Impacts of Hardware Selection in Machine Learning", "abstract": "  In the machine learning ecosystem, hardware selection is often regarded as a\nmere utility, overshadowed by the spotlight on algorithms and data. This\noversight is particularly problematic in contexts like ML-as-a-service\nplatforms, where users often lack control over the hardware used for model\ndeployment. How does the choice of hardware impact generalization properties?\nThis paper investigates the influence of hardware on the delicate balance\nbetween model performance and fairness. We demonstrate that hardware choices\ncan exacerbate existing disparities, attributing these discrepancies to\nvariations in gradient flows and loss surfaces across different demographic\ngroups. Through both theoretical and empirical analysis, the paper not only\nidentifies the underlying factors but also proposes an effective strategy for\nmitigating hardware-induced performance imbalances.\n", "date": "2023-12-06", "categories": []}
{"id": "2403.08564", "title": "Generalizing Fairness to Generative Language Models via Reformulation of\n  Non-discrimination Criteria", "abstract": "  Generative AI, such as large language models, has undergone rapid development\nwithin recent years. As these models become increasingly available to the\npublic, concerns arise about perpetuating and amplifying harmful biases in\napplications. Gender stereotypes can be harmful and limiting for the\nindividuals they target, whether they consist of misrepresentation or\ndiscrimination. Recognizing gender bias as a pervasive societal construct, this\npaper studies how to uncover and quantify the presence of gender biases in\ngenerative language models. In particular, we derive generative AI analogues of\nthree well-known non-discrimination criteria from classification, namely\nindependence, separation and sufficiency. To demonstrate these criteria in\naction, we design prompts for each of the criteria with a focus on occupational\ngender stereotype, specifically utilizing the medical test to introduce the\nground truth in the generative AI context. Our results address the presence of\noccupational gender bias within such conversational language models.\n", "date": "2024-03-13", "categories": []}
{"id": "2404.03833", "title": "An ExplainableFair Framework for Prediction of Substance Use Disorder\n  Treatment Completion", "abstract": "  Fairness of machine learning models in healthcare has drawn increasing\nattention from clinicians, researchers, and even at the highest level of\ngovernment. On the other hand, the importance of developing and deploying\ninterpretable or explainable models has been demonstrated, and is essential to\nincreasing the trustworthiness and likelihood of adoption of these models. The\nobjective of this study was to develop and implement a framework for addressing\nboth these issues - fairness and explainability. We propose an explainable\nfairness framework, first developing a model with optimized performance, and\nthen using an in-processing approach to mitigate model biases relative to the\nsensitive attributes of race and sex. We then explore and visualize\nexplanations of the model changes that lead to the fairness enhancement process\nthrough exploring the changes in importance of features. Our resulting-fairness\nenhanced models retain high sensitivity with improved fairness and explanations\nof the fairness-enhancement that may provide helpful insights for healthcare\nproviders to guide clinical decision-making and resource allocation.\n", "date": "2024-04-04", "categories": []}
{"id": "2406.14281", "title": "FairX: A comprehensive benchmarking tool for model analysis using\n  fairness, utility, and explainability", "abstract": "  We present FairX, an open-source Python-based benchmarking tool designed for\nthe comprehensive analysis of models under the umbrella of fairness, utility,\nand eXplainability (XAI). FairX enables users to train benchmarking\nbias-mitigation models and evaluate their fairness using a wide array of\nfairness metrics, data utility metrics, and generate explanations for model\npredictions, all within a unified framework. Existing benchmarking tools do not\nhave the way to evaluate synthetic data generated from fair generative models,\nalso they do not have the support for training fair generative models either.\nIn FairX, we add fair generative models in the collection of our fair-model\nlibrary (pre-processing, in-processing, post-processing) and evaluation metrics\nfor evaluating the quality of synthetic fair data. This version of FairX\nsupports both tabular and image datasets. It also allows users to provide their\nown custom datasets. The open-source FairX benchmarking package is publicly\navailable at \\url{https://github.com/fahim-sikder/FairX}.\n", "date": "2024-06-20", "categories": []}
{"id": "2407.04268", "title": "NeuFair: Neural Network Fairness Repair with Dropout", "abstract": "  This paper investigates neuron dropout as a post-processing bias mitigation\nfor deep neural networks (DNNs). Neural-driven software solutions are\nincreasingly applied in socially critical domains with significant fairness\nimplications. While neural networks are exceptionally good at finding\nstatistical patterns from data, they may encode and amplify existing biases\nfrom the historical data. Existing bias mitigation algorithms often require\nmodifying the input dataset or the learning algorithms. We posit that the\nprevalent dropout methods that prevent over-fitting during training by randomly\ndropping neurons may be an effective and less intrusive approach to improve the\nfairness of pre-trained DNNs. However, finding the ideal set of neurons to drop\nis a combinatorial problem. We propose NeuFair, a family of post-processing\nrandomized algorithms that mitigate unfairness in pre-trained DNNs via dropouts\nduring inference after training. Our randomized search is guided by an\nobjective to minimize discrimination while maintaining the model's utility. We\nshow that our design of randomized algorithms is effective and efficient in\nimproving fairness (up to 69%) with minimal or no model performance\ndegradation. We provide intuitive explanations of these phenomena and carefully\nexamine the influence of various hyperparameters of search algorithms on the\nresults. Finally, we empirically and conceptually compare NeuFair to different\nstate-of-the-art bias mitigators.\n", "date": "2024-07-05", "categories": []}
{"id": "2408.13295", "title": "Exploring Bias and Prediction Metrics to Characterise the Fairness of\n  Machine Learning for Equity-Centered Public Health Decision-Making: A\n  Narrative Review", "abstract": "  Background: The rapid advancement of Machine Learning (ML) represents novel\nopportunities to enhance public health research, surveillance, and\ndecision-making. However, there is a lack of comprehensive understanding of\nalgorithmic bias, systematic errors in predicted population health outcomes,\nresulting from the public health application of ML. The objective of this\nnarrative review is to explore the types of bias generated by ML and\nquantitative metrics to assess these biases.\n  Methods : We performed search on PubMed, MEDLINE, IEEE (Institute of\nElectrical and Electronics Engineers), ACM (Association for Computing\nMachinery) Digital Library, Science Direct, and Springer Nature. We used\nkeywords to identify studies describing types of bias and metrics to measure\nthese in the domain of ML and public and population health published in English\nbetween 2008 and 2023, inclusive.\n  Results: A total of 72 articles met the inclusion criteria. Our review\nidentified the commonly described types of bias and quantitative metrics to\nassess these biases from an equity perspective.\n  Conclusion : The review will help formalize the evaluation framework for ML\non public health from an equity perspective.\n", "date": "2024-08-23", "categories": []}
{"id": "2408.15550", "title": "Trustworthy and Responsible AI for Human-Centric Autonomous\n  Decision-Making Systems", "abstract": "  Artificial Intelligence (AI) has paved the way for revolutionary\ndecision-making processes, which if harnessed appropriately, can contribute to\nadvancements in various sectors, from healthcare to economics. However, its\nblack box nature presents significant ethical challenges related to bias and\ntransparency. AI applications are hugely impacted by biases, presenting\ninconsistent and unreliable findings, leading to significant costs and\nconsequences, highlighting and perpetuating inequalities and unequal access to\nresources. Hence, developing safe, reliable, ethical, and Trustworthy AI\nsystems is essential.\n  Our team of researchers working with Trustworthy and Responsible AI, part of\nthe Transdisciplinary Scholarship Initiative within the University of Calgary,\nconducts research on Trustworthy and Responsible AI, including fairness, bias\nmitigation, reproducibility, generalization, interpretability, and\nauthenticity. In this paper, we review and discuss the intricacies of AI\nbiases, definitions, methods of detection and mitigation, and metrics for\nevaluating bias. We also discuss open challenges with regard to the\ntrustworthiness and widespread application of AI across diverse domains of\nhuman-centric decision making, as well as guidelines to foster Responsible and\nTrustworthy AI models.\n", "date": "2024-08-28", "categories": []}
{"id": "2408.16154", "title": "Does Data-Efficient Generalization Exacerbate Bias in Foundation Models?", "abstract": "  Foundation models have emerged as robust models with label efficiency in\ndiverse domains. In medical imaging, these models contribute to the advancement\nof medical diagnoses due to the difficulty in obtaining labeled data. However,\nit is unclear whether using a large amount of unlabeled data, biased by the\npresence of sensitive attributes during pre-training, influences the fairness\nof the model. This research examines the bias in the Foundation model\n(RetFound) when it is applied to fine-tune the Brazilian Multilabel\nOphthalmological Dataset (BRSET), which has a different population than the\npre-training dataset. The model evaluation, in comparison with supervised\nlearning, shows that the Foundation Model has the potential to reduce the gap\nbetween the maximum AUC and minimum AUC evaluations across gender and age\ngroups. However, in a data-efficient generalization, the model increases the\nbias when the data amount decreases. These findings suggest that when deploying\na Foundation Model in real-life scenarios with limited data, the possibility of\nfairness issues should be considered.\n", "date": "2024-08-28", "categories": []}
{"id": "2409.01928", "title": "Comprehensive Equity Index (CEI): Definition and Application to Bias\n  Evaluation in Biometrics", "abstract": "  We present a novel metric designed, among other applications, to quantify\nbiased behaviors of machine learning models. As its core, the metric consists\nof a new similarity metric between score distributions that balances both their\ngeneral shapes and tails' probabilities. In that sense, our proposed metric may\nbe useful in many application areas. Here we focus on and apply it to the\noperational evaluation of face recognition systems, with special attention to\nquantifying demographic biases; an application where our metric is especially\nuseful. The topic of demographic bias and fairness in biometric recognition\nsystems has gained major attention in recent years. The usage of these systems\nhas spread in society, raising concerns about the extent to which these systems\ntreat different population groups. A relevant step to prevent and mitigate\ndemographic biases is first to detect and quantify them. Traditionally, two\napproaches have been studied to quantify differences between population groups\nin machine learning literature: 1) measuring differences in error rates, and 2)\nmeasuring differences in recognition score distributions. Our proposed\nComprehensive Equity Index (CEI) trade-offs both approaches combining both\nerrors from distribution tails and general distribution shapes. This new metric\nis well suited to real-world scenarios, as measured on NIST FRVT evaluations,\ninvolving high-performance systems and realistic face databases including a\nwide range of covariates and demographic groups. We first show the limitations\nof existing metrics to correctly assess the presence of biases in realistic\nsetups and then propose our new metric to tackle these limitations. We tested\nthe proposed metric with two state-of-the-art models and four widely used\ndatabases, showing its capacity to overcome the main flaws of previous bias\nmetrics.\n", "date": "2024-09-03", "categories": []}
{"id": "2408.17313", "title": "Fair Best Arm Identification with Fixed Confidence", "abstract": "  In this work, we present a novel framework for Best Arm Identification (BAI)\nunder fairness constraints, a setting that we refer to as \\textit{F-BAI} (fair\nBAI). Unlike traditional BAI, which solely focuses on identifying the optimal\narm with minimal sample complexity, F-BAI also includes a set of fairness\nconstraints. These constraints impose a lower limit on the selection rate of\neach arm and can be either model-agnostic or model-dependent. For this setting,\nwe establish an instance-specific sample complexity lower bound and analyze the\n\\textit{price of fairness}, quantifying how fairness impacts sample complexity.\nBased on the sample complexity lower bound, we propose F-TaS, an algorithm\nprovably matching the sample complexity lower bound, while ensuring that the\nfairness constraints are satisfied. Numerical results, conducted using both a\nsynthetic model and a practical wireless scheduling application, show the\nefficiency of F-TaS in minimizing the sample complexity while achieving low\nfairness violations.\n", "date": "2024-08-30", "categories": []}
{"id": "2401.04057", "title": "Unveiling Bias in Fairness Evaluations of Large Language Models: A\n  Critical Literature Review of Music and Movie Recommendation Systems", "abstract": "  The rise of generative artificial intelligence, particularly Large Language\nModels (LLMs), has intensified the imperative to scrutinize fairness alongside\naccuracy. Recent studies have begun to investigate fairness evaluations for\nLLMs within domains such as recommendations. Given that personalization is an\nintrinsic aspect of recommendation systems, its incorporation into fairness\nassessments is paramount. Yet, the degree to which current fairness evaluation\nframeworks account for personalization remains unclear. Our comprehensive\nliterature review aims to fill this gap by examining how existing frameworks\nhandle fairness evaluations of LLMs, with a focus on the integration of\npersonalization factors. Despite an exhaustive collection and analysis of\nrelevant works, we discovered that most evaluations overlook personalization, a\ncritical facet of recommendation systems, thereby inadvertently perpetuating\nunfair practices. Our findings shed light on this oversight and underscore the\nurgent need for more nuanced fairness evaluations that acknowledge\npersonalization. Such improvements are vital for fostering equitable\ndevelopment within the AI community.\n", "date": "2024-01-08", "categories": []}
{"id": "2308.11375", "title": "Standardized Interpretable Fairness Measures for Continuous Risk Scores", "abstract": "  We propose a standardized version of fairness measures for continuous scores\nwith a reasonable interpretation based on the Wasserstein distance. Our\nmeasures are easily computable and well suited for quantifying and interpreting\nthe strength of group disparities as well as for comparing biases across\ndifferent models, datasets, or time points. We derive a link between the\ndifferent families of existing fairness measures for scores and show that the\nproposed standardized fairness measures outperform ROC-based fairness measures\nbecause they are more explicit and can quantify significant biases that\nROC-based fairness measures miss.\n", "date": "2023-08-22", "categories": []}
{"id": "2312.11299", "title": "Uncertainty-based Fairness Measures", "abstract": "  Unfair predictions of machine learning (ML) models impede their broad\nacceptance in real-world settings. Tackling this arduous challenge first\nnecessitates defining what it means for an ML model to be fair. This has been\naddressed by the ML community with various measures of fairness that depend on\nthe prediction outcomes of the ML models, either at the group level or the\nindividual level. These fairness measures are limited in that they utilize\npoint predictions, neglecting their variances, or uncertainties, making them\nsusceptible to noise, missingness and shifts in data. In this paper, we first\nshow that an ML model may appear to be fair with existing point-based fairness\nmeasures but biased against a demographic group in terms of prediction\nuncertainties. Then, we introduce new fairness measures based on different\ntypes of uncertainties, namely, aleatoric uncertainty and epistemic\nuncertainty. We demonstrate on many datasets that (i) our uncertainty-based\nmeasures are complementary to existing measures of fairness, and (ii) they\nprovide more insights about the underlying issues leading to bias.\n", "date": "2023-12-18", "categories": []}
{"id": "2408.15096", "title": "Post-processing fairness with minimal changes", "abstract": "  In this paper, we introduce a novel post-processing algorithm that is both\nmodel-agnostic and does not require the sensitive attribute at test time. In\naddition, our algorithm is explicitly designed to enforce minimal changes\nbetween biased and debiased predictions; a property that, while highly\ndesirable, is rarely prioritized as an explicit objective in fairness\nliterature. Our approach leverages a multiplicative factor applied to the logit\nvalue of probability scores produced by a black-box classifier. We demonstrate\nthe efficacy of our method through empirical evaluations, comparing its\nperformance against other four debiasing algorithms on two widely used datasets\nin fairness research.\n", "date": "2024-08-27", "categories": []}
{"id": "2408.16040", "title": "Fairness, Accuracy, and Unreliable Data", "abstract": "  This thesis investigates three areas targeted at improving the reliability of\nmachine learning; fairness in machine learning, strategic classification, and\nalgorithmic robustness. Each of these domains has special properties or\nstructure that can complicate learning. A theme throughout this thesis is\nthinking about ways in which a `plain' empirical risk minimization algorithm\nwill be misleading or ineffective because of a mis-match between classical\nlearning theory assumptions and specific properties of some data distribution\nin the wild. Theoretical understanding in eachof these domains can help guide\nbest practices and allow for the design of effective, reliable, and robust\nsystems.\n", "date": "2024-08-28", "categories": []}
{"id": "2408.16088", "title": "Ensuring Equitable Financial Decisions: Leveraging Counterfactual\n  Fairness and Deep Learning for Bias", "abstract": "  Concerns regarding fairness and bias have been raised in recent years due to\nthe growing use of machine learning models in crucial decision-making\nprocesses, especially when it comes to delicate characteristics like gender. In\norder to address biases in machine learning models, this research paper\ninvestigates advanced bias mitigation techniques, with a particular focus on\ncounterfactual fairness in conjunction with data augmentation. The study looks\ninto how these integrated approaches can lessen gender bias in the financial\nindustry, specifically in loan approval procedures. We show that these\napproaches are effective in achieving more equitable results through thorough\ntesting and assessment on a skewed financial dataset. The findings emphasize\nhow crucial it is to use fairness-aware techniques when creating machine\nlearning models in order to guarantee morally righteous and impartial\ndecision-making.\n", "date": "2024-08-27", "categories": []}
{"id": "2408.16700", "title": "GradBias: Unveiling Word Influence on Bias in Text-to-Image Generative\n  Models", "abstract": "  Recent progress in Text-to-Image (T2I) generative models has enabled\nhigh-quality image generation. As performance and accessibility increase, these\nmodels are gaining significant attraction and popularity: ensuring their\nfairness and safety is a priority to prevent the dissemination and perpetuation\nof biases. However, existing studies in bias detection focus on closed sets of\npredefined biases (e.g., gender, ethnicity). In this paper, we propose a\ngeneral framework to identify, quantify, and explain biases in an open set\nsetting, i.e. without requiring a predefined set. This pipeline leverages a\nLarge Language Model (LLM) to propose biases starting from a set of captions.\nNext, these captions are used by the target generative model for generating a\nset of images. Finally, Vision Question Answering (VQA) is leveraged for bias\nevaluation. We show two variations of this framework: OpenBias and GradBias.\nOpenBias detects and quantifies biases, while GradBias determines the\ncontribution of individual prompt words on biases. OpenBias effectively detects\nboth well-known and novel biases related to people, objects, and animals and\nhighly aligns with existing closed-set bias detection methods and human\njudgment. GradBias shows that neutral words can significantly influence biases\nand it outperforms several baselines, including state-of-the-art foundation\nmodels. Code available here: https://github.com/Moreno98/GradBias.\n", "date": "2024-08-29", "categories": []}
{"id": "2408.00612", "title": "Downstream bias mitigation is all you need", "abstract": "  The advent of transformer-based architectures and large language models\n(LLMs) have significantly advanced the performance of natural language\nprocessing (NLP) models. Since these LLMs are trained on huge corpuses of data\nfrom the web and other sources, there has been a major concern about harmful\nprejudices that may potentially be transferred from the data. In many\napplications, these pre-trained LLMs are fine-tuned on task specific datasets,\nwhich can further contribute to biases. This paper studies the extent of biases\nabsorbed by LLMs during pre-training as well as task-specific behaviour after\nfine-tuning. We found that controlled interventions on pre-trained LLMs, prior\nto fine-tuning, have minimal effect on lowering biases in classifiers. However,\nthe biases present in domain-specific datasets play a much bigger role, and\nhence mitigating them at this stage has a bigger impact. While pre-training\ndoes matter, but after the model has been pre-trained, even slight changes to\nco-occurrence rates in the fine-tuning dataset has a significant effect on the\nbias of the model.\n", "date": "2024-08-01", "categories": []}
{"id": "2310.13820", "title": "FERI: A Multitask-based Fairness Achieving Algorithm with Applications\n  to Fair Organ Transplantation", "abstract": "  Liver transplantation often faces fairness challenges across subgroups\ndefined by sensitive attributes such as age group, gender, and race/ethnicity.\nMachine learning models for outcome prediction can introduce additional biases.\nTherefore, we introduce Fairness through the Equitable Rate of Improvement in\nMultitask Learning (FERI) algorithm for fair predictions of graft failure risk\nin liver transplant patients. FERI constrains subgroup loss by balancing\nlearning rates and preventing subgroup dominance in the training process. Our\nresults show that FERI maintained high predictive accuracy with AUROC and AUPRC\ncomparable to baseline models. More importantly, FERI demonstrated an ability\nto improve fairness without sacrificing accuracy. Specifically, for the gender,\nFERI reduced the demographic parity disparity by 71.74%, and for the age group,\nit decreased the equalized odds disparity by 40.46%. Therefore, the FERI\nalgorithm advanced fairness-aware predictive modeling in healthcare and\nprovides an invaluable tool for equitable healthcare systems.\n", "date": "2023-10-20", "categories": []}
{"id": "2404.04814", "title": "Inference-Time Rule Eraser: Fair Recognition via Distilling and Removing\n  Biased Rules", "abstract": "  Machine learning models often make predictions based on biased features such\nas gender, race, and other social attributes, posing significant fairness\nrisks, especially in societal applications, such as hiring, banking, and\ncriminal justice. Traditional approaches to addressing this issue involve\nretraining or fine-tuning neural networks with fairness-aware optimization\nobjectives. However, these methods can be impractical due to significant\ncomputational resources, complex industrial tests, and the associated CO2\nfootprint. Additionally, regular users often fail to fine-tune models because\nthey lack access to model parameters In this paper, we introduce the\nInference-Time Rule Eraser (Eraser), a novel method designed to address\nfairness concerns by removing biased decision-making rules from deployed models\nduring inference without altering model weights. We begin by establishing a\ntheoretical foundation for modifying model outputs to eliminate biased rules\nthrough Bayesian analysis. Next, we present a specific implementation of Eraser\nthat involves two stages: (1) distilling the biased rules from the deployed\nmodel into an additional patch model, and (2) removing these biased rules from\nthe output of the deployed model during inference. Extensive experiments\nvalidate the effectiveness of our approach, showcasing its superior performance\nin addressing fairness concerns in AI systems.\n", "date": "2024-04-07", "categories": []}
{"id": "2408.11247", "title": "Unboxing Occupational Bias: Grounded Debiasing of LLMs with U.S. Labor\n  Data", "abstract": "  Large Language Models (LLMs) are prone to inheriting and amplifying societal\nbiases embedded within their training data, potentially reinforcing harmful\nstereotypes related to gender, occupation, and other sensitive categories. This\nissue becomes particularly problematic as biased LLMs can have far-reaching\nconsequences, leading to unfair practices and exacerbating social inequalities\nacross various domains, such as recruitment, online content moderation, or even\nthe criminal justice system. Although prior research has focused on detecting\nbias in LLMs using specialized datasets designed to highlight intrinsic biases,\nthere has been a notable lack of investigation into how these findings\ncorrelate with authoritative datasets, such as those from the U.S. National\nBureau of Labor Statistics (NBLS). To address this gap, we conduct empirical\nresearch that evaluates LLMs in a ``bias-out-of-the-box\" setting, analyzing how\nthe generated outputs compare with the distributions found in NBLS data.\nFurthermore, we propose a straightforward yet effective debiasing mechanism\nthat directly incorporates NBLS instances to mitigate bias within LLMs. Our\nstudy spans seven different LLMs, including instructable, base, and\nmixture-of-expert models, and reveals significant levels of bias that are often\noverlooked by existing bias detection techniques. Importantly, our debiasing\nmethod, which does not rely on external datasets, demonstrates a substantial\nreduction in bias scores, highlighting the efficacy of our approach in creating\nfairer and more reliable LLMs.\n", "date": "2024-08-20", "categories": []}
{"id": "2408.13667", "title": "Outlier Detection Bias Busted: Understanding Sources of Algorithmic Bias\n  through Data-centric Factors", "abstract": "  The astonishing successes of ML have raised growing concern for the fairness\nof modern methods when deployed in real world settings. However, studies on\nfairness have mostly focused on supervised ML, while unsupervised outlier\ndetection (OD), with numerous applications in finance, security, etc., have\nattracted little attention. While a few studies proposed fairness-enhanced OD\nalgorithms, they remain agnostic to the underlying driving mechanisms or\nsources of unfairness. Even within the supervised ML literature, there exists\ndebate on whether unfairness stems solely from algorithmic biases (i.e. design\nchoices) or from the biases encoded in the data on which they are trained. To\nclose this gap, this work aims to shed light on the possible sources of\nunfairness in OD by auditing detection models under different data-centric\nfactors. By injecting various known biases into the input data -- as pertain to\nsample size disparity, under-representation, feature measurement noise, and\ngroup membership obfuscation -- we find that the OD algorithms under the study\nall exhibit fairness pitfalls, although differing in which types of data bias\nthey are more susceptible to. Most notable of our study is to demonstrate that\nOD algorithm bias is not merely a data bias problem. A key realization is that\nthe data properties that emerge from bias injection could as well be organic --\nas pertain to natural group differences w.r.t. sparsity, base rate, variance,\nand multi-modality. Either natural or biased, such data properties can give\nrise to unfairness as they interact with certain algorithmic design choices.\n", "date": "2024-08-24", "categories": []}
{"id": "2408.14842", "title": "From Bias to Balance: Detecting Facial Expression Recognition Biases in\n  Large Multimodal Foundation Models", "abstract": "  This study addresses the racial biases in facial expression recognition (FER)\nsystems within Large Multimodal Foundation Models (LMFMs). Despite advances in\ndeep learning and the availability of diverse datasets, FER systems often\nexhibit higher error rates for individuals with darker skin tones. Existing\nresearch predominantly focuses on traditional FER models (CNNs, RNNs, ViTs),\nleaving a gap in understanding racial biases in LMFMs. We benchmark four\nleading LMFMs: GPT-4o, PaliGemma, Gemini, and CLIP to assess their performance\nin facial emotion detection across different racial demographics. A linear\nclassifier trained on CLIP embeddings obtains accuracies of 95.9\\% for RADIATE,\n90.3\\% for Tarr, and 99.5\\% for Chicago Face. Furthermore, we identify that\nAnger is misclassified as Disgust 2.1 times more often in Black Females than\nWhite Females. This study highlights the need for fairer FER systems and\nestablishes a foundation for developing unbiased, accurate FER technologies.\nVisit https://kvjvhub.github.io/FERRacialBias/ for further information\nregarding the biases within facial expression recognition.\n", "date": "2024-08-27", "categories": []}
{"id": "2307.11892", "title": "On the Vulnerability of Fairness Constrained Learning to Malicious Noise", "abstract": "  We consider the vulnerability of fairness-constrained learning to small\namounts of malicious noise in the training data. Konstantinov and Lampert\n(2021) initiated the study of this question and presented negative results\nshowing there exist data distributions where for several fairness constraints,\nany proper learner will exhibit high vulnerability when group sizes are\nimbalanced. Here, we present a more optimistic view, showing that if we allow\nrandomized classifiers, then the landscape is much more nuanced. For example,\nfor Demographic Parity we show we can incur only a $\\Theta(\\alpha)$ loss in\naccuracy, where $\\alpha$ is the malicious noise rate, matching the best\npossible even without fairness constraints. For Equal Opportunity, we show we\ncan incur an $O(\\sqrt{\\alpha})$ loss, and give a matching\n$\\Omega(\\sqrt{\\alpha})$lower bound. In contrast, Konstantinov and Lampert\n(2021) showed for proper learners the loss in accuracy for both notions is\n$\\Omega(1)$. The key technical novelty of our work is how randomization can\nbypass simple \"tricks\" an adversary can use to amplify his power. We also\nconsider additional fairness notions including Equalized Odds and Calibration.\nFor these fairness notions, the excess accuracy clusters into three natural\nregimes $O(\\alpha)$,$O(\\sqrt{\\alpha})$ and $O(1)$. These results provide a more\nfine-grained view of the sensitivity of fairness-constrained learning to\nadversarial noise in training data.\n", "date": "2023-07-21", "categories": []}
{"id": "2408.12665", "title": "Fairness-Aware Streaming Feature Selection with Causal Graphs", "abstract": "  Its crux lies in the optimization of a tradeoff between accuracy and fairness\nof resultant models on the selected feature subset. The technical challenge of\nour setting is twofold: 1) streaming feature inputs, such that an informative\nfeature may become obsolete or redundant for prediction if its information has\nbeen covered by other similar features that arrived prior to it, and 2)\nnon-associational feature correlation, such that bias may be leaked from those\nseemingly admissible, non-protected features. To overcome this, we propose\nStreaming Feature Selection with Causal Fairness (SFCF) that builds two causal\ngraphs egocentric to prediction label and protected feature, respectively,\nstriving to model the complex correlation structure among streaming features,\nlabels, and protected information. As such, bias can be eradicated from\npredictive modeling by removing those features being causally correlated with\nthe protected feature yet independent to the labels. We theorize that the\noriginally redundant features for prediction can later become admissible, when\nthe learning accuracy is compromised by the large number of removed features\n(non-protected but can be used to reconstruct bias information). We benchmark\nSFCF\\ on five datasets widely used in streaming feature research, and the\nresults substantiate its performance superiority over six rival models in terms\nof efficiency and sparsity of feature selection and equalized odds of the\nresultant predictive models.\n", "date": "2024-08-17", "categories": []}
{"id": "2408.13208", "title": "Temporal Fairness in Decision Making Problems", "abstract": "  In this work we consider a new interpretation of fairness in decision making\nproblems. Building upon existing fairness formulations, we focus on how to\nreason over fairness from a temporal perspective, taking into account the\nfairness of a history of past decisions. After introducing the concept of\ntemporal fairness, we propose three approaches that incorporate temporal\nfairness in decision making problems formulated as optimization problems. We\npresent a qualitative evaluation of our approach in four different domains and\ncompare the solutions against a baseline approach that does not consider the\ntemporal aspect of fairness.\n", "date": "2024-08-23", "categories": []}
{"id": "2408.12055", "title": "Aligning (Medical) LLMs for (Counterfactual) Fairness", "abstract": "  Large Language Models (LLMs) have emerged as promising solutions for a\nvariety of medical and clinical decision support applications. However, LLMs\nare often subject to different types of biases, which can lead to unfair\ntreatment of individuals, worsening health disparities, and reducing trust in\nAI-augmented medical tools. Aiming to address this important issue, in this\nstudy, we present a new model alignment approach for aligning LLMs using a\npreference optimization method within a knowledge distillation framework. Prior\nto presenting our proposed method, we first use an evaluation framework to\nconduct a comprehensive (largest to our knowledge) empirical evaluation to\nreveal the type and nature of existing biases in LLMs used for medical\napplications. We then offer a bias mitigation technique to reduce the unfair\npatterns in LLM outputs across different subgroups identified by the protected\nattributes. We show that our mitigation method is effective in significantly\nreducing observed biased patterns. Our code is publicly available at\n\\url{https://github.com/healthylaife/FairAlignmentLLM}.\n", "date": "2024-08-22", "categories": []}
{"id": "2404.11457", "title": "Bias and Unfairness in Information Retrieval Systems: New Challenges in\n  the LLM Era", "abstract": "  With the rapid advancements of large language models (LLMs), information\nretrieval (IR) systems, such as search engines and recommender systems, have\nundergone a significant paradigm shift. This evolution, while heralding new\nopportunities, introduces emerging challenges, particularly in terms of biases\nand unfairness, which may threaten the information ecosystem. In this paper, we\npresent a comprehensive survey of existing works on emerging and pressing bias\nand unfairness issues in IR systems when the integration of LLMs. We first\nunify bias and unfairness issues as distribution mismatch problems, providing a\ngroundwork for categorizing various mitigation strategies through distribution\nalignment. Subsequently, we systematically delve into the specific bias and\nunfairness issues arising from three critical stages of LLMs integration into\nIR systems: data collection, model development, and result evaluation. In doing\nso, we meticulously review and analyze recent literature, focusing on the\ndefinitions, characteristics, and corresponding mitigation strategies\nassociated with these issues. Finally, we identify and highlight some open\nproblems and challenges for future work, aiming to inspire researchers and\nstakeholders in the IR field and beyond to better understand and mitigate bias\nand unfairness issues of IR in this LLM era. We also consistently maintain a\nGitHub repository for the relevant papers and resources in this rising\ndirection at https://github.com/KID-22/LLM-IR-Bias-Fairness-Survey.\n", "date": "2024-04-17", "categories": []}
{"id": "2407.20371", "title": "Gender, Race, and Intersectional Bias in Resume Screening via Language\n  Model Retrieval", "abstract": "  Artificial intelligence (AI) hiring tools have revolutionized resume\nscreening, and large language models (LLMs) have the potential to do the same.\nHowever, given the biases which are embedded within LLMs, it is unclear whether\nthey can be used in this scenario without disadvantaging groups based on their\nprotected attributes. In this work, we investigate the possibilities of using\nLLMs in a resume screening setting via a document retrieval framework that\nsimulates job candidate selection. Using that framework, we then perform a\nresume audit study to determine whether a selection of Massive Text Embedding\n(MTE) models are biased in resume screening scenarios. We simulate this for\nnine occupations, using a collection of over 500 publicly available resumes and\n500 job descriptions. We find that the MTEs are biased, significantly favoring\nWhite-associated names in 85.1\\% of cases and female-associated names in only\n11.1\\% of cases, with a minority of cases showing no statistically significant\ndifferences. Further analyses show that Black males are disadvantaged in up to\n100\\% of cases, replicating real-world patterns of bias in employment settings,\nand validate three hypotheses of intersectionality. We also find an impact of\ndocument length as well as the corpus frequency of names in the selection of\nresumes. These findings have implications for widely used AI tools that are\nautomating employment, fairness, and tech policy.\n", "date": "2024-07-29", "categories": []}
{"id": "2408.11439", "title": "BAdd: Bias Mitigation through Bias Addition", "abstract": "  Computer vision (CV) datasets often exhibit biases that are perpetuated by\ndeep learning models. While recent efforts aim to mitigate these biases and\nfoster fair representations, they fail in complex real-world scenarios. In\nparticular, existing methods excel in controlled experiments involving\nbenchmarks with single-attribute injected biases, but struggle with\nmulti-attribute biases being present in well-established CV datasets. Here, we\nintroduce BAdd, a simple yet effective method that allows for learning fair\nrepresentations invariant to the attributes introducing bias by incorporating\nfeatures representing these attributes into the backbone. BAdd is evaluated on\nseven benchmarks and exhibits competitive performance, surpassing\nstate-of-the-art methods on both single- and multi-attribute benchmarks.\nNotably, BAdd achieves +27.5% and +5.5% absolute accuracy improvements on the\nchallenging multi-attribute benchmarks, FB-Biased-MNIST and CelebA,\nrespectively.\n", "date": "2024-08-21", "categories": []}
{"id": "2408.11448", "title": "Lookism: The overlooked bias in computer vision", "abstract": "  In recent years, there have been significant advancements in computer vision\nwhich have led to the widespread deployment of image recognition and generation\nsystems in socially relevant applications, from hiring to security screening.\nHowever, the prevalence of biases within these systems has raised significant\nethical and social concerns. The most extensively studied biases in this\ncontext are related to gender, race and age. Yet, other biases are equally\npervasive and harmful, such as lookism, i.e., the preferential treatment of\nindividuals based on their physical appearance. Lookism remains under-explored\nin computer vision but can have profound implications not only by perpetuating\nharmful societal stereotypes but also by undermining the fairness and\ninclusivity of AI technologies. Thus, this paper advocates for the systematic\nstudy of lookism as a critical bias in computer vision models. Through a\ncomprehensive review of existing literature, we identify three areas of\nintersection between lookism and computer vision. We illustrate them by means\nof examples and a user study. We call for an interdisciplinary approach to\naddress lookism, urging researchers, developers, and policymakers to prioritize\nthe development of equitable computer vision systems that respect and reflect\nthe diversity of human appearances.\n", "date": "2024-08-21", "categories": []}
{"id": "2401.00763", "title": "New Job, New Gender? Measuring the Social Bias in Image Generation\n  Models", "abstract": "  Image generation models can generate or edit images from a given text. Recent\nadvancements in image generation technology, exemplified by DALL-E and\nMidjourney, have been groundbreaking. These advanced models, despite their\nimpressive capabilities, are often trained on massive Internet datasets, making\nthem susceptible to generating content that perpetuates social stereotypes and\nbiases, which can lead to severe consequences. Prior research on assessing bias\nwithin image generation models suffers from several shortcomings, including\nlimited accuracy, reliance on extensive human labor, and lack of comprehensive\nanalysis. In this paper, we propose BiasPainter, a novel evaluation framework\nthat can accurately, automatically and comprehensively trigger social bias in\nimage generation models. BiasPainter uses a diverse range of seed images of\nindividuals and prompts the image generation models to edit these images using\ngender, race, and age-neutral queries. These queries span 62 professions, 39\nactivities, 57 types of objects, and 70 personality traits. The framework then\ncompares the edited images to the original seed images, focusing on the\nsignificant changes related to gender, race, and age. BiasPainter adopts a key\ninsight that these characteristics should not be modified when subjected to\nneutral prompts. Built upon this design, BiasPainter can trigger the social\nbias and evaluate the fairness of image generation models. We use BiasPainter\nto evaluate six widely-used image generation models, such as stable diffusion\nand Midjourney. Experimental results show that BiasPainter can successfully\ntrigger social bias in image generation models. According to our human\nevaluation, BiasPainter can achieve 90.8% accuracy on automatic bias detection,\nwhich is significantly higher than the results reported in previous work.\n", "date": "2024-01-01", "categories": []}
{"id": "2408.10608", "title": "Promoting Equality in Large Language Models: Identifying and Mitigating\n  the Implicit Bias based on Bayesian Theory", "abstract": "  Large language models (LLMs) are trained on extensive text corpora, which\ninevitably include biased information. Although techniques such as Affective\nAlignment can mitigate some negative impacts of these biases, existing\nprompt-based attack methods can still extract these biases from the model's\nweights. Moreover, these biases frequently appear subtly when LLMs are prompted\nto perform identical tasks across different demographic groups, thereby\ncamouflaging their presence. To address this issue, we have formally defined\nthe implicit bias problem and developed an innovative framework for bias\nremoval based on Bayesian theory, Bayesian-Theory based Bias Removal (BTBR).\nBTBR employs likelihood ratio screening to pinpoint data entries within\npublicly accessible biased datasets that represent biases inadvertently\nincorporated during the LLM training phase. It then automatically constructs\nrelevant knowledge triples and expunges bias information from LLMs using model\nediting techniques. Through extensive experimentation, we have confirmed the\npresence of the implicit bias problem in LLMs and demonstrated the\neffectiveness of our BTBR approach.\n", "date": "2024-08-20", "categories": []}
{"id": "2402.08522", "title": "Fairness Auditing with Multi-Agent Collaboration", "abstract": "  Existing work in fairness auditing assumes that each audit is performed\nindependently. In this paper, we consider multiple agents working together,\neach auditing the same platform for different tasks. Agents have two levers:\ntheir collaboration strategy, with or without coordination beforehand, and\ntheir strategy for sampling appropriate data points. We theoretically compare\nthe interplay of these levers. Our main findings are that (i) collaboration is\ngenerally beneficial for accurate audits, (ii) basic sampling methods often\nprove to be effective, and (iii) counter-intuitively, extensive coordination on\nqueries often deteriorates audits accuracy as the number of agents increases.\nExperiments on three large datasets confirm our theoretical results. Our\nfindings motivate collaboration during fairness audits of platforms that use ML\nmodels for decision-making.\n", "date": "2024-02-13", "categories": []}
{"id": "2408.09312", "title": "Learning Fair Invariant Representations under Covariate and Correlation\n  Shifts Simultaneously", "abstract": "  Achieving the generalization of an invariant classifier from training domains\nto shifted test domains while simultaneously considering model fairness is a\nsubstantial and complex challenge in machine learning. Existing methods address\nthe problem of fairness-aware domain generalization, focusing on either\ncovariate shift or correlation shift, but rarely consider both at the same\ntime. In this paper, we introduce a novel approach that focuses on learning a\nfairness-aware domain-invariant predictor within a framework addressing both\ncovariate and correlation shifts simultaneously, ensuring its generalization to\nunknown test domains inaccessible during training. In our approach, data are\nfirst disentangled into content and style factors in latent spaces.\nFurthermore, fairness-aware domain-invariant content representations can be\nlearned by mitigating sensitive information and retaining as much other\ninformation as possible. Extensive empirical studies on benchmark datasets\ndemonstrate that our approach surpasses state-of-the-art methods with respect\nto model accuracy as well as both group and individual fairness.\n", "date": "2024-08-18", "categories": []}
{"id": "2408.09757", "title": "Strategic Demonstration Selection for Improved Fairness in LLM\n  In-Context Learning", "abstract": "  Recent studies highlight the effectiveness of using in-context learning (ICL)\nto steer large language models (LLMs) in processing tabular data, a challenging\ntask given the structured nature of such data. Despite advancements in\nperformance, the fairness implications of these methods are less understood.\nThis study investigates how varying demonstrations within ICL prompts influence\nthe fairness outcomes of LLMs. Our findings reveal that deliberately including\nminority group samples in prompts significantly boosts fairness without\nsacrificing predictive accuracy. Further experiments demonstrate that the\nproportion of minority to majority samples in demonstrations affects the\ntrade-off between fairness and prediction accuracy. Based on these insights, we\nintroduce a mitigation technique that employs clustering and evolutionary\nstrategies to curate a diverse and representative sample set from the training\ndata. This approach aims to enhance both predictive performance and fairness in\nICL applications. Experimental results validate that our proposed method\ndramatically improves fairness across various metrics, showing its efficacy in\nreal-world scenarios.\n", "date": "2024-08-19", "categories": []}
{"id": "2408.10002", "title": "The Fairness-Quality Trade-off in Clustering", "abstract": "  Fairness in clustering has been considered extensively in the past; however,\nthe trade-off between the two objectives -- e.g., can we sacrifice just a\nlittle in the quality of the clustering to significantly increase fairness, or\nvice-versa? -- has rarely been addressed. We introduce novel algorithms for\ntracing the complete trade-off curve, or Pareto front, between quality and\nfairness in clustering problems; that is, computing all clusterings that are\nnot dominated in both objectives by other clusterings. Unlike previous work\nthat deals with specific objectives for quality and fairness, we deal with all\nobjectives for fairness and quality in two general classes encompassing most of\nthe special cases addressed in previous work. Our algorithm must take\nexponential time in the worst case as the Pareto front itself can be\nexponential. Even when the Pareto front is polynomial, our algorithm may take\nexponential time, and we prove that this is inevitable unless P = NP. However,\nwe also present a new polynomial-time algorithm for computing the entire Pareto\nfront when the cluster centers are fixed, and for perhaps the most natural\nfairness objective: minimizing the sum, over all clusters, of the imbalance\nbetween the two groups in each cluster.\n", "date": "2024-08-19", "categories": []}
{"id": "2408.10175", "title": "Fairness Under Cover: Evaluating the Impact of Occlusions on Demographic\n  Bias in Facial Recognition", "abstract": "  This study investigates the effects of occlusions on the fairness of face\nrecognition systems, particularly focusing on demographic biases. Using the\nRacial Faces in the Wild (RFW) dataset and synthetically added realistic\nocclusions, we evaluate their effect on the performance of face recognition\nmodels trained on the BUPT-Balanced and BUPT-GlobalFace datasets. We note\nincreases in the dispersion of FMR, FNMR, and accuracy alongside decreases in\nfairness according to Equilized Odds, Demographic Parity, STD of Accuracy, and\nFairness Discrepancy Rate. Additionally, we utilize a pixel attribution method\nto understand the importance of occlusions in model predictions, proposing a\nnew metric, Face Occlusion Impact Ratio (FOIR), that quantifies the extent to\nwhich occlusions affect model performance across different demographic groups.\nOur results indicate that occlusions exacerbate existing demographic biases,\nwith models placing higher importance on occlusions in an unequal fashion,\nparticularly affecting African individuals more severely.\n", "date": "2024-08-19", "categories": []}
{"id": "2408.08214", "title": "Federated Fairness Analytics: Quantifying Fairness in Federated Learning", "abstract": "  Federated Learning (FL) is a privacy-enhancing technology for distributed ML.\nBy training models locally and aggregating updates - a federation learns\ntogether, while bypassing centralised data collection. FL is increasingly\npopular in healthcare, finance and personal computing. However, it inherits\nfairness challenges from classical ML and introduces new ones, resulting from\ndifferences in data quality, client participation, communication constraints,\naggregation methods and underlying hardware. Fairness remains an unresolved\nissue in FL and the community has identified an absence of succinct definitions\nand metrics to quantify fairness; to address this, we propose Federated\nFairness Analytics - a methodology for measuring fairness. Our definition of\nfairness comprises four notions with novel, corresponding metrics. They are\nsymptomatically defined and leverage techniques originating from XAI,\ncooperative game-theory and networking engineering. We tested a range of\nexperimental settings, varying the FL approach, ML task and data settings. The\nresults show that statistical heterogeneity and client participation affect\nfairness and fairness conscious approaches such as Ditto and q-FedAvg\nmarginally improve fairness-performance trade-offs. Using our techniques, FL\npractitioners can uncover previously unobtainable insights into their system's\nfairness, at differing levels of granularity in order to address fairness\nchallenges in FL. We have open-sourced our work at:\nhttps://github.com/oscardilley/federated-fairness.\n", "date": "2024-08-15", "categories": []}
{"id": "2306.08158", "title": "Sociodemographic Bias in Language Models: A Survey and Forward Path", "abstract": "  Sociodemographic bias in language models (LMs) has the potential for harm\nwhen deployed in real-world settings. This paper presents a comprehensive\nsurvey of the past decade of research on sociodemographic bias in LMs,\norganized into a typology that facilitates examining the different aims: types\nof bias, quantifying bias, and debiasing techniques. We track the evolution of\nthe latter two questions, then identify current trends and their limitations,\nas well as emerging techniques. To guide future research towards more effective\nand reliable solutions, and to help authors situate their work within this\nbroad landscape, we conclude with a checklist of open questions.\n", "date": "2023-06-13", "categories": []}
{"id": "2405.17034", "title": "FUGNN: Harmonizing Fairness and Utility in Graph Neural Networks", "abstract": "  Fairness-aware Graph Neural Networks (GNNs) often face a challenging\ntrade-off, where prioritizing fairness may require compromising utility. In\nthis work, we re-examine fairness through the lens of spectral graph theory,\naiming to reconcile fairness and utility within the framework of spectral graph\nlearning. We explore the correlation between sensitive features and spectrum in\nGNNs, using theoretical analysis to delineate the similarity between original\nsensitive features and those after convolution under different spectra. Our\nanalysis reveals a reduction in the impact of similarity when the eigenvectors\nassociated with the largest magnitude eigenvalue exhibit directional\nsimilarity. Based on these theoretical insights, we propose FUGNN, a novel\nspectral graph learning approach that harmonizes the conflict between fairness\nand utility. FUGNN ensures algorithmic fairness and utility by truncating the\nspectrum and optimizing eigenvector distribution during the encoding process.\nThe fairness-aware eigenvector selection reduces the impact of convolution on\nsensitive features while concurrently minimizing the sacrifice of utility.\nFUGNN further optimizes the distribution of eigenvectors through a transformer\narchitecture. By incorporating the optimized spectrum into the graph\nconvolution network, FUGNN effectively learns node representations. Experiments\non six real-world datasets demonstrate the superiority of FUGNN over baseline\nmethods. The codes are available at https://github.com/yushuowiki/FUGNN.\n", "date": "2024-05-27", "categories": []}
{"id": "2306.04489", "title": "Fair Column Subset Selection", "abstract": "  The problem of column subset selection asks for a subset of columns from an\ninput matrix such that the matrix can be reconstructed as accurately as\npossible within the span of the selected columns. A natural extension is to\nconsider a setting where the matrix rows are partitioned into two groups, and\nthe goal is to choose a subset of columns that minimizes the maximum\nreconstruction error of both groups, relative to their respective best rank-k\napproximation. Extending the known results of column subset selection to this\nfair setting is not straightforward: in certain scenarios it is unavoidable to\nchoose columns separately for each group, resulting in double the expected\ncolumn count. We propose a deterministic leverage-score sampling strategy for\nthe fair setting and show that sampling a column subset of minimum size becomes\nNP-hard in the presence of two groups. Despite these negative results, we give\nan approximation algorithm that guarantees a solution within 1.5 times the\noptimal solution size. We also present practical heuristic algorithms based on\nrank-revealing QR factorization. Finally, we validate our methods through an\nextensive set of experiments using real-world data.\n", "date": "2023-06-07", "categories": []}
{"id": "2312.03027", "title": "Stable Diffusion Exposed: Gender Bias from Prompt to Image", "abstract": "  Several studies have raised awareness about social biases in image generative\nmodels, demonstrating their predisposition towards stereotypes and imbalances.\nThis paper contributes to this growing body of research by introducing an\nevaluation protocol that analyzes the impact of gender indicators at every step\nof the generation process on Stable Diffusion images. Leveraging insights from\nprior work, we explore how gender indicators not only affect gender\npresentation but also the representation of objects and layouts within the\ngenerated images. Our findings include the existence of differences in the\ndepiction of objects, such as instruments tailored for specific genders, and\nshifts in overall layouts. We also reveal that neutral prompts tend to produce\nimages more aligned with masculine prompts than their feminine counterparts. We\nfurther explore where bias originates through representational disparities and\nhow it manifests in the images via prompt-image dependencies, and provide\nrecommendations for developers and users to mitigate potential bias in image\ngeneration.\n", "date": "2023-12-05", "categories": []}
{"id": "2408.06102", "title": "Contexts Matter: An Empirical Study on Contextual Influence in Fairness\n  Testing for Deep Learning Systems", "abstract": "  Background: Fairness testing for deep learning systems has been becoming\nincreasingly important. However, much work assumes perfect context and\nconditions from the other parts: well-tuned hyperparameters for accuracy;\nrectified bias in data, and mitigated bias in the labeling. Yet, these are\noften difficult to achieve in practice due to their resource-/labour-intensive\nnature. Aims: In this paper, we aim to understand how varying contexts affect\nfairness testing outcomes. Method:We conduct an extensive empirical study,\nwhich covers $10,800$ cases, to investigate how contexts can change the\nfairness testing result at the model level against the existing assumptions. We\nalso study why the outcomes were observed from the lens of correlation/fitness\nlandscape analysis. Results: Our results show that different context types and\nsettings generally lead to a significant impact on the testing, which is mainly\ncaused by the shifts of the fitness landscape under varying contexts.\nConclusions: Our findings provide key insights for practitioners to evaluate\nthe test generators and hint at future research directions.\n", "date": "2024-08-12", "categories": []}
{"id": "2408.04671", "title": "Prompt and Prejudice", "abstract": "  This paper investigates the impact of using first names in Large Language\nModels (LLMs) and Vision Language Models (VLMs), particularly when prompted\nwith ethical decision-making tasks. We propose an approach that appends first\nnames to ethically annotated text scenarios to reveal demographic biases in\nmodel outputs. Our study involves a curated list of more than 300 names\nrepresenting diverse genders and ethnic backgrounds, tested across thousands of\nmoral scenarios. Following the auditing methodologies from social sciences we\npropose a detailed analysis involving popular LLMs/VLMs to contribute to the\nfield of responsible AI by emphasizing the importance of recognizing and\nmitigating biases in these systems. Furthermore, we introduce a novel\nbenchmark, the Pratical Scenarios Benchmark (PSB), designed to assess the\npresence of biases involving gender or demographic prejudices in everyday\ndecision-making scenarios as well as practical scenarios where an LLM might be\nused to make sensible decisions (e.g., granting mortgages or insurances). This\nbenchmark allows for a comprehensive comparison of model behaviors across\ndifferent demographic categories, highlighting the risks and biases that may\narise in practical applications of LLMs and VLMs.\n", "date": "2024-08-07", "categories": []}
{"id": "2408.04955", "title": "Model Debiasing by Learnable Data Augmentation", "abstract": "  Deep Neural Networks are well known for efficiently fitting training data,\nyet experiencing poor generalization capabilities whenever some kind of bias\ndominates over the actual task labels, resulting in models learning\n\"shortcuts\". In essence, such models are often prone to learn spurious\ncorrelations between data and labels. In this work, we tackle the problem of\nlearning from biased data in the very realistic unsupervised scenario, i.e.,\nwhen the bias is unknown. This is a much harder task as compared to the\nsupervised case, where auxiliary, bias-related annotations, can be exploited in\nthe learning process. This paper proposes a novel 2-stage learning pipeline\nfeaturing a data augmentation strategy able to regularize the training. First,\nbiased/unbiased samples are identified by training over-biased models. Second,\nsuch subdivision (typically noisy) is exploited within a data augmentation\nframework, properly combining the original samples while learning mixing\nparameters, which has a regularization effect. Experiments on synthetic and\nrealistic biased datasets show state-of-the-art classification accuracy,\noutperforming competing methods, ultimately proving robust performance on both\nbiased and unbiased examples. Notably, being our training method totally\nagnostic to the level of bias, it also positively affects performance for any,\neven apparently unbiased, dataset, thus improving the model generalization\nregardless of the level of bias (or its absence) in the data.\n", "date": "2024-08-09", "categories": []}
{"id": "2408.00992", "title": "Fairness in Large Language Models in Three Hours", "abstract": "  Large Language Models (LLMs) have demonstrated remarkable success across\nvarious domains but often lack fairness considerations, potentially leading to\ndiscriminatory outcomes against marginalized populations. Unlike fairness in\ntraditional machine learning, fairness in LLMs involves unique backgrounds,\ntaxonomies, and fulfillment techniques. This tutorial provides a systematic\noverview of recent advances in the literature concerning fair LLMs, beginning\nwith real-world case studies to introduce LLMs, followed by an analysis of bias\ncauses therein. The concept of fairness in LLMs is then explored, summarizing\nthe strategies for evaluating bias and the algorithms designed to promote\nfairness. Additionally, resources for assessing bias in LLMs, including\ntoolkits and datasets, are compiled, and current research challenges and open\nquestions in the field are discussed. The repository is available at\n\\url{https://github.com/LavinWong/Fairness-in-Large-Language-Models}.\n", "date": "2024-08-02", "categories": []}
{"id": "2408.04026", "title": "Multimodal Gender Fairness in Depression Prediction: Insights on Data\n  from the USA & China", "abstract": "  Social agents and robots are increasingly being used in wellbeing settings.\nHowever, a key challenge is that these agents and robots typically rely on\nmachine learning (ML) algorithms to detect and analyse an individual's mental\nwellbeing. The problem of bias and fairness in ML algorithms is becoming an\nincreasingly greater source of concern. In concurrence, existing literature has\nalso indicated that mental health conditions can manifest differently across\ngenders and cultures. We hypothesise that the representation of features\n(acoustic, textual, and visual) and their inter-modal relations would vary\namong subjects from different cultures and genders, thus impacting the\nperformance and fairness of various ML models. We present the very first\nevaluation of multimodal gender fairness in depression manifestation by\nundertaking a study on two different datasets from the USA and China. We\nundertake thorough statistical and ML experimentation and repeat the\nexperiments for several different algorithms to ensure that the results are not\nalgorithm-dependent. Our findings indicate that though there are differences\nbetween both datasets, it is not conclusive whether this is due to the\ndifference in depression manifestation as hypothesised or other external\nfactors such as differences in data collection methodology. Our findings\nfurther motivate a call for a more consistent and culturally aware data\ncollection process in order to address the problem of ML bias in depression\ndetection and to promote the development of fairer agents and robots for\nwellbeing.\n", "date": "2024-08-07", "categories": []}
{"id": "2302.04262", "title": "Algorithmic Collective Action in Machine Learning", "abstract": "  We initiate a principled study of algorithmic collective action on digital\nplatforms that deploy machine learning algorithms. We propose a simple\ntheoretical model of a collective interacting with a firm's learning algorithm.\nThe collective pools the data of participating individuals and executes an\nalgorithmic strategy by instructing participants how to modify their own data\nto achieve a collective goal. We investigate the consequences of this model in\nthree fundamental learning-theoretic settings: the case of a nonparametric\noptimal learning algorithm, a parametric risk minimizer, and gradient-based\noptimization. In each setting, we come up with coordinated algorithmic\nstrategies and characterize natural success criteria as a function of the\ncollective's size. Complementing our theory, we conduct systematic experiments\non a skill classification task involving tens of thousands of resumes from a\ngig platform for freelancers. Through more than two thousand model training\nruns of a BERT-like language model, we see a striking correspondence emerge\nbetween our empirical observations and the predictions made by our theory.\nTaken together, our theory and experiments broadly support the conclusion that\nalgorithmic collectives of exceedingly small fractional size can exert\nsignificant control over a platform's learning algorithm.\n", "date": "2023-02-08", "categories": []}
{"id": "2407.02702", "title": "Practical Guide for Causal Pathways and Sub-group Disparity Analysis", "abstract": "  In this study, we introduce the application of causal disparity analysis to\nunveil intricate relationships and causal pathways between sensitive attributes\nand the targeted outcomes within real-world observational data. Our methodology\ninvolves employing causal decomposition analysis to quantify and examine the\ncausal interplay between sensitive attributes and outcomes. We also emphasize\nthe significance of integrating heterogeneity assessment in causal disparity\nanalysis to gain deeper insights into the impact of sensitive attributes within\nspecific sub-groups on outcomes. Our two-step investigation focuses on datasets\nwhere race serves as the sensitive attribute. The results on two datasets\nindicate the benefit of leveraging causal analysis and heterogeneity assessment\nnot only for quantifying biases in the data but also for disentangling their\ninfluences on outcomes. We demonstrate that the sub-groups identified by our\napproach to be affected the most by disparities are the ones with the largest\nML classification errors. We also show that grouping the data only based on a\nsensitive attribute is not enough, and through these analyses, we can find\nsub-groups that are directly affected by disparities. We hope that our findings\nwill encourage the adoption of such methodologies in future ethical AI\npractices and bias audits, fostering a more equitable and fair technological\nlandscape.\n", "date": "2024-07-02", "categories": []}
{"id": "2305.13088", "title": "Should We Attend More or Less? Modulating Attention for Fairness", "abstract": "  The advances in natural language processing (NLP) pose both opportunities and\nchallenges. While recent progress enables the development of high-performing\nmodels for a variety of tasks, it also poses the risk of models learning\nharmful biases from the data, such as gender stereotypes. In this work, we\ninvestigate the role of attention, a widely-used technique in current\nstate-of-the-art NLP models, in the propagation of social biases. Specifically,\nwe study the relationship between the entropy of the attention distribution and\nthe model's performance and fairness. We then propose a novel method for\nmodulating attention weights to improve model fairness after training. Since\nour method is only applied post-training and pre-inference, it is an\nintra-processing method and is, therefore, less computationally expensive than\nexisting in-processing and pre-processing approaches. Our results show an\nincrease in fairness and minimal performance loss on different text\nclassification and generation tasks using language models of varying sizes.\nWARNING: This work uses language that is offensive.\n", "date": "2023-05-22", "categories": []}
{"id": "2310.19981", "title": "'Person' == Light-skinned, Western Man, and Sexualization of Women of\n  Color: Stereotypes in Stable Diffusion", "abstract": "  We study stereotypes embedded within one of the most popular text-to-image\ngenerators: Stable Diffusion. We examine what stereotypes of gender and\nnationality/continental identity does Stable Diffusion display in the absence\nof such information i.e. what gender and nationality/continental identity is\nassigned to `a person', or to `a person from Asia'. Using vision-language model\nCLIP's cosine similarity to compare images generated by CLIP-based Stable\nDiffusion v2.1 verified by manual examination, we chronicle results from 136\nprompts (50 results/prompt) of front-facing images of persons from 6 different\ncontinents, 27 nationalities and 3 genders. We observe how Stable Diffusion\noutputs of `a person' without any additional gender/nationality information\ncorrespond closest to images of men and least with persons of nonbinary gender,\nand to persons from Europe/North America over Africa/Asia, pointing towards\nStable Diffusion having a concerning representation of personhood to be a\nEuropean/North American man. We also show continental stereotypes and resultant\nharms e.g. a person from Oceania is deemed to be Australian/New Zealander over\nPapua New Guinean, pointing to the erasure of Indigenous Oceanic peoples, who\nform a majority over descendants of colonizers both in Papua New Guinea and in\nOceania overall. Finally, we unexpectedly observe a pattern of\noversexualization of women, specifically Latin American, Mexican, Indian and\nEgyptian women relative to other nationalities, measured through an NSFW\ndetector. This demonstrates how Stable Diffusion perpetuates Western\nfetishization of women of color through objectification in media, which if left\nunchecked will amplify this stereotypical representation. Image datasets are\nmade publicly available.\n", "date": "2023-10-30", "categories": []}
{"id": "2402.00955", "title": "FairEHR-CLP: Towards Fairness-Aware Clinical Predictions with\n  Contrastive Learning in Multimodal Electronic Health Records", "abstract": "  In the high-stakes realm of healthcare, ensuring fairness in predictive\nmodels is crucial. Electronic Health Records (EHRs) have become integral to\nmedical decision-making, yet existing methods for enhancing model fairness\nrestrict themselves to unimodal data and fail to address the multifaceted\nsocial biases intertwined with demographic factors in EHRs. To mitigate these\nbiases, we present FairEHR-CLP: a general framework for Fairness-aware Clinical\nPredictions with Contrastive Learning in EHRs. FairEHR-CLP operates through a\ntwo-stage process, utilizing patient demographics, longitudinal data, and\nclinical notes. First, synthetic counterparts are generated for each patient,\nallowing for diverse demographic identities while preserving essential health\ninformation. Second, fairness-aware predictions employ contrastive learning to\nalign patient representations across sensitive attributes, jointly optimized\nwith an MLP classifier with a softmax layer for clinical classification tasks.\nAcknowledging the unique challenges in EHRs, such as varying group sizes and\nclass imbalance, we introduce a novel fairness metric to effectively measure\nerror rate disparities across subgroups. Extensive experiments on three diverse\nEHR datasets on three tasks demonstrate the effectiveness of FairEHR-CLP in\nterms of fairness and utility compared with competitive baselines. FairEHR-CLP\nrepresents an advancement towards ensuring both accuracy and equity in\npredictive healthcare models.\n", "date": "2024-02-01", "categories": []}
{"id": "2404.07990", "title": "OpenBias: Open-set Bias Detection in Text-to-Image Generative Models", "abstract": "  Text-to-image generative models are becoming increasingly popular and\naccessible to the general public. As these models see large-scale deployments,\nit is necessary to deeply investigate their safety and fairness to not\ndisseminate and perpetuate any kind of biases. However, existing works focus on\ndetecting closed sets of biases defined a priori, limiting the studies to\nwell-known concepts. In this paper, we tackle the challenge of open-set bias\ndetection in text-to-image generative models presenting OpenBias, a new\npipeline that identifies and quantifies the severity of biases agnostically,\nwithout access to any precompiled set. OpenBias has three stages. In the first\nphase, we leverage a Large Language Model (LLM) to propose biases given a set\nof captions. Secondly, the target generative model produces images using the\nsame set of captions. Lastly, a Vision Question Answering model recognizes the\npresence and extent of the previously proposed biases. We study the behavior of\nStable Diffusion 1.5, 2, and XL emphasizing new biases, never investigated\nbefore. Via quantitative experiments, we demonstrate that OpenBias agrees with\ncurrent closed-set bias detection methods and human judgement.\n", "date": "2024-04-11", "categories": []}
{"id": "2408.01630", "title": "Fair Risk Minimization under Causal Path-Specific Effect Constraints", "abstract": "  This paper introduces a framework for estimating fair optimal predictions\nusing machine learning where the notion of fairness can be quantified using\npath-specific causal effects. We use a recently developed approach based on\nLagrange multipliers for infinite-dimensional functional estimation to derive\nclosed-form solutions for constrained optimization based on mean squared error\nand cross-entropy risk criteria. The theoretical forms of the solutions are\nanalyzed in detail and described as nuanced adjustments to the unconstrained\nminimizer. This analysis highlights important trade-offs between risk\nminimization and achieving fairnes. The theoretical solutions are also used as\nthe basis for construction of flexible semiparametric estimation strategies for\nthese nuisance components. We describe the robustness properties of our\nestimators in terms of achieving the optimal constrained risk, as well as in\nterms of controlling the value of the constraint. We study via simulation the\nimpact of using robust estimators of pathway-specific effects to validate our\ntheory. This work advances the discourse on algorithmic fairness by integrating\ncomplex causal considerations into model training, thus providing strategies\nfor implementing fair models in real-world applications.\n", "date": "2024-08-03", "categories": []}
{"id": "2408.02464", "title": "Fairness and Bias Mitigation in Computer Vision: A Survey", "abstract": "  Computer vision systems have witnessed rapid progress over the past two\ndecades due to multiple advances in the field. As these systems are\nincreasingly being deployed in high-stakes real-world applications, there is a\ndire need to ensure that they do not propagate or amplify any discriminatory\ntendencies in historical or human-curated data or inadvertently learn biases\nfrom spurious correlations. This paper presents a comprehensive survey on\nfairness that summarizes and sheds light on ongoing trends and successes in the\ncontext of computer vision. The topics we discuss include 1) The origin and\ntechnical definitions of fairness drawn from the wider fair machine learning\nliterature and adjacent disciplines. 2) Work that sought to discover and\nanalyze biases in computer vision systems. 3) A summary of methods proposed to\nmitigate bias in computer vision systems in recent years. 4) A comprehensive\nsummary of resources and datasets produced by researchers to measure, analyze,\nand mitigate bias and enhance fairness. 5) Discussion of the field's success,\ncontinuing trends in the context of multimodal foundation and generative\nmodels, and gaps that still need to be addressed. The presented\ncharacterization should help researchers understand the importance of\nidentifying and mitigating bias in computer vision and the state of the field\nand identify potential directions for future research.\n", "date": "2024-08-05", "categories": []}
{"id": "2403.15481", "title": "Navigating Fairness: Practitioners' Understanding, Challenges, and\n  Strategies in AI/ML Development", "abstract": "  The rise in the use of AI/ML applications across industries has sparked more\ndiscussions about the fairness of AI/ML in recent times. While prior research\non the fairness of AI/ML exists, there is a lack of empirical studies focused\non understanding the perspectives and experiences of AI practitioners in\ndeveloping a fair AI/ML system. Understanding AI practitioners' perspectives\nand experiences on the fairness of AI/ML systems are important because they are\ndirectly involved in its development and deployment and their insights can\noffer valuable real-world perspectives on the challenges associated with\nensuring fairness in AI/ML systems. We conducted semi-structured interviews\nwith 22 AI practitioners to investigate their understanding of what a 'fair\nAI/ML' is, the challenges they face in developing a fair AI/ML system, the\nconsequences of developing an unfair AI/ML system, and the strategies they\nemploy to ensure AI/ML system fairness. We developed a framework showcasing the\nrelationship between AI practitioners' understanding of 'fair AI/ML' system and\n(i) their challenges in its development, (ii) the consequences of developing an\nunfair AI/ML system, and (iii) strategies used to ensure AI/ML system fairness.\nBy exploring AI practitioners' perspectives and experiences, this study\nprovides actionable insights to enhance AI/ML fairness, which may promote\nfairer systems, reduce bias, and foster public trust in AI technologies.\nAdditionally, we also identify areas for further investigation and offer\nrecommendations to aid AI practitioners and AI companies in navigating\nfairness.\n", "date": "2024-03-21", "categories": []}
{"id": "2407.11733", "title": "How Are LLMs Mitigating Stereotyping Harms? Learning from Search Engine\n  Studies", "abstract": "  With the widespread availability of LLMs since the release of ChatGPT and\nincreased public scrutiny, commercial model development appears to have focused\ntheir efforts on 'safety' training concerning legal liabilities at the expense\nof social impact evaluation. This mimics a similar trend which we could observe\nfor search engine autocompletion some years prior. We draw on scholarship from\nNLP and search engine auditing and present a novel evaluation task in the style\nof autocompletion prompts to assess stereotyping in LLMs. We assess LLMs by\nusing four metrics, namely refusal rates, toxicity, sentiment and regard, with\nand without safety system prompts. Our findings indicate an improvement to\nstereotyping outputs with the system prompt, but overall a lack of attention by\nLLMs under study to certain harms classified as toxic, particularly for prompts\nabout peoples/ethnicities and sexual orientation. Mentions of intersectional\nidentities trigger a disproportionate amount of stereotyping. Finally, we\ndiscuss the implications of these findings about stereotyping harms in light of\nthe coming intermingling of LLMs and search and the choice of stereotyping\nmitigation policy to adopt. We address model builders, academics, NLP\npractitioners and policy makers, calling for accountability and awareness\nconcerning stereotyping harms, be it for training data curation, leader board\ndesign and usage, or social impact measurement.\n", "date": "2024-07-16", "categories": []}
{"id": "2408.00330", "title": "\"Patriarchy Hurts Men Too.\" Does Your Model Agree? A Discussion on\n  Fairness Assumptions", "abstract": "  The pipeline of a fair ML practitioner is generally divided into three\nphases: 1) Selecting a fairness measure. 2) Choosing a model that minimizes\nthis measure. 3) Maximizing the model's performance on the data. In the context\nof group fairness, this approach often obscures implicit assumptions about how\nbias is introduced into the data. For instance, in binary classification, it is\noften assumed that the best model, with equal fairness, is the one with better\nperformance. However, this belief already imposes specific properties on the\nprocess that introduced bias. More precisely, we are already assuming that the\nbiasing process is a monotonic function of the fair scores, dependent solely on\nthe sensitive attribute. We formally prove this claim regarding several\nimplicit fairness assumptions. This leads, in our view, to two possible\nconclusions: either the behavior of the biasing process is more complex than\nmere monotonicity, which means we need to identify and reject our implicit\nassumptions in order to develop models capable of tackling more complex\nsituations; or the bias introduced in the data behaves predictably, implying\nthat many of the developed models are superfluous.\n", "date": "2024-08-01", "categories": []}
{"id": "2407.15439", "title": "Merit-based Fair Combinatorial Semi-Bandit with Unrestricted Feedback\n  Delays", "abstract": "  We study the stochastic combinatorial semi-bandit problem with unrestricted\nfeedback delays under merit-based fairness constraints. This is motivated by\napplications such as crowdsourcing, and online advertising, where immediate\nfeedback is not immediately available and fairness among different choices (or\narms) is crucial. We consider two types of unrestricted feedback delays:\nreward-independent delays where the feedback delays are independent of the\nrewards, and reward-dependent delays where the feedback delays are correlated\nwith the rewards. Furthermore, we introduce merit-based fairness constraints to\nensure a fair selection of the arms. We define the reward regret and the\nfairness regret and present new bandit algorithms to select arms under\nunrestricted feedback delays based on their merits. We prove that our\nalgorithms all achieve sublinear expected reward regret and expected fairness\nregret, with a dependence on the quantiles of the delay distribution. We also\nconduct extensive experiments using synthetic and real-world data and show that\nour algorithms can fairly select arms with different feedback delays.\n", "date": "2024-07-22", "categories": []}
{"id": "2407.19646", "title": "Foundations for Unfairness in Anomaly Detection -- Case Studies in\n  Facial Imaging Data", "abstract": "  Deep anomaly detection (AD) is perhaps the most controversial of data\nanalytic tasks as it identifies entities that are then specifically targeted\nfor further investigation or exclusion. Also controversial is the application\nof AI to facial imaging data. This work explores the intersection of these two\nareas to understand two core questions: \"Who\" these algorithms are being unfair\nto and equally important \"Why\". Recent work has shown that deep AD can be\nunfair to different groups despite being unsupervised with a recent study\nshowing that for portraits of people: men of color are far more likely to be\nchosen to be outliers. We study the two main categories of AD algorithms:\nautoencoder-based and single-class-based which effectively try to compress all\nthe instances with those that can not be easily compressed being deemed to be\noutliers. We experimentally verify sources of unfairness such as the\nunder-representation of a group (e.g. people of color are relatively rare),\nspurious group features (e.g. men are often photographed with hats), and group\nlabeling noise (e.g. race is subjective). We conjecture that lack of\ncompressibility is the main foundation and the others cause it but experimental\nresults show otherwise and we present a natural hierarchy amongst them.\n", "date": "2024-07-29", "categories": []}
{"id": "2407.18376", "title": "Exploring Bengali Religious Dialect Biases in Large Language Models with\n  Evaluation Perspectives", "abstract": "  While Large Language Models (LLM) have created a massive technological impact\nin the past decade, allowing for human-enabled applications, they can produce\noutput that contains stereotypes and biases, especially when using low-resource\nlanguages. This can be of great ethical concern when dealing with sensitive\ntopics such as religion. As a means toward making LLMS more fair, we explore\nbias from a religious perspective in Bengali, focusing specifically on two main\nreligious dialects: Hindu and Muslim-majority dialects. Here, we perform\ndifferent experiments and audit showing the comparative analysis of different\nsentences using three commonly used LLMs: ChatGPT, Gemini, and Microsoft\nCopilot, pertaining to the Hindu and Muslim dialects of specific words and\nshowcasing which ones catch the social biases and which do not. Furthermore, we\nanalyze our findings and relate them to potential reasons and evaluation\nperspectives, considering their global impact with over 300 million speakers\nworldwide. With this work, we hope to establish the rigor for creating more\nfairness in LLMs, as these are widely used as creative writing agents.\n", "date": "2024-07-25", "categories": []}
{"id": "2407.18524", "title": "She Works, He Works: A Curious Exploration of Gender Bias in\n  AI-Generated Imagery", "abstract": "  This paper examines gender bias in AI-generated imagery of construction\nworkers, highlighting discrepancies in the portrayal of male and female\nfigures. Grounded in Griselda Pollock's theories on visual culture and gender,\nthe analysis reveals that AI models tend to sexualize female figures while\nportraying male figures as more authoritative and competent. These findings\nunderscore AI's potential to mirror and perpetuate societal biases, emphasizing\nthe need for critical engagement with AI-generated content. The project\ncontributes to discussions on the ethical implications of AI in creative\npractices and its broader impact on cultural perceptions of gender.\n", "date": "2024-07-26", "categories": []}
{"id": "2407.18689", "title": "The BIAS Detection Framework: Bias Detection in Word Embeddings and\n  Language Models for European Languages", "abstract": "  The project BIAS: Mitigating Diversity Biases of AI in the Labor Market is a\nfour-year project funded by the European commission and supported by the Swiss\nState Secretariat for Education, Research and Innovation (SERI). As part of the\nproject, novel bias detection methods to identify societal bias in language\nmodels and word embeddings in European languages are developed, with particular\nattention to linguistic and geographic particularities. This technical report\ndescribes the overall architecture and components of the BIAS Detection\nFramework. The code described in this technical report is available and will be\nupdated and expanded continuously with upcoming results from the BIAS project.\nThe details about the datasets for the different languages are described in\ncorresponding papers at scientific venues.\n", "date": "2024-07-26", "categories": []}
{"id": "2407.16892", "title": "Exploring Fusion Techniques in Multimodal AI-Based Recruitment: Insights\n  from FairCVdb", "abstract": "  Despite the large body of work on fairness-aware learning for individual\nmodalities like tabular data, images, and text, less work has been done on\nmultimodal data, which fuses various modalities for a comprehensive analysis.\nIn this work, we investigate the fairness and bias implications of multimodal\nfusion techniques in the context of multimodal AI-based recruitment systems\nusing the FairCVdb dataset. Our results show that early-fusion closely matches\nthe ground truth for both demographics, achieving the lowest MAEs by\nintegrating each modality's unique characteristics. In contrast, late-fusion\nleads to highly generalized mean scores and higher MAEs. Our findings emphasise\nthe significant potential of early-fusion for accurate and fair applications,\neven in the presence of demographic biases, compared to late-fusion. Future\nresearch could explore alternative fusion strategies and incorporate\nmodality-related fairness constraints to improve fairness. For code and\nadditional insights, visit:\nhttps://github.com/Swati17293/Multimodal-AI-Based-Recruitment-FairCVdb\n", "date": "2024-06-17", "categories": []}
{"id": "2407.16895", "title": "(Unfair) Norms in Fairness Research: A Meta-Analysis", "abstract": "  Algorithmic fairness has emerged as a critical concern in artificial\nintelligence (AI) research. However, the development of fair AI systems is not\nan objective process. Fairness is an inherently subjective concept, shaped by\nthe values, experiences, and identities of those involved in research and\ndevelopment. To better understand the norms and values embedded in current\nfairness research, we conduct a meta-analysis of algorithmic fairness papers\nfrom two leading conferences on AI fairness and ethics, AIES and FAccT,\ncovering a final sample of 139 papers over the period from 2018 to 2022. Our\ninvestigation reveals two concerning trends: first, a US-centric perspective\ndominates throughout fairness research; and second, fairness studies exhibit a\nwidespread reliance on binary codifications of human identity (e.g.,\n\"Black/White\", \"male/female\"). These findings highlight how current research\noften overlooks the complexities of identity and lived experiences, ultimately\nfailing to represent diverse global contexts when defining algorithmic bias and\nfairness. We discuss the limitations of these research design choices and offer\nrecommendations for fostering more inclusive and representative approaches to\nfairness in AI systems, urging a paradigm shift that embraces nuanced, global\nunderstandings of human identity and values.\n", "date": "2024-06-17", "categories": []}
{"id": "2407.16953", "title": "Open Challenges on Fairness of Artificial Intelligence in Medical\n  Imaging Applications", "abstract": "  Recently, the research community of computerized medical imaging has started\nto discuss and address potential fairness issues that may emerge when\ndeveloping and deploying AI systems for medical image analysis. This chapter\ncovers some of the pressing challenges encountered when doing research in this\narea, and it is intended to raise questions and provide food for thought for\nthose aiming to enter this research field. The chapter first discusses various\nsources of bias, including data collection, model training, and clinical\ndeployment, and their impact on the fairness of machine learning algorithms in\nmedical image computing. We then turn to discussing open challenges that we\nbelieve require attention from researchers and practitioners, as well as\npotential pitfalls of naive application of common methods in the field. We\ncover a variety of topics including the impact of biased metrics when auditing\nfor fairness, the leveling down effect, task difficulty variations among\nsubgroups, discovering biases in unseen populations, and explaining biases\nbeyond standard demographic attributes.\n", "date": "2024-07-24", "categories": []}
{"id": "2407.17459", "title": "Hidden or Inferred: Fair Learning-To-Rank with Unknown Demographics", "abstract": "  As learning-to-rank models are increasingly deployed for decision-making in\nareas with profound life implications, the FairML community has been developing\nfair learning-to-rank (LTR) models. These models rely on the availability of\nsensitive demographic features such as race or sex. However, in practice,\nregulatory obstacles and privacy concerns protect this data from collection and\nuse. As a result, practitioners may either need to promote fairness despite the\nabsence of these features or turn to demographic inference tools to attempt to\ninfer them. Given that these tools are fallible, this paper aims to further\nunderstand how errors in demographic inference impact the fairness performance\nof popular fair LTR strategies. In which cases would it be better to keep such\ndemographic attributes hidden from models versus infer them? We examine a\nspectrum of fair LTR strategies ranging from fair LTR with and without\ndemographic features hidden versus inferred to fairness-unaware LTR followed by\nfair re-ranking. We conduct a controlled empirical investigation modeling\ndifferent levels of inference errors by systematically perturbing the inferred\nsensitive attribute. We also perform three case studies with real-world\ndatasets and popular open-source inference methods. Our findings reveal that as\ninference noise grows, LTR-based methods that incorporate fairness\nconsiderations into the learning process may increase bias. In contrast, fair\nre-ranking strategies are more robust to inference errors. All source code,\ndata, and experimental artifacts of our experimental study are available here:\nhttps://github.com/sewen007/hoiltr.git\n", "date": "2024-07-24", "categories": []}
{"id": "2407.16431", "title": "FairFlow: An Automated Approach to Model-based Counterfactual Data\n  Augmentation For NLP", "abstract": "  Despite the evolution of language models, they continue to portray harmful\nsocietal biases and stereotypes inadvertently learned from training data. These\ninherent biases often result in detrimental effects in various applications.\nCounterfactual Data Augmentation (CDA), which seeks to balance demographic\nattributes in training data, has been a widely adopted approach to mitigate\nbias in natural language processing. However, many existing CDA approaches rely\non word substitution techniques using manually compiled word-pair dictionaries.\nThese techniques often lead to out-of-context substitutions, resulting in\npotential quality issues. The advancement of model-based techniques, on the\nother hand, has been challenged by the need for parallel training data. Works\nin this area resort to manually generated parallel data that are expensive to\ncollect and are consequently limited in scale. This paper proposes FairFlow, an\nautomated approach to generating parallel data for training counterfactual text\ngenerator models that limits the need for human intervention. Furthermore, we\nshow that FairFlow significantly overcomes the limitations of dictionary-based\nword-substitution approaches whilst maintaining good performance.\n", "date": "2024-07-23", "categories": []}
{"id": "2407.14087", "title": "Score Normalization for Demographic Fairness in Face Recognition", "abstract": "  Fair biometric algorithms have similar verification performance across\ndifferent demographic groups given a single decision threshold. Unfortunately,\nfor state-of-the-art face recognition networks, score distributions differ\nbetween demographics. Contrary to work that tries to align those distributions\nby extra training or fine-tuning, we solely focus on score post-processing\nmethods. As proved, well-known sample-centered score normalization techniques,\nZ-norm and T-norm, do not improve fairness for high-security operating points.\nThus, we extend the standard Z/T-norm to integrate demographic information in\nnormalization. Additionally, we investigate several possibilities to\nincorporate cohort similarities for both genuine and impostor pairs per\ndemographic to improve fairness across different operating points. We run\nexperiments on two datasets with different demographics (gender and ethnicity)\nand show that our techniques generally improve the overall fairness of five\nstate-of-the-art pre-trained face recognition networks, without downgrading\nverification performance. We also indicate that an equal contribution of False\nMatch Rate (FMR) and False Non-Match Rate (FNMR) in fairness evaluation is\nrequired for the highest gains. Code and protocols are available.\n", "date": "2024-07-19", "categories": []}
{"id": "2407.14799", "title": "FairViT: Fair Vision Transformer via Adaptive Masking", "abstract": "  Vision Transformer (ViT) has achieved excellent performance and demonstrated\nits promising potential in various computer vision tasks. The wide deployment\nof ViT in real-world tasks requires a thorough understanding of the societal\nimpact of the model. However, most ViT-based works do not take fairness into\naccount and it is unclear whether directly applying CNN-oriented debiased\nalgorithm to ViT is feasible. Moreover, previous works typically sacrifice\naccuracy for fairness. Therefore, we aim to develop an algorithm that improves\naccuracy without sacrificing fairness. In this paper, we propose FairViT, a\nnovel accurate and fair ViT framework. To this end, we introduce a novel\ndistance loss and deploy adaptive fairness-aware masks on attention layers\nupdating with model parameters. Experimental results show \\sys can achieve\naccuracy better than other alternatives, even with competitive computational\nefficiency. Furthermore, \\sys achieves appreciable fairness results.\n", "date": "2024-07-20", "categories": []}
{"id": "2407.15006", "title": "Fair Machine Learning for Healthcare Requires Recognizing the\n  Intersectionality of Sociodemographic Factors, a Case Study", "abstract": "  As interest in implementing artificial intelligence (AI) in medical systems\ngrows, discussion continues on how to evaluate the fairness of these systems,\nor the disparities they may perpetuate. Socioeconomic status (SES) is commonly\nincluded in machine learning models to control for health inequities, with the\nunderlying assumption that increased SES is associated with better health. In\nthis work, we considered a large cohort of patients from the Mount Sinai Health\nSystem in New York City to investigate the effect of patient SES, race, and sex\non schizophrenia (SCZ) diagnosis rates via a logistic regression model. Within\nan intersectional framework, patient SES, race, and sex were found to have\nsignificant interactions. Our findings showed that increased SES is associated\nwith a higher probability of obtaining a SCZ diagnosis in Black Americans\n($\\beta=4.1\\times10^{-8}$, $SE=4.5\\times10^{-9}$, $p < 0.001$). Whereas high\nSES acts as a protective factor for SCZ diagnosis in White Americans\n($\\beta=-4.1\\times10^{-8}$, $SE=6.7\\times10^{-9}$, $p < 0.001$). Further\ninvestigation is needed to reliably explain and quantify health disparities.\nNevertheless, we advocate that building fair AI tools for the health care space\nrequires recognizing the intersectionality of sociodemographic factors.\n", "date": "2024-05-30", "categories": []}
{"id": "2407.15453", "title": "Regression under demographic parity constraints via unlabeled\n  post-processing", "abstract": "  We address the problem of performing regression while ensuring demographic\nparity, even without access to sensitive attributes during inference. We\npresent a general-purpose post-processing algorithm that, using accurate\nestimates of the regression function and a sensitive attribute predictor,\ngenerates predictions that meet the demographic parity constraint. Our method\ninvolves discretization and stochastic minimization of a smooth convex\nfunction. It is suitable for online post-processing and multi-class\nclassification tasks only involving unlabeled data for the post-processing.\nUnlike prior methods, our approach is fully theory-driven. We require precise\ncontrol over the gradient norm of the convex function, and thus, we rely on\nmore advanced techniques than standard stochastic gradient descent. Our\nalgorithm is backed by finite-sample analysis and post-processing bounds, with\nexperimental results validating our theoretical findings.\n", "date": "2024-07-22", "categories": []}
{"id": "2207.13665", "title": "Causal foundations of bias, disparity and fairness", "abstract": "  The study of biases, such as gender or racial biases, is an important topic\nin the social and behavioural sciences. However, the literature does not always\nclearly define the concept. Definitions of bias are often ambiguous or not\nprovided at all. To study biases in a precise manner, it is important to have a\nwell-defined concept of bias. We propose to define bias as a direct causal\neffect that is unjustified. We propose to define the closely related concept of\ndisparity as a direct or indirect causal effect that includes a bias. Our\nproposed definitions can be used to study biases and disparities in a more\nrigorous and systematic way. We compare our definitions of bias and disparity\nwith various criteria of fairness introduced in the artificial intelligence\nliterature. In addition, we discuss how our definitions relate to\ndiscrimination. We illustrate our definitions of bias and disparity in two case\nstudies, focusing on gender bias in science and racial bias in police\nshootings. Our proposed definitions aim to contribute to a better appreciation\nof the causal intricacies of studies of biases and disparities. We hope that\nthis will also promote an improved understanding of the policy implications of\nsuch studies.\n", "date": "2022-07-27", "categories": []}
{"id": "2403.14200", "title": "Debiasing surgeon: fantastic weights and how to find them", "abstract": "  Nowadays an ever-growing concerning phenomenon, the emergence of algorithmic\nbiases that can lead to unfair models, emerges. Several debiasing approaches\nhave been proposed in the realm of deep learning, employing more or less\nsophisticated approaches to discourage these models from massively employing\nthese biases. However, a question emerges: is this extra complexity really\nnecessary? Is a vanilla-trained model already embodying some ``unbiased\nsub-networks'' that can be used in isolation and propose a solution without\nrelying on the algorithmic biases? In this work, we show that such a\nsub-network typically exists, and can be extracted from a vanilla-trained model\nwithout requiring additional training. We further validate that such specific\narchitecture is incapable of learning a specific bias, suggesting that there\nare possible architectural countermeasures to the problem of biases in deep\nneural networks.\n", "date": "2024-03-21", "categories": []}
{"id": "2404.07354", "title": "FairEM360: A Suite for Responsible Entity Matching", "abstract": "  Entity matching is one the earliest tasks that occur in the big data pipeline\nand is alarmingly exposed to unintentional biases that affect the quality of\ndata. Identifying and mitigating the biases that exist in the data or are\nintroduced by the matcher at this stage can contribute to promoting fairness in\ndownstream tasks. This demonstration showcases FairEM360, a framework for 1)\nauditing the output of entity matchers across a wide range of fairness measures\nand paradigms, 2) providing potential explanations for the underlying reasons\nfor unfairness, and 3) providing resolutions for the unfairness issues through\nan exploratory process with human-in-the-loop feedback, utilizing an ensemble\nof matchers. We aspire for FairEM360 to contribute to the prioritization of\nfairness as a key consideration in the evaluation of EM pipelines.\n", "date": "2024-04-10", "categories": []}
{"id": "2407.08813", "title": "FairDomain: Achieving Fairness in Cross-Domain Medical Image\n  Segmentation and Classification", "abstract": "  Addressing fairness in artificial intelligence (AI), particularly in medical\nAI, is crucial for ensuring equitable healthcare outcomes. Recent efforts to\nenhance fairness have introduced new methodologies and datasets in medical AI.\nHowever, the fairness issue under the setting of domain transfer is almost\nunexplored, while it is common that clinics rely on different imaging\ntechnologies (e.g., different retinal imaging modalities) for patient\ndiagnosis. This paper presents FairDomain, a pioneering systemic study into\nalgorithmic fairness under domain shifts, employing state-of-the-art domain\nadaptation (DA) and generalization (DG) algorithms for both medical\nsegmentation and classification tasks to understand how biases are transferred\nbetween different domains. We also introduce a novel plug-and-play fair\nidentity attention (FIA) module that adapts to various DA and DG algorithms to\nimprove fairness by using self-attention to adjust feature importance based on\ndemographic attributes. Additionally, we curate the first fairness-focused\ndataset with two paired imaging modalities for the same patient cohort on\nmedical segmentation and classification tasks, to rigorously assess fairness in\ndomain-shift scenarios. Excluding the confounding impact of demographic\ndistribution variation between source and target domains will allow clearer\nquantification of the performance of domain transfer models. Our extensive\nevaluations reveal that the proposed FIA significantly enhances both model\nperformance accounted for fairness across all domain shift settings (i.e., DA\nand DG) with respect to different demographics, which outperforms existing\nmethods on both segmentation and classification. The code and data can be\naccessed at https://ophai.hms.harvard.edu/datasets/harvard-fairdomain20k.\n", "date": "2024-07-11", "categories": []}
{"id": "2407.13896", "title": "Data-Algorithm-Architecture Co-Optimization for Fair Neural Networks on\n  Skin Lesion Dataset", "abstract": "  As Artificial Intelligence (AI) increasingly integrates into our daily lives,\nfairness has emerged as a critical concern, particularly in medical AI, where\ndatasets often reflect inherent biases due to social factors like the\nunderrepresentation of marginalized communities and socioeconomic barriers to\ndata collection. Traditional approaches to mitigating these biases have focused\non data augmentation and the development of fairness-aware training algorithms.\nHowever, this paper argues that the architecture of neural networks, a core\ncomponent of Machine Learning (ML), plays a crucial role in ensuring fairness.\nWe demonstrate that addressing fairness effectively requires a holistic\napproach that simultaneously considers data, algorithms, and architecture.\nUtilizing Automated ML (AutoML) technology, specifically Neural Architecture\nSearch (NAS), we introduce a novel framework, BiaslessNAS, designed to achieve\nfair outcomes in analyzing skin lesion datasets. BiaslessNAS incorporates\nfairness considerations at every stage of the NAS process, leading to the\nidentification of neural networks that are not only more accurate but also\nsignificantly fairer. Our experiments show that BiaslessNAS achieves a 2.55%\nincrease in accuracy and a 65.50% improvement in fairness compared to\ntraditional NAS methods, underscoring the importance of integrating fairness\ninto neural network architecture for better outcomes in medical AI\napplications.\n", "date": "2024-07-18", "categories": []}
{"id": "2407.13982", "title": "Reexamining Racial Disparities in Automatic Speech Recognition\n  Performance: The Role of Confounding by Provenance", "abstract": "  Automatic speech recognition (ASR) models trained on large amounts of audio\ndata are now widely used to convert speech to written text in a variety of\napplications from video captioning to automated assistants used in healthcare\nand other domains. As such, it is important that ASR models and their use is\nfair and equitable. Prior work examining the performance of commercial ASR\nsystems on the Corpus of Regional African American Language (CORAAL)\ndemonstrated significantly worse ASR performance on African American English\n(AAE). The current study seeks to understand the factors underlying this\ndisparity by examining the performance of the current state-of-the-art neural\nnetwork based ASR system (Whisper, OpenAI) on the CORAAL dataset. Two key\nfindings have been identified as a result of the current study. The first\nconfirms prior findings of significant dialectal variation even across\nneighboring communities, and worse ASR performance on AAE that can be improved\nto some extent with fine-tuning of ASR models. The second is a novel finding\nnot discussed in prior work on CORAAL: differences in audio recording practices\nwithin the dataset have a significant impact on ASR accuracy resulting in a\n``confounding by provenance'' effect in which both language use and recording\nquality differ by study location. These findings highlight the need for further\nsystematic investigation to disentangle the effects of recording quality and\ninherent linguistic diversity when examining the fairness and bias present in\nneural ASR models, as any bias in ASR accuracy may have negative downstream\neffects on disparities in various domains of life in which ASR technology is\nused.\n", "date": "2024-07-19", "categories": []}
{"id": "2212.11672", "title": "Trustworthy Social Bias Measurement", "abstract": "  How do we design measures of social bias that we trust? While prior work has\nintroduced several measures, no measure has gained widespread trust: instead,\nmounting evidence argues we should distrust these measures. In this work, we\ndesign bias measures that warrant trust based on the cross-disciplinary theory\nof measurement modeling. To combat the frequently fuzzy treatment of social\nbias in NLP, we explicitly define social bias, grounded in principles drawn\nfrom social science research. We operationalize our definition by proposing a\ngeneral bias measurement framework DivDist, which we use to instantiate 5\nconcrete bias measures. To validate our measures, we propose a rigorous testing\nprotocol with 8 testing criteria (e.g. predictive validity: do measures predict\nbiases in US employment?). Through our testing, we demonstrate considerable\nevidence to trust our measures, showing they overcome conceptual, technical,\nand empirical deficiencies present in prior measures.\n", "date": "2022-12-20", "categories": []}
{"id": "2407.13052", "title": "Matchings, Predictions and Counterfactual Harm in Refugee Resettlement\n  Processes", "abstract": "  Resettlement agencies have started to adopt data-driven algorithmic matching\nto match refugees to locations using employment rate as a measure of utility.\nGiven a pool of refugees, data-driven algorithmic matching utilizes a\nclassifier to predict the probability that each refugee would find employment\nat any given location. Then, it uses the predicted probabilities to estimate\nthe expected utility of all possible placement decisions. Finally, it finds the\nplacement decisions that maximize the predicted utility by solving a maximum\nweight bipartite matching problem. In this work, we argue that, using existing\nsolutions, there may be pools of refugees for which data-driven algorithmic\nmatching is (counterfactually) harmful -- it would have achieved lower utility\nthan a given default policy used in the past, had it been used. Then, we\ndevelop a post-processing algorithm that, given placement decisions made by a\ndefault policy on a pool of refugees and their employment outcomes, solves an\ninverse~matching problem to minimally modify the predictions made by a given\nclassifier. Under these modified predictions, the optimal matching policy that\nmaximizes predicted utility on the pool is guaranteed to be not harmful.\nFurther, we introduce a Transformer model that, given placement decisions made\nby a default policy on multiple pools of refugees and their employment\noutcomes, learns to modify the predictions made by a classifier so that the\noptimal matching policy that maximizes predicted utility under the modified\npredictions on an unseen pool of refugees is less likely to be harmful than\nunder the original predictions. Experiments on simulated resettlement processes\nusing synthetic refugee data created from a variety of publicly available data\nsuggest that our methodology may be effective in making algorithmic placement\ndecisions that are less likely to be harmful than existing solutions.\n", "date": "2024-05-24", "categories": []}
{"id": "2312.01261", "title": "TIBET: Identifying and Evaluating Biases in Text-to-Image Generative\n  Models", "abstract": "  Text-to-Image (TTI) generative models have shown great progress in the past\nfew years in terms of their ability to generate complex and high-quality\nimagery. At the same time, these models have been shown to suffer from harmful\nbiases, including exaggerated societal biases (e.g., gender, ethnicity), as\nwell as incidental correlations that limit such a model's ability to generate\nmore diverse imagery. In this paper, we propose a general approach to study and\nquantify a broad spectrum of biases, for any TTI model and for any prompt,\nusing counterfactual reasoning. Unlike other works that evaluate generated\nimages on a predefined set of bias axes, our approach automatically identifies\npotential biases that might be relevant to the given prompt, and measures those\nbiases. In addition, we complement quantitative scores with post-hoc\nexplanations in terms of semantic concepts in the images generated. We show\nthat our method is uniquely capable of explaining complex multi-dimensional\nbiases through semantic concepts, as well as the intersectionality between\ndifferent biases for any given prompt. We perform extensive user studies to\nillustrate that the results of our method and analysis are consistent with\nhuman judgements.\n", "date": "2023-12-03", "categories": []}
{"id": "2407.12031", "title": "Evaluation of Bias Towards Medical Professionals in Large Language\n  Models", "abstract": "  This study evaluates whether large language models (LLMs) exhibit biases\ntowards medical professionals. Fictitious candidate resumes were created to\ncontrol for identity factors while maintaining consistent qualifications. Three\nLLMs (GPT-4, Claude-3-haiku, and Mistral-Large) were tested using a\nstandardized prompt to evaluate resumes for specific residency programs.\nExplicit bias was tested by changing gender and race information, while\nimplicit bias was tested by changing names while hiding race and gender.\nPhysician data from the Association of American Medical Colleges was used to\ncompare with real-world demographics. 900,000 resumes were evaluated. All LLMs\nexhibited significant gender and racial biases across medical specialties.\nGender preferences varied, favoring male candidates in surgery and orthopedics,\nwhile preferring females in dermatology, family medicine, obstetrics and\ngynecology, pediatrics, and psychiatry. Claude-3 and Mistral-Large generally\nfavored Asian candidates, while GPT-4 preferred Black and Hispanic candidates\nin several specialties. Tests revealed strong preferences towards Hispanic\nfemales and Asian males in various specialties. Compared to real-world data,\nLLMs consistently chose higher proportions of female and underrepresented\nracial candidates than their actual representation in the medical workforce.\nGPT-4, Claude-3, and Mistral-Large showed significant gender and racial biases\nwhen evaluating medical professionals for residency selection. These findings\nhighlight the potential for LLMs to perpetuate biases and compromise healthcare\nworkforce diversity if used without proper bias mitigation strategies.\n", "date": "2024-06-30", "categories": []}
{"id": "2407.12112", "title": "A Benchmark for Fairness-Aware Graph Learning", "abstract": "  Fairness-aware graph learning has gained increasing attention in recent\nyears. Nevertheless, there lacks a comprehensive benchmark to evaluate and\ncompare different fairness-aware graph learning methods, which blocks\npractitioners from choosing appropriate ones for broader real-world\napplications. In this paper, we present an extensive benchmark on ten\nrepresentative fairness-aware graph learning methods. Specifically, we design a\nsystematic evaluation protocol and conduct experiments on seven real-world\ndatasets to evaluate these methods from multiple perspectives, including group\nfairness, individual fairness, the balance between different fairness criteria,\nand computational efficiency. Our in-depth analysis reveals key insights into\nthe strengths and limitations of existing methods. Additionally, we provide\npractical guidance for applying fairness-aware graph learning methods in\napplications. To the best of our knowledge, this work serves as an initial step\ntowards comprehensively understanding representative fairness-aware graph\nlearning methods to facilitate future advancements in this area.\n", "date": "2024-07-16", "categories": []}
{"id": "2402.12572", "title": "FairProof : Confidential and Certifiable Fairness for Neural Networks", "abstract": "  Machine learning models are increasingly used in societal applications, yet\nlegal and privacy concerns demand that they very often be kept confidential.\nConsequently, there is a growing distrust about the fairness properties of\nthese models in the minds of consumers, who are often at the receiving end of\nmodel predictions. To this end, we propose \\name -- a system that uses\nZero-Knowledge Proofs (a cryptographic primitive) to publicly verify the\nfairness of a model, while maintaining confidentiality. We also propose a\nfairness certification algorithm for fully-connected neural networks which is\nbefitting to ZKPs and is used in this system. We implement \\name in Gnark and\ndemonstrate empirically that our system is practically feasible. Code is\navailable at https://github.com/infinite-pursuits/FairProof.\n", "date": "2024-02-19", "categories": []}
{"id": "2407.11203", "title": "The Life Cycle of Large Language Models: A Review of Biases in Education", "abstract": "  Large Language Models (LLMs) are increasingly adopted in educational contexts\nto provide personalized support to students and teachers. The unprecedented\ncapacity of LLM-based applications to understand and generate natural language\ncan potentially improve instructional effectiveness and learning outcomes, but\nthe integration of LLMs in education technology has renewed concerns over\nalgorithmic bias which may exacerbate educational inequities. In this review,\nbuilding on prior work on mapping the traditional machine learning life cycle,\nwe provide a holistic map of the LLM life cycle from the initial development of\nLLMs to customizing pre-trained models for various applications in educational\nsettings. We explain each step in the LLM life cycle and identify potential\nsources of bias that may arise in the context of education. We discuss why\ncurrent measures of bias from traditional machine learning fail to transfer to\nLLM-generated content in education, such as tutoring conversations because the\ntext is high-dimensional, there can be multiple correct responses, and\ntailoring responses may be pedagogically desirable rather than unfair. This\nreview aims to clarify the complex nature of bias in LLM applications and\nprovide practical guidance for their evaluation to promote educational equity.\n", "date": "2024-06-03", "categories": []}
{"id": "2407.11624", "title": "Rethinking Fair Graph Neural Networks from Re-balancing", "abstract": "  Driven by the powerful representation ability of Graph Neural Networks\n(GNNs), plentiful GNN models have been widely deployed in many real-world\napplications. Nevertheless, due to distribution disparities between different\ndemographic groups, fairness in high-stake decision-making systems is receiving\nincreasing attention. Although lots of recent works devoted to improving the\nfairness of GNNs and achieved considerable success, they all require\nsignificant architectural changes or additional loss functions requiring more\nhyper-parameter tuning. Surprisingly, we find that simple re-balancing methods\ncan easily match or surpass existing fair GNN methods. We claim that the\nimbalance across different demographic groups is a significant source of\nunfairness, resulting in imbalanced contributions from each group to the\nparameters updating. However, these simple re-balancing methods have their own\nshortcomings during training. In this paper, we propose FairGB, Fair Graph\nNeural Network via re-Balancing, which mitigates the unfairness of GNNs by\ngroup balancing. Technically, FairGB consists of two modules: counterfactual\nnode mixup and contribution alignment loss. Firstly, we select counterfactual\npairs across inter-domain and inter-class, and interpolate the ego-networks to\ngenerate new samples. Guided by analysis, we can reveal the debiasing mechanism\nof our model by the causal view and prove that our strategy can make sensitive\nattributes statistically independent from target labels. Secondly, we reweigh\nthe contribution of each group according to gradients. By combining these two\nmodules, they can mutually promote each other. Experimental results on\nbenchmark datasets show that our method can achieve state-of-the-art results\nconcerning both utility and fairness metrics. Code is available at\nhttps://github.com/ZhixunLEE/FairGB.\n", "date": "2024-07-16", "categories": []}
{"id": "2309.00770", "title": "Bias and Fairness in Large Language Models: A Survey", "abstract": "  Rapid advancements of large language models (LLMs) have enabled the\nprocessing, understanding, and generation of human-like text, with increasing\nintegration into systems that touch our social sphere. Despite this success,\nthese models can learn, perpetuate, and amplify harmful social biases. In this\npaper, we present a comprehensive survey of bias evaluation and mitigation\ntechniques for LLMs. We first consolidate, formalize, and expand notions of\nsocial bias and fairness in natural language processing, defining distinct\nfacets of harm and introducing several desiderata to operationalize fairness\nfor LLMs. We then unify the literature by proposing three intuitive taxonomies,\ntwo for bias evaluation, namely metrics and datasets, and one for mitigation.\nOur first taxonomy of metrics for bias evaluation disambiguates the\nrelationship between metrics and evaluation datasets, and organizes metrics by\nthe different levels at which they operate in a model: embeddings,\nprobabilities, and generated text. Our second taxonomy of datasets for bias\nevaluation categorizes datasets by their structure as counterfactual inputs or\nprompts, and identifies the targeted harms and social groups; we also release a\nconsolidation of publicly-available datasets for improved access. Our third\ntaxonomy of techniques for bias mitigation classifies methods by their\nintervention during pre-processing, in-training, intra-processing, and\npost-processing, with granular subcategories that elucidate research trends.\nFinally, we identify open problems and challenges for future work. Synthesizing\na wide range of recent research, we aim to provide a clear guide of the\nexisting literature that empowers researchers and practitioners to better\nunderstand and prevent the propagation of bias in LLMs.\n", "date": "2023-09-02", "categories": []}
{"id": "2401.06310", "title": "ViSAGe: A Global-Scale Analysis of Visual Stereotypes in Text-to-Image\n  Generation", "abstract": "  Recent studies have shown that Text-to-Image (T2I) model generations can\nreflect social stereotypes present in the real world. However, existing\napproaches for evaluating stereotypes have a noticeable lack of coverage of\nglobal identity groups and their associated stereotypes. To address this gap,\nwe introduce the ViSAGe (Visual Stereotypes Around the Globe) dataset to enable\nthe evaluation of known nationality-based stereotypes in T2I models, across 135\nnationalities. We enrich an existing textual stereotype resource by\ndistinguishing between stereotypical associations that are more likely to have\nvisual depictions, such as `sombrero', from those that are less visually\nconcrete, such as 'attractive'. We demonstrate ViSAGe's utility through a\nmulti-faceted evaluation of T2I generations. First, we show that stereotypical\nattributes in ViSAGe are thrice as likely to be present in generated images of\ncorresponding identities as compared to other attributes, and that the\noffensiveness of these depictions is especially higher for identities from\nAfrica, South America, and South East Asia. Second, we assess the stereotypical\npull of visual depictions of identity groups, which reveals how the 'default'\nrepresentations of all identity groups in ViSAGe have a pull towards\nstereotypical depictions, and that this pull is even more prominent for\nidentity groups from the Global South. CONTENT WARNING: Some examples contain\noffensive stereotypes.\n", "date": "2024-01-12", "categories": []}
{"id": "2407.09551", "title": "Diminishing Stereotype Bias in Image Generation Model using\n  Reinforcemenlent Learning Feedback", "abstract": "  This study addresses gender bias in image generation models using\nReinforcement Learning from Artificial Intelligence Feedback (RLAIF) with a\nnovel Denoising Diffusion Policy Optimization (DDPO) pipeline. By employing a\npretrained stable diffusion model and a highly accurate gender classification\nTransformer, the research introduces two reward functions: Rshift for shifting\ngender imbalances, and Rbalance for achieving and maintaining gender balance.\nExperiments demonstrate the effectiveness of this approach in mitigating bias\nwithout compromising image quality or requiring additional data or prompt\nmodifications. While focusing on gender bias, this work establishes a\nfoundation for addressing various forms of bias in AI systems, emphasizing the\nneed for responsible AI development. Future research directions include\nextending the methodology to other bias types, enhancing the RLAIF pipeline's\nrobustness, and exploring multi-prompt fine-tuning to further advance fairness\nand inclusivity in AI.\n", "date": "2024-06-27", "categories": []}
{"id": "2407.10629", "title": "Balancing the Scales: Reinforcement Learning for Fair Classification", "abstract": "  Fairness in classification tasks has traditionally focused on bias removal\nfrom neural representations, but recent trends favor algorithmic methods that\nembed fairness into the training process. These methods steer models towards\nfair performance, preventing potential elimination of valuable information that\narises from representation manipulation. Reinforcement Learning (RL), with its\ncapacity for learning through interaction and adjusting reward functions to\nencourage desired behaviors, emerges as a promising tool in this domain. In\nthis paper, we explore the usage of RL to address bias in imbalanced\nclassification by scaling the reward function to mitigate bias. We employ the\ncontextual multi-armed bandit framework and adapt three popular RL algorithms\nto suit our objectives, demonstrating a novel approach to mitigating bias.\n", "date": "2024-07-15", "categories": []}
{"id": "2407.10633", "title": "Evaluating Model Bias Requires Characterizing its Mistakes", "abstract": "  The ability to properly benchmark model performance in the face of spurious\ncorrelations is important to both build better predictors and increase\nconfidence that models are operating as intended. We demonstrate that\ncharacterizing (as opposed to simply quantifying) model mistakes across\nsubgroups is pivotal to properly reflect model biases, which are ignored by\nstandard metrics such as worst-group accuracy or accuracy gap. Inspired by the\nhypothesis testing framework, we introduce SkewSize, a principled and flexible\nmetric that captures bias from mistakes in a model's predictions. It can be\nused in multi-class settings or generalised to the open vocabulary setting of\ngenerative models. SkewSize is an aggregation of the effect size of the\ninteraction between two categorical variables: the spurious variable\nrepresenting the bias attribute and the model's prediction. We demonstrate the\nutility of SkewSize in multiple settings including: standard vision models\ntrained on synthetic data, vision models trained on ImageNet, and large scale\nvision-and-language models from the BLIP-2 family. In each case, the proposed\nSkewSize is able to highlight biases not captured by other metrics, while also\nproviding insights on the impact of recently proposed techniques, such as\ninstruction tuning.\n", "date": "2024-07-15", "categories": []}
{"id": "2309.07085", "title": "Mitigating Group Bias in Federated Learning for Heterogeneous Devices", "abstract": "  Federated Learning is emerging as a privacy-preserving model training\napproach in distributed edge applications. As such, most edge deployments are\nheterogeneous in nature i.e., their sensing capabilities and environments vary\nacross deployments. This edge heterogeneity violates the independence and\nidentical distribution (IID) property of local data across clients and produces\nbiased global models i.e. models that contribute to unfair decision-making and\ndiscrimination against a particular community or a group. Existing bias\nmitigation techniques only focus on bias generated from label heterogeneity in\nnon-IID data without accounting for domain variations due to feature\nheterogeneity and do not address global group-fairness property.\n  Our work proposes a group-fair FL framework that minimizes group-bias while\npreserving privacy and without resource utilization overhead. Our main idea is\nto leverage average conditional probabilities to compute a cross-domain group\n\\textit{importance weights} derived from heterogeneous training data to\noptimize the performance of the worst-performing group using a modified\nmultiplicative weights update method. Additionally, we propose regularization\ntechniques to minimize the difference between the worst and best-performing\ngroups while making sure through our thresholding mechanism to strike a balance\nbetween bias reduction and group performance degradation. Our evaluation of\nhuman emotion recognition and image classification benchmarks assesses the fair\ndecision-making of our framework in real-world heterogeneous settings.\n", "date": "2023-09-13", "categories": []}
{"id": "2301.13803", "title": "Fairness-aware Vision Transformer via Debiased Self-Attention", "abstract": "  Vision Transformer (ViT) has recently gained significant attention in solving\ncomputer vision (CV) problems due to its capability of extracting informative\nfeatures and modeling long-range dependencies through the attention mechanism.\nWhereas recent works have explored the trustworthiness of ViT, including its\nrobustness and explainability, the issue of fairness has not yet been\nadequately addressed. We establish that the existing fairness-aware algorithms\ndesigned for CNNs do not perform well on ViT, which highlights the need to\ndevelop our novel framework via Debiased Self-Attention (DSA). DSA is a\nfairness-through-blindness approach that enforces ViT to eliminate spurious\nfeatures correlated with the sensitive label for bias mitigation and\nsimultaneously retain real features for target prediction. Notably, DSA\nleverages adversarial examples to locate and mask the spurious features in the\ninput image patches with an additional attention weights alignment regularizer\nin the training objective to encourage learning real features for target\nprediction. Importantly, our DSA framework leads to improved fairness\nguarantees over prior works on multiple prediction tasks without compromising\ntarget prediction performance. Code is available at\n\\href{https://github.com/qiangyao1988/DSA}{https://github.com/qiangyao1988/DSA}.\n", "date": "2023-01-31", "categories": []}
{"id": "2407.03623", "title": "Resampled Datasets Are Not Enough: Mitigating Societal Bias Beyond\n  Single Attributes", "abstract": "  We tackle societal bias in image-text datasets by removing spurious\ncorrelations between protected groups and image attributes. Traditional methods\nonly target labeled attributes, ignoring biases from unlabeled ones. Using\ntext-guided inpainting models, our approach ensures protected group\nindependence from all attributes and mitigates inpainting biases through data\nfiltering. Evaluations on multi-label image classification and image captioning\ntasks show our method effectively reduces bias without compromising performance\nacross various models.\n", "date": "2024-07-04", "categories": []}
{"id": "2407.08189", "title": "fairBERTs: Erasing Sensitive Information Through Semantic and\n  Fairness-aware Perturbations", "abstract": "  Pre-trained language models (PLMs) have revolutionized both the natural\nlanguage processing research and applications. However, stereotypical biases\n(e.g., gender and racial discrimination) encoded in PLMs have raised negative\nethical implications for PLMs, which critically limits their broader\napplications. To address the aforementioned unfairness issues, we present\nfairBERTs, a general framework for learning fair fine-tuned BERT series models\nby erasing the protected sensitive information via semantic and fairness-aware\nperturbations generated by a generative adversarial network. Through extensive\nqualitative and quantitative experiments on two real-world tasks, we\ndemonstrate the great superiority of fairBERTs in mitigating unfairness while\nmaintaining the model utility. We also verify the feasibility of transferring\nadversarial components in fairBERTs to other conventionally trained BERT-like\nmodels for yielding fairness improvements. Our findings may shed light on\nfurther research on building fairer fine-tuned PLMs.\n", "date": "2024-07-11", "categories": []}
{"id": "2407.07350", "title": "Long-Term Fairness in Sequential Multi-Agent Selection with Positive\n  Reinforcement", "abstract": "  While much of the rapidly growing literature on fair decision-making focuses\non metrics for one-shot decisions, recent work has raised the intriguing\npossibility of designing sequential decision-making to positively impact\nlong-term social fairness. In selection processes such as college admissions or\nhiring, biasing slightly towards applicants from under-represented groups is\nhypothesized to provide positive feedback that increases the pool of\nunder-represented applicants in future selection rounds, thus enhancing\nfairness in the long term. In this paper, we examine this hypothesis and its\nconsequences in a setting in which multiple agents are selecting from a common\npool of applicants. We propose the Multi-agent Fair-Greedy policy, that\nbalances greedy score maximization and fairness. Under this policy, we prove\nthat the resource pool and the admissions converge to a long-term fairness\ntarget set by the agents when the score distributions across the groups in the\npopulation are identical. We provide empirical evidence of existence of\nequilibria under non-identical score distributions through synthetic and\nadapted real-world datasets. We then sound a cautionary note for more complex\napplicant pool evolution models, under which uncoordinated behavior by the\nagents can cause negative reinforcement, leading to a reduction in the fraction\nof under-represented applicants. Our results indicate that, while positive\nreinforcement is a promising mechanism for long-term fairness, policies must be\ndesigned carefully to be robust to variations in the evolution model, with a\nnumber of open issues that remain to be explored by algorithm designers, social\nscientists, and policymakers.\n", "date": "2024-07-10", "categories": []}
{"id": "2209.13177", "title": "Addressing Fairness Issues in Deep Learning-Based Medical Image\n  Analysis: A Systematic Review", "abstract": "  Deep learning algorithms have demonstrated remarkable efficacy in various\nmedical image analysis (MedIA) applications. However, recent research\nhighlights a performance disparity in these algorithms when applied to specific\nsubgroups, such as exhibiting poorer predictive performance in elderly females.\nAddressing this fairness issue has become a collaborative effort involving AI\nscientists and clinicians seeking to understand its origins and develop\nsolutions for mitigation within MedIA. In this survey, we thoroughly examine\nthe current advancements in addressing fairness issues in MedIA, focusing on\nmethodological approaches. We introduce the basics of group fairness and\nsubsequently categorize studies on fair MedIA into fairness evaluation and\nunfairness mitigation. Detailed methods employed in these studies are presented\ntoo. Our survey concludes with a discussion of existing challenges and\nopportunities in establishing a fair MedIA and healthcare system. By offering\nthis comprehensive review, we aim to foster a shared understanding of fairness\namong AI researchers and clinicians, enhance the development of unfairness\nmitigation methods, and contribute to the creation of an equitable MedIA\nsociety.\n", "date": "2022-09-27", "categories": []}
{"id": "2310.15097", "title": "A Canonical Data Transformation for Achieving Inter- and Within-group\n  Fairness", "abstract": "  Increases in the deployment of machine learning algorithms for applications\nthat deal with sensitive data have brought attention to the issue of fairness\nin machine learning. Many works have been devoted to applications that require\ndifferent demographic groups to be treated fairly. However, algorithms that aim\nto satisfy inter-group fairness (also called group fairness) may inadvertently\ntreat individuals within the same demographic group unfairly. To address this\nissue, we introduce a formal definition of within-group fairness that maintains\nfairness among individuals from within the same group. We propose a\npre-processing framework to meet both inter- and within-group fairness criteria\nwith little compromise in accuracy. The framework maps the feature vectors of\nmembers from different groups to an inter-group-fair canonical domain before\nfeeding them into a scoring function. The mapping is constructed to preserve\nthe relative relationship between the scores obtained from the unprocessed\nfeature vectors of individuals from the same demographic group, guaranteeing\nwithin-group fairness. We apply this framework to the COMPAS risk assessment\nand Law School datasets and compare its performance in achieving inter-group\nand within-group fairness to two regularization-based methods.\n", "date": "2023-10-23", "categories": []}
{"id": "2407.04804", "title": "Fair Submodular Cover", "abstract": "  Submodular optimization is a fundamental problem with many applications in\nmachine learning, often involving decision-making over datasets with sensitive\nattributes such as gender or age. In such settings, it is often desirable to\nproduce a diverse solution set that is fairly distributed with respect to these\nattributes. Motivated by this, we initiate the study of Fair Submodular Cover\n(FSC), where given a ground set $U$, a monotone submodular function\n$f:2^U\\to\\mathbb{R}_{\\ge 0}$, a threshold $\\tau$, the goal is to find a\nbalanced subset of $S$ with minimum cardinality such that $f(S)\\ge\\tau$. We\nfirst introduce discrete algorithms for FSC that achieve a bicriteria\napproximation ratio of $(\\frac{1}{\\epsilon}, 1-O(\\epsilon))$. We then present a\ncontinuous algorithm that achieves a $(\\ln\\frac{1}{\\epsilon},\n1-O(\\epsilon))$-bicriteria approximation ratio, which matches the best\napproximation guarantee of submodular cover without a fairness constraint.\nFinally, we complement our theoretical results with a number of empirical\nevaluations that demonstrate the effectiveness of our algorithms on instances\nof maximum coverage.\n", "date": "2024-07-05", "categories": []}
{"id": "2407.05271", "title": "Beyond Binary Gender Labels: Revealing Gender Biases in LLMs through\n  Gender-Neutral Name Predictions", "abstract": "  Name-based gender prediction has traditionally categorized individuals as\neither female or male based on their names, using a binary classification\nsystem. That binary approach can be problematic in the cases of gender-neutral\nnames that do not align with any one gender, among other reasons. Relying\nsolely on binary gender categories without recognizing gender-neutral names can\nreduce the inclusiveness of gender prediction tasks. We introduce an additional\ngender category, i.e., \"neutral\", to study and address potential gender biases\nin Large Language Models (LLMs). We evaluate the performance of several\nfoundational and large language models in predicting gender based on first\nnames only. Additionally, we investigate the impact of adding birth years to\nenhance the accuracy of gender prediction, accounting for shifting associations\nbetween names and genders over time. Our findings indicate that most LLMs\nidentify male and female names with high accuracy (over 80%) but struggle with\ngender-neutral names (under 40%), and the accuracy of gender prediction is\nhigher for English-based first names than non-English names. The experimental\nresults show that incorporating the birth year does not improve the overall\naccuracy of gender prediction, especially for names with evolving gender\nassociations. We recommend using caution when applying LLMs for gender\nidentification in downstream tasks, particularly when dealing with non-binary\ngender labels.\n", "date": "2024-07-07", "categories": []}
{"id": "2407.05398", "title": "A Fair Post-Processing Method based on the MADD Metric for Predictive\n  Student Models", "abstract": "  Predictive student models are increasingly used in learning environments.\nHowever, due to the rising social impact of their usage, it is now all the more\nimportant for these models to be both sufficiently accurate and fair in their\npredictions. To evaluate algorithmic fairness, a new metric has been developed\nin education, namely the Model Absolute Density Distance (MADD). This metric\nenables us to measure how different a predictive model behaves regarding two\ngroups of students, in order to quantify its algorithmic unfairness. In this\npaper, we thus develop a post-processing method based on this metric, that aims\nat improving the fairness while preserving the accuracy of relevant predictive\nmodels' results. We experiment with our approach on the task of predicting\nstudent success in an online course, using both simulated and real-world\neducational data, and obtain successful results. Our source code and data are\nin open access at https://github.com/melinaverger/MADD .\n", "date": "2024-07-07", "categories": []}
{"id": "2407.05732", "title": "FairPFN: Transformers Can do Counterfactual Fairness", "abstract": "  Machine Learning systems are increasingly prevalent across healthcare, law\nenforcement, and finance but often operate on historical data, which may carry\nbiases against certain demographic groups. Causal and counterfactual fairness\nprovides an intuitive way to define fairness that closely aligns with legal\nstandards. Despite its theoretical benefits, counterfactual fairness comes with\nseveral practical limitations, largely related to the reliance on domain\nknowledge and approximate causal discovery techniques in constructing a causal\nmodel. In this study, we take a fresh perspective on counterfactually fair\nprediction, building upon recent work in in context learning (ICL) and prior\nfitted networks (PFNs) to learn a transformer called FairPFN. This model is\npretrained using synthetic fairness data to eliminate the causal effects of\nprotected attributes directly from observational data, removing the requirement\nof access to the correct causal model in practice. In our experiments, we\nthoroughly assess the effectiveness of FairPFN in eliminating the causal impact\nof protected attributes on a series of synthetic case studies and real world\ndatasets. Our findings pave the way for a new and promising research area:\ntransformers for causal and counterfactual fairness.\n", "date": "2024-07-08", "categories": []}
{"id": "2407.06003", "title": "Surprising gender biases in GPT", "abstract": "  We present seven experiments exploring gender biases in GPT. Initially, GPT\nwas asked to generate demographics of a potential writer of twenty phrases\ncontaining feminine stereotypes and twenty with masculine stereotypes. Results\nshow a strong asymmetry, with stereotypically masculine sentences attributed to\na female more often than vice versa. For example, the sentence \"I love playing\nfotbal! Im practicing with my cosin Michael\" was constantly assigned by ChatGPT\nto a female writer. This phenomenon likely reflects that while initiatives to\nintegrate women in traditionally masculine roles have gained momentum, the\nreverse movement remains relatively underdeveloped. Subsequent experiments\ninvestigate the same issue in high-stakes moral dilemmas. GPT-4 finds it more\nappropriate to abuse a man to prevent a nuclear apocalypse than to abuse a\nwoman. This bias extends to other forms of violence central to the gender\nparity debate (abuse), but not to those less central (torture). Moreover, this\nbias increases in cases of mixed-sex violence for the greater good: GPT-4\nagrees with a woman using violence against a man to prevent a nuclear\napocalypse but disagrees with a man using violence against a woman for the same\npurpose. Finally, these biases are implicit, as they do not emerge when GPT-4\nis directly asked to rank moral violations. These results highlight the\nnecessity of carefully managing inclusivity efforts to prevent unintended\ndiscrimination.\n", "date": "2024-07-08", "categories": []}
{"id": "2305.13935", "title": "Distribution-aware Fairness Test Generation", "abstract": "  Ensuring that all classes of objects are detected with equal accuracy is\nessential in AI systems. For instance, being unable to identify any one class\nof objects could have fatal consequences in autonomous driving systems. Hence,\nensuring the reliability of image recognition systems is crucial. This work\naddresses how to validate group fairness in image recognition software. We\npropose a distribution-aware fairness testing approach (called DistroFair) that\nsystematically exposes class-level fairness violations in image classifiers via\na synergistic combination of out-of-distribution (OOD) testing and\nsemantic-preserving image mutation. DistroFair automatically learns the\ndistribution (e.g., number/orientation) of objects in a set of images. Then it\nsystematically mutates objects in the images to become OOD using three\nsemantic-preserving image mutations - object deletion, object insertion and\nobject rotation. We evaluate DistroFair using two well-known datasets\n(CityScapes and MS-COCO) and three major, commercial image recognition software\n(namely, Amazon Rekognition, Google Cloud Vision and Azure Computer Vision).\nResults show that about 21% of images generated by DistroFair reveal\nclass-level fairness violations using either ground truth or metamorphic\noracles. DistroFair is up to 2.3x more effective than two main baselines, i.e.,\n(a) an approach which focuses on generating images only within the distribution\n(ID) and (b) fairness analysis using only the original image dataset. We\nfurther observed that DistroFair is efficient, it generates 460 images per\nhour, on average. Finally, we evaluate the semantic validity of our approach\nvia a user study with 81 participants, using 30 real images and 30\ncorresponding mutated images generated by DistroFair. We found that images\ngenerated by DistroFair are 80% as realistic as real-world images.\n", "date": "2023-05-08", "categories": []}
{"id": "2305.01397", "title": "Are demographically invariant models and representations in medical\n  imaging fair?", "abstract": "  Medical imaging models have been shown to encode information about patient\ndemographics such as age, race, and sex in their latent representation, raising\nconcerns about their potential for discrimination. Here, we ask whether\nrequiring models not to encode demographic attributes is desirable. We point\nout that marginal and class-conditional representation invariance imply the\nstandard group fairness notions of demographic parity and equalized odds,\nrespectively. In addition, however, they require matching the risk\ndistributions, thus potentially equalizing away important group differences.\nEnforcing the traditional fairness notions directly instead does not entail\nthese strong constraints. Moreover, representationally invariant models may\nstill take demographic attributes into account for deriving predictions,\nimplying unequal treatment - in fact, achieving representation invariance may\nrequire doing so. In theory, this can be prevented using counterfactual notions\nof (individual) fairness or invariance. We caution, however, that properly\ndefining medical image counterfactuals with respect to demographic attributes\nis fraught with challenges. Finally, we posit that encoding demographic\nattributes may even be advantageous if it enables learning a task-specific\nencoding of demographic features that does not rely on social constructs such\nas 'race' and 'gender.' We conclude that demographically invariant\nrepresentations are neither necessary nor sufficient for fairness in medical\nimaging. Models may need to encode demographic attributes, lending further\nurgency to calls for comprehensive model fairness assessments in terms of\npredictive performance across diverse patient groups.\n", "date": "2023-05-02", "categories": []}
{"id": "2407.02805", "title": "Efficient DNN-Powered Software with Fair Sparse Models", "abstract": "  With the emergence of the Software 3.0 era, there is a growing trend of\ncompressing and integrating large models into software systems, with\nsignificant societal implications. Regrettably, in numerous instances, model\ncompression techniques impact the fairness performance of these models and thus\nthe ethical behavior of DNN-powered software. One of the most notable example\nis the Lottery Ticket Hypothesis (LTH), a prevailing model pruning approach.\nThis paper demonstrates that fairness issue of LTHbased pruning arises from\nboth its subnetwork selection and training procedures, highlighting the\ninadequacy of existing remedies. To address this, we propose a novel pruning\nframework, Ballot, which employs a novel conflict-detection-based subnetwork\nselection to find accurate and fair subnetworks, coupled with a refined\ntraining process to attain a high-performance model, thereby improving the\nfairness of DNN-powered software. By means of this procedure, Ballot improves\nthe fairness of pruning by 38.00%, 33.91%, 17.96%, and 35.82% compared to\nstate-of-the-art baselines, namely Magnitude Pruning, Standard LTH,\nSafeCompress, and FairScratch respectively, based on our evaluation of five\npopular datasets and three widely used models. Our code is available at\nhttps://anonymous.4open.science/r/Ballot-506E.\n", "date": "2024-07-03", "categories": []}
{"id": "2311.12560", "title": "Benchmarking bias: Expanding clinical AI model card to incorporate bias\n  reporting of social and non-social factors", "abstract": "  Clinical AI model reporting cards should be expanded to incorporate a broad\nbias reporting of both social and non-social factors. Non-social factors\nconsider the role of other factors, such as disease dependent, anatomic, or\ninstrument factors on AI model bias, which are essential to ensure safe\ndeployment.\n", "date": "2023-11-21", "categories": []}
{"id": "2407.01595", "title": "Fairpriori: Improving Biased Subgroup Discovery for Deep Neural Network\n  Fairness", "abstract": "  While deep learning has become a core functional module of most software\nsystems, concerns regarding the fairness of ML predictions have emerged as a\nsignificant issue that affects prediction results due to discrimination.\nIntersectional bias, which disproportionately affects members of subgroups, is\na prime example of this. For instance, a machine learning model might exhibit\nbias against darker-skinned women, while not showing bias against individuals\nwith darker skin or women. This problem calls for effective fairness testing\nbefore the deployment of such deep learning models in real-world scenarios.\nHowever, research into detecting such bias is currently limited compared to\nresearch on individual and group fairness. Existing tools to investigate\nintersectional bias lack important features such as support for multiple\nfairness metrics, fast and efficient computation, and user-friendly\ninterpretation. This paper introduces Fairpriori, a novel biased subgroup\ndiscovery method, which aims to address these limitations. Fairpriori\nincorporates the frequent itemset generation algorithm to facilitate effective\nand efficient investigation of intersectional bias by producing fast fairness\nmetric calculations on subgroups of a dataset. Through comparison with the\nstate-of-the-art methods (e.g., Themis, FairFictPlay, and TestSGD) under\nsimilar conditions, Fairpriori demonstrates superior effectiveness and\nefficiency when identifying intersectional bias. Specifically, Fairpriori is\neasier to use and interpret, supports a wider range of use cases by\naccommodating multiple fairness metrics, and exhibits higher efficiency in\ncomputing fairness metrics. These findings showcase Fairpriori's potential for\neffectively uncovering subgroups affected by intersectional bias, supported by\nits open-source tooling at https://anonymous.4open.science/r/Fairpriori-0320.\n", "date": "2024-06-25", "categories": []}
{"id": "2303.01504", "title": "Backdoor for Debias: Mitigating Model Bias with Backdoor Attack-based\n  Artificial Bias", "abstract": "  With the swift advancement of deep learning, state-of-the-art algorithms have\nbeen utilized in various social situations. Nonetheless, some algorithms have\nbeen discovered to exhibit biases and provide unequal results. The current\ndebiasing methods face challenges such as poor utilization of data or intricate\ntraining requirements. In this work, we found that the backdoor attack can\nconstruct an artificial bias similar to the model bias derived in standard\ntraining. Considering the strong adjustability of backdoor triggers, we are\nmotivated to mitigate the model bias by carefully designing reverse artificial\nbias created from backdoor attack. Based on this, we propose a backdoor\ndebiasing framework based on knowledge distillation, which effectively reduces\nthe model bias from original data and minimizes security risks from the\nbackdoor attack. The proposed solution is validated on both image and\nstructured datasets, showing promising results. This work advances the\nunderstanding of backdoor attacks and highlights its potential for beneficial\napplications. The code for the study can be found at\n\\url{https://anonymous.4open.science/r/DwB-BC07/}.\n", "date": "2023-03-01", "categories": []}
{"id": "2305.11602", "title": "Latent Imitator: Generating Natural Individual Discriminatory Instances\n  for Black-Box Fairness Testing", "abstract": "  Machine learning (ML) systems have achieved remarkable performance across a\nwide area of applications. However, they frequently exhibit unfair behaviors in\nsensitive application domains, raising severe fairness concerns. To evaluate\nand test fairness, engineers often generate individual discriminatory instances\nto expose unfair behaviors before model deployment. However, existing baselines\nignore the naturalness of generation and produce instances that deviate from\nthe real data distribution, which may fail to reveal the actual model fairness\nsince these unnatural discriminatory instances are unlikely to appear in\npractice. To address the problem, this paper proposes a framework named Latent\nImitator (LIMI) to generate more natural individual discriminatory instances\nwith the help of a generative adversarial network (GAN), where we imitate the\ndecision boundary of the target model in the semantic latent space of GAN and\nfurther samples latent instances on it. Specifically, we first derive a\nsurrogate linear boundary to coarsely approximate the decision boundary of the\ntarget model, which reflects the nature of the original data distribution.\nSubsequently, to obtain more natural instances, we manipulate random latent\nvectors to the surrogate boundary with a one-step movement, and further conduct\nvector calculation to probe two potential discriminatory candidates that may be\nmore closely located in the real decision boundary. Extensive experiments on\nvarious datasets demonstrate that our LIMI outperforms other baselines largely\nin effectiveness ($\\times$9.42 instances), efficiency ($\\times$8.71 speeds),\nand naturalness (+19.65%) on average. In addition, we empirically demonstrate\nthat retraining on test samples generated by our approach can lead to\nimprovements in both individual fairness (45.67% on $IF_r$ and 32.81% on\n$IF_o$) and group fairness (9.86% on $SPD$ and 28.38% on $AOD$}).\n", "date": "2023-05-19", "categories": []}
{"id": "2311.02115", "title": "Towards objective and systematic evaluation of bias in artificial\n  intelligence for medical imaging", "abstract": "  Artificial intelligence (AI) models trained using medical images for clinical\ntasks often exhibit bias in the form of disparities in performance between\nsubgroups. Since not all sources of biases in real-world medical imaging data\nare easily identifiable, it is challenging to comprehensively assess how those\nbiases are encoded in models, and how capable bias mitigation methods are at\nameliorating performance disparities. In this article, we introduce a novel\nanalysis framework for systematically and objectively investigating the impact\nof biases in medical images on AI models. We developed and tested this\nframework for conducting controlled in silico trials to assess bias in medical\nimaging AI using a tool for generating synthetic magnetic resonance images with\nknown disease effects and sources of bias. The feasibility is showcased by\nusing three counterfactual bias scenarios to measure the impact of simulated\nbias effects on a convolutional neural network (CNN) classifier and the\nefficacy of three bias mitigation strategies. The analysis revealed that the\nsimulated biases resulted in expected subgroup performance disparities when the\nCNN was trained on the synthetic datasets. Moreover, reweighing was identified\nas the most successful bias mitigation strategy for this setup, and we\ndemonstrated how explainable AI methods can aid in investigating the\nmanifestation of bias in the model using this framework. Developing fair AI\nmodels is a considerable challenge given that many and often unknown sources of\nbiases can be present in medical imaging datasets. In this work, we present a\nnovel methodology to objectively study the impact of biases and mitigation\nstrategies on deep learning pipelines, which can support the development of\nclinical AI that is robust and responsible.\n", "date": "2023-11-03", "categories": []}
{"id": "2404.19100", "title": "Predicting Fairness of ML Software Configurations", "abstract": "  This paper investigates the relationships between hyperparameters of machine\nlearning and fairness. Data-driven solutions are increasingly used in critical\nsocio-technical applications where ensuring fairness is important. Rather than\nexplicitly encoding decision logic via control and data structures, the ML\ndevelopers provide input data, perform some pre-processing, choose ML\nalgorithms, and tune hyperparameters (HPs) to infer a program that encodes the\ndecision logic. Prior works report that the selection of HPs can significantly\ninfluence fairness. However, tuning HPs to find an ideal trade-off between\naccuracy, precision, and fairness has remained an expensive and tedious task.\nCan we predict fairness of HP configuration for a given dataset? Are the\npredictions robust to distribution shifts?\n  We focus on group fairness notions and investigate the HP space of 5 training\nalgorithms. We first find that tree regressors and XGBoots significantly\noutperformed deep neural networks and support vector machines in accurately\npredicting the fairness of HPs. When predicting the fairness of ML\nhyperparameters under temporal distribution shift, the tree regressors\noutperforms the other algorithms with reasonable accuracy. However, the\nprecision depends on the ML training algorithm, dataset, and protected\nattributes. For example, the tree regressor model was robust for training data\nshift from 2014 to 2018 on logistic regression and discriminant analysis HPs\nwith sex as the protected attribute; but not for race and other training\nalgorithms. Our method provides a sound framework to efficiently perform\nfine-tuning of ML training algorithms and understand the relationships between\nHPs and fairness.\n", "date": "2024-04-29", "categories": []}
{"id": "2405.09341", "title": "Large Language Model Bias Mitigation from the Perspective of Knowledge\n  Editing", "abstract": "  Existing debiasing methods inevitably make unreasonable or undesired\npredictions as they are designated and evaluated to achieve parity across\ndifferent social groups but leave aside individual facts, resulting in modified\nexisting knowledge. In this paper, we first establish a new bias mitigation\nbenchmark BiasKE leveraging existing and additional constructed datasets, which\nsystematically assesses debiasing performance by complementary metrics on\nfairness, specificity, and generalization. Meanwhile, we propose a novel\ndebiasing method, Fairness Stamp (FAST), which enables editable fairness\nthrough fine-grained calibration on individual biased knowledge. Comprehensive\nexperiments demonstrate that FAST surpasses state-of-the-art baselines with\nremarkable debiasing performance while not hampering overall model capability\nfor knowledge preservation, highlighting the prospect of fine-grained debiasing\nstrategies for editable fairness in LLMs.\n", "date": "2024-05-15", "categories": []}
{"id": "2407.00170", "title": "Dataset Representativeness and Downstream Task Fairness", "abstract": "  Our society collects data on people for a wide range of applications, from\nbuilding a census for policy evaluation to running meaningful clinical trials.\nTo collect data, we typically sample individuals with the goal of accurately\nrepresenting a population of interest. However, current sampling processes\noften collect data opportunistically from data sources, which can lead to\ndatasets that are biased and not representative, i.e., the collected dataset\ndoes not accurately reflect the distribution of demographics of the true\npopulation. This is a concern because subgroups within the population can be\nunder- or over-represented in a dataset, which may harm generalizability and\nlead to an unequal distribution of benefits and harms from downstream tasks\nthat use such datasets (e.g., algorithmic bias in medical decision-making\nalgorithms). In this paper, we assess the relationship between dataset\nrepresentativeness and group-fairness of classifiers trained on that dataset.\nWe demonstrate that there is a natural tension between dataset\nrepresentativeness and classifier fairness; empirically we observe that\ntraining datasets with better representativeness can frequently result in\nclassifiers with higher rates of unfairness. We provide some intuition as to\nwhy this occurs via a set of theoretical results in the case of univariate\nclassifiers. We also find that over-sampling underrepresented groups can result\nin classifiers which exhibit greater bias to those groups. Lastly, we observe\nthat fairness-aware sampling strategies (i.e., those which are specifically\ndesigned to select data with high downstream fairness) will often over-sample\nmembers of majority groups. These results demonstrate that the relationship\nbetween dataset representativeness and downstream classifier fairness is\ncomplex; balancing these two quantities requires special care from both model-\nand dataset-designers.\n", "date": "2024-06-28", "categories": []}
{"id": "2404.11819", "title": "Utilizing Adversarial Examples for Bias Mitigation and Accuracy\n  Enhancement", "abstract": "  We propose a novel approach to mitigate biases in computer vision models by\nutilizing counterfactual generation and fine-tuning. While counterfactuals have\nbeen used to analyze and address biases in DNN models, the counterfactuals\nthemselves are often generated from biased generative models, which can\nintroduce additional biases or spurious correlations. To address this issue, we\npropose using adversarial images, that is images that deceive a deep neural\nnetwork but not humans, as counterfactuals for fair model training. Our\napproach leverages a curriculum learning framework combined with a fine-grained\nadversarial loss to fine-tune the model using adversarial examples. By\nincorporating adversarial images into the training data, we aim to prevent\nbiases from propagating through the pipeline. We validate our approach through\nboth qualitative and quantitative assessments, demonstrating improved bias\nmitigation and accuracy compared to existing methods. Qualitatively, our\nresults indicate that post-training, the decisions made by the model are less\ndependent on the sensitive attribute and our model better disentangles the\nrelationship between sensitive attributes and classification variables.\n", "date": "2024-04-18", "categories": []}
{"id": "2406.19668", "title": "PopAlign: Population-Level Alignment for Fair Text-to-Image Generation", "abstract": "  Text-to-image (T2I) models achieve high-fidelity generation through extensive\ntraining on large datasets. However, these models may unintentionally pick up\nundesirable biases of their training data, such as over-representation of\nparticular identities in gender or ethnicity neutral prompts. Existing\nalignment methods such as Reinforcement Learning from Human Feedback (RLHF) and\nDirect Preference Optimization (DPO) fail to address this problem effectively\nbecause they operate on pairwise preferences consisting of individual samples,\nwhile the aforementioned biases can only be measured at a population level. For\nexample, a single sample for the prompt \"doctor\" could be male or female, but a\nmodel generating predominantly male doctors even with repeated sampling\nreflects a gender bias. To address this limitation, we introduce PopAlign, a\nnovel approach for population-level preference optimization, while standard\noptimization would prefer entire sets of samples over others. We further derive\na stochastic lower bound that directly optimizes for individual samples from\npreferred populations over others for scalable training. Using human evaluation\nand standard image quality and bias metrics, we show that PopAlign\nsignificantly mitigates the bias of pretrained T2I models while largely\npreserving the generation quality. Code is available at\nhttps://github.com/jacklishufan/PopAlignSDXL.\n", "date": "2024-06-28", "categories": []}
{"id": "2404.18736", "title": "Mapping the Potential of Explainable AI for Fairness Along the AI\n  Lifecycle", "abstract": "  The widespread use of artificial intelligence (AI) systems across various\ndomains is increasingly surfacing issues related to algorithmic fairness,\nespecially in high-stakes scenarios. Thus, critical considerations of how\nfairness in AI systems might be improved -- and what measures are available to\naid this process -- are overdue. Many researchers and policymakers see\nexplainable AI (XAI) as a promising way to increase fairness in AI systems.\nHowever, there is a wide variety of XAI methods and fairness conceptions\nexpressing different desiderata, and the precise connections between XAI and\nfairness remain largely nebulous. Besides, different measures to increase\nalgorithmic fairness might be applicable at different points throughout an AI\nsystem's lifecycle. Yet, there currently is no coherent mapping of fairness\ndesiderata along the AI lifecycle. In this paper, we we distill eight fairness\ndesiderata, map them along the AI lifecycle, and discuss how XAI could help\naddress each of them. We hope to provide orientation for practical applications\nand to inspire XAI research specifically focused on these fairness desiderata.\n", "date": "2024-04-29", "categories": []}
{"id": "2406.18939", "title": "Evaluating AI Group Fairness: a Fuzzy Logic Perspective", "abstract": "  Artificial intelligence systems often address fairness concerns by evaluating\nand mitigating measures of group discrimination, for example that indicate\nbiases against certain genders or races. However, what constitutes group\nfairness depends on who is asked and the social context, whereas definitions\nare often relaxed to accept small deviations from the statistical constraints\nthey set out to impose. Here we decouple definitions of group fairness both\nfrom the context and from relaxation-related uncertainty by expressing them in\nthe axiomatic system of Basic fuzzy Logic (BL) with loosely understood\npredicates, like encountering group members. We then evaluate the definitions\nin subclasses of BL, such as Product or Lukasiewicz logics. Evaluation produces\ncontinuous instead of binary truth values by choosing the logic subclass and\ntruth values for predicates that reflect uncertain context-specific beliefs,\nsuch as stakeholder opinions gathered through questionnaires. Internally, it\nfollows logic-specific rules to compute the truth values of definitions. We\nshow that commonly held propositions standardize the resulting mathematical\nformulas and we transcribe logic and truth value choices to layperson terms, so\nthat anyone can answer them. We also use our framework to study several\nliterature definitions of algorithmic fairness, for which we rationalize\nprevious expedient practices that are non-probabilistic and show how to\nre-interpret their formulas and parameters in new contexts.\n", "date": "2024-06-27", "categories": []}
{"id": "2406.19066", "title": "Dancing in the Shadows: Harnessing Ambiguity for Fairer Classifiers", "abstract": "  This paper introduces a novel approach to bolster algorithmic fairness in\nscenarios where sensitive information is only partially known. In particular,\nwe propose to leverage instances with uncertain identity with regards to the\nsensitive attribute to train a conventional machine learning classifier. The\nenhanced fairness observed in the final predictions of this classifier\nhighlights the promising potential of prioritizing ambiguity (i.e.,\nnon-normativity) as a means to improve fairness guarantees in real-world\nclassification tasks.\n", "date": "2024-06-27", "categories": []}
{"id": "2402.11039", "title": "Robustness to Subpopulation Shift with Domain Label Noise via\n  Regularized Annotation of Domains", "abstract": "  Existing methods for last layer retraining that aim to optimize worst-group\naccuracy (WGA) rely heavily on well-annotated groups in the training data. We\nshow, both in theory and practice, that annotation-based data augmentations\nusing either downsampling or upweighting for WGA are susceptible to domain\nannotation noise, and in high-noise regimes approach the WGA of a model trained\nwith vanilla empirical risk minimization. We introduce Regularized Annotation\nof Domains (RAD) in order to train robust last layer classifiers without the\nneed for explicit domain annotations. Our results show that RAD is competitive\nwith other recently proposed domain annotation-free techniques. Most\nimportantly, RAD outperforms state-of-the-art annotation-reliant methods even\nwith only 5% noise in the training data for several publicly available\ndatasets.\n", "date": "2024-02-16", "categories": []}
{"id": "2403.20089", "title": "Implications of the AI Act for Non-Discrimination Law and Algorithmic\n  Fairness", "abstract": "  The topic of fairness in AI, as debated in the FATE (Fairness,\nAccountability, Transparency, and Ethics in AI) communities, has sparked\nmeaningful discussions in the past years. However, from a legal perspective,\nparticularly from the perspective of European Union law, many open questions\nremain. Whereas algorithmic fairness aims to mitigate structural inequalities\nat design-level, European non-discrimination law is tailored to individual\ncases of discrimination after an AI model has been deployed. The AI Act might\npresent a tremendous step towards bridging these two approaches by shifting\nnon-discrimination responsibilities into the design stage of AI models. Based\non an integrative reading of the AI Act, we comment on legal as well as\ntechnical enforcement problems and propose practical implications on bias\ndetection and bias correction in order to specify and comply with specific\ntechnical requirements.\n", "date": "2024-03-29", "categories": []}
{"id": "2406.15009", "title": "Fair, Manipulation-Robust, and Transparent Sortition", "abstract": "  Sortition, the random selection of political representatives, is increasingly\nbeing used around the world to choose participants of deliberative processes\nlike Citizens' Assemblies. Motivated by sortition's practical importance, there\nhas been a recent flurry of research on sortition algorithms, whose task it is\nto select a panel from among a pool of volunteers. This panel must satisfy\nquotas enforcing representation of key population subgroups. Past work has\ncontributed an algorithmic approach for fulfilling this task while ensuring\nthat volunteers' chances of selection are maximally equal, as measured by any\nconvex equality objective. The question, then, is: which equality objective is\nthe right one? Past work has mainly studied the objectives Minimax and Leximin,\nwhich respectively minimize the maximum and maximize the minimum chance of\nselection given to any volunteer. Recent work showed that both of these\nobjectives have key weaknesses: Minimax is highly robust to manipulation but is\narbitrarily unfair; oppositely, Leximin is highly fair but arbitrarily\nmanipulable.\n  In light of this gap, we propose a new equality objective, Goldilocks, that\naims to achieve these ideals simultaneously by ensuring that no volunteer\nreceives too little or too much chance of selection. We theoretically bound the\nextent to which Goldilocks achieves these ideals, finding that in an important\nsense, Goldilocks recovers among the best available solutions in a given\ninstance. We then extend our bounds to the case where the output of Goldilocks\nis transformed to achieve a third goal, Transparency. Our empirical analysis of\nGoldilocks in real data is even more promising: we find that this objective\nachieves nearly instance-optimal minimum and maximum selection probabilities\nsimultaneously in most real instances -- an outcome not even guaranteed to be\npossible for any algorithm.\n", "date": "2024-06-21", "categories": []}
{"id": "2406.17906", "title": "Unbiasing on the Fly: Explanation-Guided Human Oversight of Machine\n  Learning System Decisions", "abstract": "  The widespread adoption of ML systems across critical domains like hiring,\nfinance, and healthcare raises growing concerns about their potential for\ndiscriminatory decision-making based on protected attributes. While efforts to\nensure fairness during development are crucial, they leave deployed ML systems\nvulnerable to potentially exhibiting discrimination during their operations. To\naddress this gap, we propose a novel framework for on-the-fly tracking and\ncorrection of discrimination in deployed ML systems. Leveraging counterfactual\nexplanations, the framework continuously monitors the predictions made by an ML\nsystem and flags discriminatory outcomes. When flagged, post-hoc explanations\nrelated to the original prediction and the counterfactual alternatives are\npresented to a human reviewer for real-time intervention. This\nhuman-in-the-loop approach empowers reviewers to accept or override the ML\nsystem decision, enabling fair and responsible ML operation under dynamic\nsettings. While further work is needed for validation and refinement, this\nframework offers a promising avenue for mitigating discrimination and building\ntrust in ML systems deployed in a wide range of domains.\n", "date": "2024-06-25", "categories": []}
{"id": "2307.12797", "title": "Causal Fair Machine Learning via Rank-Preserving Interventional\n  Distributions", "abstract": "  A decision can be defined as fair if equal individuals are treated equally\nand unequals unequally. Adopting this definition, the task of designing machine\nlearning (ML) models that mitigate unfairness in automated decision-making\nsystems must include causal thinking when introducing protected attributes:\nFollowing a recent proposal, we define individuals as being normatively equal\nif they are equal in a fictitious, normatively desired (FiND) world, where the\nprotected attributes have no (direct or indirect) causal effect on the target.\nWe propose rank-preserving interventional distributions to define a specific\nFiND world in which this holds and a warping method for estimation. Evaluation\ncriteria for both the method and the resulting ML model are presented and\nvalidated through simulations. Experiments on empirical data showcase the\npractical application of our method and compare results with \"fairadapt\"\n(Ple\\v{c}ko and Meinshausen, 2020), a different approach for mitigating\nunfairness by causally preprocessing data that uses quantile regression\nforests. With this, we show that our warping approach effectively identifies\nthe most discriminated individuals and mitigates unfairness.\n", "date": "2023-07-24", "categories": []}
{"id": "2402.14208", "title": "LLM-Assisted Content Conditional Debiasing for Fair Text Embedding", "abstract": "  Mitigating biases in machine learning models has become an increasing concern\nin Natural Language Processing (NLP), particularly in developing fair text\nembeddings, which are crucial yet challenging for real-world applications like\nsearch engines. In response, this paper proposes a novel method for learning\nfair text embeddings. First, we define a novel content-conditional equal\ndistance (CCED) fairness for text embeddings, ensuring content-conditional\nindependence between sensitive attributes and text embeddings. Building on\nCCED, we introduce a content-conditional debiasing (CCD) loss to ensure that\nembeddings of texts with different sensitive attributes but identical content\nmaintain the same distance from the embedding of their corresponding neutral\ntext. Additionally, we tackle the issue of insufficient training data by using\nLarge Language Models (LLMs) with instructions to fairly augment texts into\ndifferent sensitive groups. Our extensive evaluations show that our approach\neffectively enhances fairness while maintaining the utility of embeddings.\nFurthermore, our augmented dataset, combined with the CCED metric, serves as an\nnew benchmark for evaluating fairness.\n", "date": "2024-02-22", "categories": []}
{"id": "2405.13025", "title": "A survey on fairness of large language models in e-commerce: progress,\n  application, and challenge", "abstract": "  This survey explores the fairness of large language models (LLMs) in\ne-commerce, examining their progress, applications, and the challenges they\nface. LLMs have become pivotal in the e-commerce domain, offering innovative\nsolutions and enhancing customer experiences. This work presents a\ncomprehensive survey on the applications and challenges of LLMs in e-commerce.\nThe paper begins by introducing the key principles underlying the use of LLMs\nin e-commerce, detailing the processes of pretraining, fine-tuning, and\nprompting that tailor these models to specific needs. It then explores the\nvaried applications of LLMs in e-commerce, including product reviews, where\nthey synthesize and analyze customer feedback; product recommendations, where\nthey leverage consumer data to suggest relevant items; product information\ntranslation, enhancing global accessibility; and product question and answer\nsections, where they automate customer support. The paper critically addresses\nthe fairness challenges in e-commerce, highlighting how biases in training data\nand algorithms can lead to unfair outcomes, such as reinforcing stereotypes or\ndiscriminating against certain groups. These issues not only undermine consumer\ntrust, but also raise ethical and legal concerns. Finally, the work outlines\nfuture research directions, emphasizing the need for more equitable and\ntransparent LLMs in e-commerce. It advocates for ongoing efforts to mitigate\nbiases and improve the fairness of these systems, ensuring they serve diverse\nglobal markets effectively and ethically. Through this comprehensive analysis,\nthe survey provides a holistic view of the current landscape of LLMs in\ne-commerce, offering insights into their potential and limitations, and guiding\nfuture endeavors in creating fairer and more inclusive e-commerce environments.\n", "date": "2024-05-15", "categories": []}
{"id": "2406.15960", "title": "Fair Clustering: Critique, Caveats, and Future Directions", "abstract": "  Clustering is a fundamental problem in machine learning and operations\nresearch. Therefore, given the fact that fairness considerations have become of\nparamount importance in algorithm design, fairness in clustering has received\nsignificant attention from the research community. The literature on fair\nclustering has resulted in a collection of interesting fairness notions and\nelaborate algorithms. In this paper, we take a critical view of fair\nclustering, identifying a collection of ignored issues such as the lack of a\nclear utility characterization and the difficulty in accounting for the\ndownstream effects of a fair clustering algorithm in machine learning settings.\nIn some cases, we demonstrate examples where the application of a fair\nclustering algorithm can have significant negative impacts on social welfare.\nWe end by identifying a collection of steps that would lead towards more\nimpactful research in fair clustering.\n", "date": "2024-06-22", "categories": []}
{"id": "2406.16557", "title": "Efficient k-means with Individual Fairness via Exponential Tilting", "abstract": "  In location-based resource allocation scenarios, the distances between each\nindividual and the facility are desired to be approximately equal, thereby\nensuring fairness. Individually fair clustering is often employed to achieve\nthe principle of treating all points equally, which can be applied in these\nscenarios. This paper proposes a novel algorithm, tilted k-means (TKM), aiming\nto achieve individual fairness in clustering. We integrate the exponential\ntilting into the sum of squared errors (SSE) to formulate a novel objective\nfunction called tilted SSE. We demonstrate that the tilted SSE can generalize\nto SSE and employ the coordinate descent and first-order gradient method for\noptimization. We propose a novel fairness metric, the variance of the distances\nwithin each cluster, which can alleviate the Matthew Effect typically caused by\nexisting fairness metrics. Our theoretical analysis demonstrates that the\nwell-known k-means++ incurs a multiplicative error of O(k log k), and we\nestablish the convergence of TKM under mild conditions. In terms of fairness,\nwe prove that the variance generated by TKM decreases with a scaled\nhyperparameter. In terms of efficiency, we demonstrate the time complexity is\nlinear with the dataset size. Our experiments demonstrate that TKM outperforms\nstate-of-the-art methods in effectiveness, fairness, and efficiency.\n", "date": "2024-06-24", "categories": []}
{"id": "2406.16846", "title": "Data Debiasing with Datamodels (D3M): Improving Subgroup Robustness via\n  Data Selection", "abstract": "  Machine learning models can fail on subgroups that are underrepresented\nduring training. While techniques such as dataset balancing can improve\nperformance on underperforming groups, they require access to training group\nannotations and can end up removing large portions of the dataset. In this\npaper, we introduce Data Debiasing with Datamodels (D3M), a debiasing approach\nwhich isolates and removes specific training examples that drive the model's\nfailures on minority groups. Our approach enables us to efficiently train\ndebiased classifiers while removing only a small number of examples, and does\nnot require training group annotations or additional hyperparameter tuning.\n", "date": "2024-06-24", "categories": []}
{"id": "2406.14993", "title": "Disability Representations: Finding Biases in Automatic Image Generation", "abstract": "  Recent advancements in image generation technology have enabled widespread\naccess to AI-generated imagery, prominently used in advertising, entertainment,\nand progressively in every form of visual content. However, these technologies\noften perpetuate societal biases. This study investigates the representation\nbiases in popular image generation models towards people with disabilities\n(PWD). Through a comprehensive experiment involving several popular\ntext-to-image models, we analyzed the depiction of disability. The results\nindicate a significant bias, with most generated images portraying disabled\nindividuals as old, sad, and predominantly using manual wheelchairs. These\nfindings highlight the urgent need for more inclusive AI development, ensuring\ndiverse and accurate representation of PWD in generated images. This research\nunderscores the importance of addressing and mitigating biases in AI models to\nfoster equitable and realistic representations.\n", "date": "2024-06-21", "categories": []}
{"id": "2308.16681", "title": "One Model Many Scores: Using Multiverse Analysis to Prevent Fairness\n  Hacking and Evaluate the Influence of Model Design Decisions", "abstract": "  A vast number of systems across the world use algorithmic decision making\n(ADM) to (partially) automate decisions that have previously been made by\nhumans. The downstream effects of ADM systems critically depend on the\ndecisions made during a systems' design, implementation, and evaluation, as\nbiases in data can be mitigated or reinforced along the modeling pipeline. Many\nof these decisions are made implicitly, without knowing exactly how they will\ninfluence the final system. To study this issue, we draw on insights from the\nfield of psychology and introduce the method of multiverse analysis for\nalgorithmic fairness. In our proposed method, we turn implicit decisions during\ndesign and evaluation into explicit ones and demonstrate their fairness\nimplications. By combining decisions, we create a grid of all possible\n\"universes\" of decision combinations. For each of these universes, we compute\nmetrics of fairness and performance. Using the resulting dataset, one can\ninvestigate the variability and robustness of fairness scores and see how and\nwhich decisions impact fairness. We demonstrate how multiverse analyses can be\nused to better understand fairness implications of design and evaluation\ndecisions using an exemplary case study of predicting public health care\ncoverage for vulnerable populations. Our results highlight how decisions\nregarding the evaluation of a system can lead to vastly different fairness\nmetrics for the same model. This is problematic, as a nefarious actor could\noptimise or \"hack\" a fairness metric to portray a discriminating model as fair\nmerely by changing how it is evaluated. We illustrate how a multiverse analysis\ncan help to address this issue.\n", "date": "2023-08-31", "categories": []}
{"id": "2311.11900", "title": "Measuring and Mitigating Biases in Motor Insurance Pricing", "abstract": "  The non-life insurance sector operates within a highly competitive and\ntightly regulated framework, confronting a pivotal juncture in the formulation\nof pricing strategies. Insurers are compelled to harness a range of statistical\nmethodologies and available data to construct optimal pricing structures that\nalign with the overarching corporate strategy while accommodating the dynamics\nof market competition. Given the fundamental societal role played by insurance,\npremium rates are subject to rigorous scrutiny by regulatory authorities. These\nrates must conform to principles of transparency, explainability, and ethical\nconsiderations. Consequently, the act of pricing transcends mere statistical\ncalculations and carries the weight of strategic and societal factors. These\nmultifaceted concerns may drive insurers to establish equitable premiums,\ntaking into account various variables. For instance, regulations mandate the\nprovision of equitable premiums, considering factors such as policyholder\ngender or mutualist group dynamics in accordance with respective corporate\nstrategies. Age-based premium fairness is also mandated. In certain insurance\ndomains, variables such as the presence of serious illnesses or disabilities\nare emerging as new dimensions for evaluating fairness. Regardless of the\nmotivating factor prompting an insurer to adopt fairer pricing strategies for a\nspecific variable, the insurer must possess the capability to define, measure,\nand ultimately mitigate any ethical biases inherent in its pricing practices\nwhile upholding standards of consistency and performance. This study seeks to\nprovide a comprehensive set of tools for these endeavors and assess their\neffectiveness through practical application in the context of automobile\ninsurance.\n", "date": "2023-11-20", "categories": []}
{"id": "2312.02592", "title": "FRAPPE: A Group Fairness Framework for Post-Processing Everything", "abstract": "  Despite achieving promising fairness-error trade-offs, in-processing\nmitigation techniques for group fairness cannot be employed in numerous\npractical applications with limited computation resources or no access to the\ntraining pipeline of the prediction model. In these situations, post-processing\nis a viable alternative. However, current methods are tailored to specific\nproblem settings and fairness definitions and hence, are not as broadly\napplicable as in-processing. In this work, we propose a framework that turns\nany regularized in-processing method into a post-processing approach. This\nprocedure prescribes a way to obtain post-processing techniques for a much\nbroader range of problem settings than the prior post-processing literature. We\nshow theoretically and through extensive experiments that our framework\npreserves the good fairness-error trade-offs achieved with in-processing and\ncan improve over the effectiveness of prior post-processing methods. Finally,\nwe demonstrate several advantages of a modular mitigation strategy that\ndisentangles the training of the prediction model from the fairness mitigation,\nincluding better performance on tasks with partial group labels.\n", "date": "2023-12-05", "categories": []}
{"id": "2312.04772", "title": "Remembering to Be Fair: Non-Markovian Fairness in Sequential Decision\n  Making", "abstract": "  Fair decision making has largely been studied with respect to a single\ndecision. Here we investigate the notion of fairness in the context of\nsequential decision making where multiple stakeholders can be affected by the\noutcomes of decisions. We observe that fairness often depends on the history of\nthe sequential decision-making process, and in this sense that it is inherently\nnon-Markovian. We further observe that fairness often needs to be assessed at\ntime points within the process, not just at the end of the process. To advance\nour understanding of this class of fairness problems, we explore the notion of\nnon-Markovian fairness in the context of sequential decision making. We\nidentify properties of non-Markovian fairness, including notions of long-term,\nanytime, periodic, and bounded fairness. We explore the interplay between\nnon-Markovian fairness and memory and how memory can support construction of\nfair policies. Finally, we introduce the FairQCM algorithm, which can\nautomatically augment its training data to improve sample efficiency in the\nsynthesis of fair policies via reinforcement learning.\n", "date": "2023-12-08", "categories": []}
{"id": "2402.18129", "title": "On the Inductive Biases of Demographic Parity-based Fair Learning\n  Algorithms", "abstract": "  Fair supervised learning algorithms assigning labels with little dependence\non a sensitive attribute have attracted great attention in the machine learning\ncommunity. While the demographic parity (DP) notion has been frequently used to\nmeasure a model's fairness in training fair classifiers, several studies in the\nliterature suggest potential impacts of enforcing DP in fair learning\nalgorithms. In this work, we analytically study the effect of standard DP-based\nregularization methods on the conditional distribution of the predicted label\ngiven the sensitive attribute. Our analysis shows that an imbalanced training\ndataset with a non-uniform distribution of the sensitive attribute could lead\nto a classification rule biased toward the sensitive attribute outcome holding\nthe majority of training data. To control such inductive biases in DP-based\nfair learning, we propose a sensitive attribute-based distributionally robust\noptimization (SA-DRO) method improving robustness against the marginal\ndistribution of the sensitive attribute. Finally, we present several numerical\nresults on the application of DP-based learning methods to standard centralized\nand distributed learning problems. The empirical findings support our\ntheoretical results on the inductive biases in DP-based fair learning\nalgorithms and the debiasing effects of the proposed SA-DRO method.\n", "date": "2024-02-28", "categories": []}
{"id": "2404.17293", "title": "Lazy Data Practices Harm Fairness Research", "abstract": "  Data practices shape research and practice on fairness in machine learning\n(fair ML). Critical data studies offer important reflections and critiques for\nthe responsible advancement of the field by highlighting shortcomings and\nproposing recommendations for improvement. In this work, we present a\ncomprehensive analysis of fair ML datasets, demonstrating how unreflective yet\ncommon practices hinder the reach and reliability of algorithmic fairness\nfindings. We systematically study protected information encoded in tabular\ndatasets and their usage in 280 experiments across 142 publications.\n  Our analyses identify three main areas of concern: (1) a \\textbf{lack of\nrepresentation for certain protected attributes} in both data and evaluations;\n(2) the widespread \\textbf{exclusion of minorities} during data preprocessing;\nand (3) \\textbf{opaque data processing} threatening the generalization of\nfairness research. By conducting exemplary analyses on the utilization of\nprominent datasets, we demonstrate how unreflective data decisions\ndisproportionately affect minority groups, fairness metrics, and resultant\nmodel comparisons. Additionally, we identify supplementary factors such as\nlimitations in publicly available data, privacy considerations, and a general\nlack of awareness, which exacerbate these challenges. To address these issues,\nwe propose a set of recommendations for data usage in fairness research\ncentered on transparency and responsible inclusion. This study underscores the\nneed for a critical reevaluation of data practices in fair ML and offers\ndirections to improve both the sourcing and usage of datasets.\n", "date": "2024-04-26", "categories": []}
{"id": "2406.12033", "title": "Unveiling and Mitigating Bias in Mental Health Analysis with Large\n  Language Models", "abstract": "  The advancement of large language models (LLMs) has demonstrated strong\ncapabilities across various applications, including mental health analysis.\nHowever, existing studies have focused on predictive performance, leaving the\ncritical issue of fairness underexplored, posing significant risks to\nvulnerable populations. Despite acknowledging potential biases, previous works\nhave lacked thorough investigations into these biases and their impacts. To\naddress this gap, we systematically evaluate biases across seven social factors\n(e.g., gender, age, religion) using ten LLMs with different prompting methods\non eight diverse mental health datasets. Our results show that GPT-4 achieves\nthe best overall balance in performance and fairness among LLMs, although it\nstill lags behind domain-specific models like MentalRoBERTa in some cases.\nAdditionally, our tailored fairness-aware prompts can effectively mitigate bias\nin mental health predictions, highlighting the great potential for fair\nanalysis in this field.\n", "date": "2024-06-17", "categories": []}
{"id": "2406.13681", "title": "On the Consistency of Fairness Measurement Methods for Regression Tasks", "abstract": "  With growing applications of Machine Learning (ML) techniques in the real\nworld, it is highly important to ensure that these models work in an equitable\nmanner. One main step in ensuring fairness is to effectively measure fairness,\nand to this end, various metrics have been proposed in the past literature.\nWhile the computation of those metrics are straightforward in the\nclassification set-up, it is computationally intractable in the regression\ndomain. To address the challenge of computational intractability, past\nliterature proposed various methods to approximate such metrics. However, they\ndid not verify the extent to which the output of such approximation algorithms\nare consistent with each other. To fill this gap, this paper comprehensively\nstudies the consistency of the output of various fairness measurement methods\nthrough conducting an extensive set of experiments on various regression tasks.\nAs a result, it finds that while some fairness measurement approaches show\nstrong consistency across various regression tasks, certain methods show a\nrelatively poor consistency in certain regression tasks. This, in turn, calls\nfor a more principled approach for measuring fairness in the regression domain.\n", "date": "2024-06-19", "categories": []}
{"id": "2406.13882", "title": "Allocation Requires Prediction Only if Inequality Is Low", "abstract": "  Algorithmic predictions are emerging as a promising solution concept for\nefficiently allocating societal resources. Fueling their use is an underlying\nassumption that such systems are necessary to identify individuals for\ninterventions. We propose a principled framework for assessing this assumption:\nUsing a simple mathematical model, we evaluate the efficacy of prediction-based\nallocations in settings where individuals belong to larger units such as\nhospitals, neighborhoods, or schools. We find that prediction-based allocations\noutperform baseline methods using aggregate unit-level statistics only when\nbetween-unit inequality is low and the intervention budget is high. Our results\nhold for a wide range of settings for the price of prediction, treatment effect\nheterogeneity, and unit-level statistics' learnability. Combined, we highlight\nthe potential limits to improving the efficacy of interventions through\nprediction.\n", "date": "2024-06-19", "categories": []}
{"id": "2406.14154", "title": "Watching the Watchers: A Comparative Fairness Audit of Cloud-based\n  Content Moderation Services", "abstract": "  Online platforms face the challenge of moderating an ever-increasing volume\nof content, including harmful hate speech. In the absence of clear legal\ndefinitions and a lack of transparency regarding the role of algorithms in\nshaping decisions on content moderation, there is a critical need for external\naccountability. Our study contributes to filling this gap by systematically\nevaluating four leading cloud-based content moderation services through a\nthird-party audit, highlighting issues such as biases against minorities and\nvulnerable groups that may arise through over-reliance on these services. Using\na black-box audit approach and four benchmark data sets, we measure performance\nin explicit and implicit hate speech detection as well as counterfactual\nfairness through perturbation sensitivity analysis and present disparities in\nperformance for certain target identity groups and data sets. Our analysis\nreveals that all services had difficulties detecting implicit hate speech,\nwhich relies on more subtle and codified messages. Moreover, our results point\nto the need to remove group-specific bias. It seems that biases towards some\ngroups, such as Women, have been mostly rectified, while biases towards other\ngroups, such as LGBTQ+ and PoC remain.\n", "date": "2024-06-20", "categories": []}
{"id": "2406.14401", "title": "Fair Streaming Feature Selection", "abstract": "  Streaming feature selection techniques have become essential in processing\nreal-time data streams, as they facilitate the identification of the most\nrelevant attributes from continuously updating information. Despite their\nperformance, current algorithms to streaming feature selection frequently fall\nshort in managing biases and avoiding discrimination that could be perpetuated\nby sensitive attributes, potentially leading to unfair outcomes in the\nresulting models. To address this issue, we propose FairSFS, a novel algorithm\nfor Fair Streaming Feature Selection, to uphold fairness in the feature\nselection process without compromising the ability to handle data in an online\nmanner. FairSFS adapts to incoming feature vectors by dynamically adjusting the\nfeature set and discerns the correlations between classification attributes and\nsensitive attributes from this revised set, thereby forestalling the\npropagation of sensitive data. Empirical evaluations show that FairSFS not only\nmaintains accuracy that is on par with leading streaming feature selection\nmethods and existing fair feature techniques but also significantly improves\nfairness metrics.\n", "date": "2024-06-20", "categories": []}
{"id": "2306.15083", "title": "Balanced Filtering via Disclosure-Controlled Proxies", "abstract": "  We study the problem of collecting a cohort or set that is balanced with\nrespect to sensitive groups when group membership is unavailable or prohibited\nfrom use at deployment time. Specifically, our deployment-time collection\nmechanism does not reveal significantly more about the group membership of any\nindividual sample than can be ascertained from base rates alone. To do this, we\nstudy a learner that can use a small set of labeled data to train a proxy\nfunction that can later be used for this filtering or selection task. We then\nassociate the range of the proxy function with sampling probabilities; given a\nnew example, we classify it using our proxy function and then select it with\nprobability corresponding to its proxy classification. Importantly, we require\nthat the proxy classification does not reveal significantly more information\nabout the sensitive group membership of any individual example compared to\npopulation base rates alone (i.e., the level of disclosure should be\ncontrolled) and show that we can find such a proxy in a sample- and\noracle-efficient manner. Finally, we experimentally evaluate our algorithm and\nanalyze its generalization properties.\n", "date": "2023-06-26", "categories": []}
{"id": "2403.06017", "title": "Addressing Shortcomings in Fair Graph Learning Datasets: Towards a New\n  Benchmark", "abstract": "  Fair graph learning plays a pivotal role in numerous practical applications.\nRecently, many fair graph learning methods have been proposed; however, their\nevaluation often relies on poorly constructed semi-synthetic datasets or\nsubstandard real-world datasets. In such cases, even a basic Multilayer\nPerceptron (MLP) can outperform Graph Neural Networks (GNNs) in both utility\nand fairness. In this work, we illustrate that many datasets fail to provide\nmeaningful information in the edges, which may challenge the necessity of using\ngraph structures in these problems. To address these issues, we develop and\nintroduce a collection of synthetic, semi-synthetic, and real-world datasets\nthat fulfill a broad spectrum of requirements. These datasets are thoughtfully\ndesigned to include relevant graph structures and bias information crucial for\nthe fair evaluation of models. The proposed synthetic and semi-synthetic\ndatasets offer the flexibility to create data with controllable bias\nparameters, thereby enabling the generation of desired datasets with\nuser-defined bias values with ease. Moreover, we conduct systematic evaluations\nof these proposed datasets and establish a unified evaluation approach for fair\ngraph learning models. Our extensive experimental results with fair graph\nlearning methods across our datasets demonstrate their effectiveness in\nbenchmarking the performance of these methods. Our datasets and the code for\nreproducing our experiments are available at\nhttps://github.com/XweiQ/Benchmark-GraphFairness.\n", "date": "2024-03-09", "categories": []}
{"id": "2405.09360", "title": "The Unfairness of $\\varepsilon$-Fairness", "abstract": "  Fairness in decision-making processes is often quantified using probabilistic\nmetrics. However, these metrics may not fully capture the real-world\nconsequences of unfairness. In this article, we adopt a utility-based approach\nto more accurately measure the real-world impacts of decision-making process.\nIn particular, we show that if the concept of $\\varepsilon$-fairness is\nemployed, it can possibly lead to outcomes that are maximally unfair in the\nreal-world context. Additionally, we address the common issue of unavailable\ndata on false negatives by proposing a reduced setting that still captures\nessential fairness considerations. We illustrate our findings with two\nreal-world examples: college admissions and credit risk assessment. Our\nanalysis reveals that while traditional probability-based evaluations might\nsuggest fairness, a utility-based approach uncovers the necessary actions to\ntruly achieve equality. For instance, in the college admission case, we find\nthat enhancing completion rates is crucial for ensuring fairness. Summarizing,\nthis paper highlights the importance of considering the real-world context when\nevaluating fairness.\n", "date": "2024-05-15", "categories": []}
{"id": "2406.08819", "title": "AIM: Attributing, Interpreting, Mitigating Data Unfairness", "abstract": "  Data collected in the real world often encapsulates historical discrimination\nagainst disadvantaged groups and individuals. Existing fair machine learning\n(FairML) research has predominantly focused on mitigating discriminative bias\nin the model prediction, with far less effort dedicated towards exploring how\nto trace biases present in the data, despite its importance for the\ntransparency and interpretability of FairML. To fill this gap, we investigate a\nnovel research problem: discovering samples that reflect biases/prejudices from\nthe training data. Grounding on the existing fairness notions, we lay out a\nsample bias criterion and propose practical algorithms for measuring and\ncountering sample bias. The derived bias score provides intuitive sample-level\nattribution and explanation of historical bias in data. On this basis, we\nfurther design two FairML strategies via sample-bias-informed minimal data\nediting. They can mitigate both group and individual unfairness at the cost of\nminimal or zero predictive utility loss. Extensive experiments and analyses on\nmultiple real-world datasets demonstrate the effectiveness of our methods in\nexplaining and mitigating unfairness. Code is available at\nhttps://github.com/ZhiningLiu1998/AIM.\n", "date": "2024-06-13", "categories": []}
{"id": "2406.12011", "title": "The Benefits and Risks of Transductive Approaches for AI Fairness", "abstract": "  Recently, transductive learning methods, which leverage holdout sets during\ntraining, have gained popularity for their potential to improve speed,\naccuracy, and fairness in machine learning models. Despite this, the\ncomposition of the holdout set itself, particularly the balance of sensitive\nsub-groups, has been largely overlooked. Our experiments on CIFAR and CelebA\ndatasets show that compositional changes in the holdout set can substantially\ninfluence fairness metrics. Imbalanced holdout sets exacerbate existing\ndisparities, while balanced holdouts can mitigate issues introduced by\nimbalanced training data. These findings underline the necessity of\nconstructing holdout sets that are both diverse and representative.\n", "date": "2024-06-17", "categories": []}
{"id": "2406.10262", "title": "Fast solution to the fair ranking problem using the Sinkhorn algorithm", "abstract": "  In two-sided marketplaces such as online flea markets, recommender systems\nfor providing consumers with personalized item rankings play a key role in\npromoting transactions between providers and consumers. Meanwhile, two-sided\nmarketplaces face the problem of balancing consumer satisfaction and fairness\namong items to stimulate activity of item providers. Saito and Joachims (2022)\ndevised an impact-based fair ranking method for maximizing the Nash social\nwelfare based on fair division; however, this method, which requires solving a\nlarge-scale constrained nonlinear optimization problem, is very difficult to\napply to practical-scale recommender systems. We thus propose a fast solution\nto the impact-based fair ranking problem. We first transform the fair ranking\nproblem into an unconstrained optimization problem and then design a gradient\nascent method that repeatedly executes the Sinkhorn algorithm. Experimental\nresults demonstrate that our algorithm provides fair rankings of high quality\nand is about 1000 times faster than application of commercial optimization\nsoftware.\n", "date": "2024-06-11", "categories": []}
{"id": "2406.10486", "title": "Do Large Language Models Discriminate in Hiring Decisions on the Basis\n  of Race, Ethnicity, and Gender?", "abstract": "  We examine whether large language models (LLMs) exhibit race- and\ngender-based name discrimination in hiring decisions, similar to classic\nfindings in the social sciences (Bertrand and Mullainathan, 2004). We design a\nseries of templatic prompts to LLMs to write an email to a named job applicant\ninforming them of a hiring decision. By manipulating the applicant's first\nname, we measure the effect of perceived race, ethnicity, and gender on the\nprobability that the LLM generates an acceptance or rejection email. We find\nthat the hiring decisions of LLMs in many settings are more likely to favor\nWhite applicants over Hispanic applicants. In aggregate, the groups with the\nhighest and lowest acceptance rates respectively are masculine White names and\nmasculine Hispanic names. However, the comparative acceptance rates by group\nvary under different templatic settings, suggesting that LLMs' race- and\ngender-sensitivity may be idiosyncratic and prompt-sensitive.\n", "date": "2024-06-15", "categories": []}
{"id": "2402.19226", "title": "Investigating Gender Fairness in Machine Learning-driven Personalized\n  Care for Chronic Pain", "abstract": "  Chronic pain significantly diminishes the quality of life for millions\nworldwide. While psychoeducation and therapy can improve pain outcomes, many\nindividuals experiencing pain lack access to evidence-based treatments or fail\nto complete the necessary number of sessions to achieve benefit. Reinforcement\nlearning (RL) shows potential in tailoring personalized pain management\ninterventions according to patients' individual needs while ensuring the\nefficient use of scarce clinical resources. However, clinicians, patients, and\nhealthcare decision-makers are concerned that RL solutions could exacerbate\ndisparities associated with patient characteristics like race or gender. In\nthis article, we study gender fairness in personalized pain care\nrecommendations using a real-world application of reinforcement learning\n(Piette et al., 2022a). Here, adhering to gender fairness translates to minimal\nor no disparity in the utility received by subpopulations as defined by gender.\nWe investigate whether the selection of relevant patient information (referred\nto as features) used to assist decision-making affects gender fairness. Our\nexperiments, conducted using real-world data Piette, 2022), indicate that\nincluded features can impact gender fairness. Moreover, we propose an RL\nsolution, NestedRecommendation, that demonstrates the ability: i) to adaptively\nlearn to select the features that optimize for utility and fairness, and ii) to\naccelerate feature selection and in turn, improve pain care recommendations\nfrom early on, by leveraging clinicians' domain expertise.\n", "date": "2024-02-29", "categories": []}
{"id": "2406.08183", "title": "Underneath the Numbers: Quantitative and Qualitative Gender Fairness in\n  LLMs for Depression Prediction", "abstract": "  Recent studies show bias in many machine learning models for depression\ndetection, but bias in LLMs for this task remains unexplored. This work\npresents the first attempt to investigate the degree of gender bias present in\nexisting LLMs (ChatGPT, LLaMA 2, and Bard) using both quantitative and\nqualitative approaches. From our quantitative evaluation, we found that ChatGPT\nperforms the best across various performance metrics and LLaMA 2 outperforms\nother LLMs in terms of group fairness metrics. As qualitative fairness\nevaluation remains an open research question we propose several strategies\n(e.g., word count, thematic analysis) to investigate whether and how a\nqualitative evaluation can provide valuable insights for bias analysis beyond\nwhat is possible with quantitative evaluation. We found that ChatGPT\nconsistently provides a more comprehensive, well-reasoned explanation for its\nprediction compared to LLaMA 2. We have also identified several themes adopted\nby LLMs to qualitatively evaluate gender fairness. We hope our results can be\nused as a stepping stone towards future attempts at improving qualitative\nevaluation of fairness for LLMs especially for high-stakes tasks such as\ndepression detection.\n", "date": "2024-06-12", "categories": []}
{"id": "2406.09513", "title": "Fair GLASSO: Estimating Fair Graphical Models with Unbiased Statistical\n  Behavior", "abstract": "  We propose estimating Gaussian graphical models (GGMs) that are fair with\nrespect to sensitive nodal attributes. Many real-world models exhibit unfair\ndiscriminatory behavior due to biases in data. Such discrimination is known to\nbe exacerbated when data is equipped with pairwise relationships encoded in a\ngraph. Additionally, the effect of biased data on graphical models is largely\nunderexplored. We thus introduce fairness for graphical models in the form of\ntwo bias metrics to promote balance in statistical similarities across nodal\ngroups with different sensitive attributes. Leveraging these metrics, we\npresent Fair GLASSO, a regularized graphical lasso approach to obtain sparse\nGaussian precision matrices with unbiased statistical dependencies across\ngroups. We also propose an efficient proximal gradient algorithm to obtain the\nestimates. Theoretically, we express the tradeoff between fair and accurate\nestimated precision matrices. Critically, this includes demonstrating when\naccuracy can be preserved in the presence of a fairness regularizer. On top of\nthis, we study the complexity of Fair GLASSO and demonstrate that our algorithm\nenjoys a fast convergence rate. Our empirical validation includes synthetic and\nreal-world simulations that illustrate the value and effectiveness of our\nproposed optimization problem and iterative algorithm.\n", "date": "2024-06-13", "categories": []}
{"id": "2406.09561", "title": "Label Noise Robustness for Domain-Agnostic Fair Corrections via Nearest\n  Neighbors Label Spreading", "abstract": "  Last-layer retraining methods have emerged as an efficient framework for\ncorrecting existing base models. Within this framework, several methods have\nbeen proposed to deal with correcting models for subgroup fairness with and\nwithout group membership information. Importantly, prior work has demonstrated\nthat many methods are susceptible to noisy labels. To this end, we propose a\ndrop-in correction for label noise in last-layer retraining, and demonstrate\nthat it achieves state-of-the-art worst-group accuracy for a broad range of\nsymmetric label noise and across a wide variety of datasets exhibiting spurious\ncorrelations. Our proposed approach uses label spreading on a latent nearest\nneighbors graph and has minimal computational overhead compared to existing\nmethods.\n", "date": "2024-06-13", "categories": []}
{"id": "2406.09977", "title": "Disentangling Dialect from Social Bias via Multitask Learning to Improve\n  Fairness", "abstract": "  Dialects introduce syntactic and lexical variations in language that occur in\nregional or social groups. Most NLP methods are not sensitive to such\nvariations. This may lead to unfair behavior of the methods, conveying negative\nbias towards dialect speakers. While previous work has studied dialect-related\nfairness for aspects like hate speech, other aspects of biased language, such\nas lewdness, remain fully unexplored. To fill this gap, we investigate\nperformance disparities between dialects in the detection of five aspects of\nbiased language and how to mitigate them. To alleviate bias, we present a\nmultitask learning approach that models dialect language as an auxiliary task\nto incorporate syntactic and lexical variations. In our experiments with\nAfrican-American English dialect, we provide empirical evidence that\ncomplementing common learning approaches with dialect modeling improves their\nfairness. Furthermore, the results suggest that multitask learning achieves\nstate-of-the-art performance and helps to detect properties of biased language\nmore reliably.\n", "date": "2024-06-14", "categories": []}
{"id": "2406.04382", "title": "Improving the Fairness of Deep-Learning, Short-term Crime Prediction\n  with Under-reporting-aware Models", "abstract": "  Deep learning crime predictive tools use past crime data and additional\nbehavioral datasets to forecast future crimes. Nevertheless, these tools have\nbeen shown to suffer from unfair predictions across minority racial and ethnic\ngroups. Current approaches to address this unfairness generally propose either\npre-processing methods that mitigate the bias in the training datasets by\napplying corrections to crime counts based on domain knowledge or in-processing\nmethods that are implemented as fairness regularizers to optimize for both\naccuracy and fairness. In this paper, we propose a novel deep learning\narchitecture that combines the power of these two approaches to increase\nprediction fairness. Our results show that the proposed model improves the\nfairness of crime predictions when compared to models with in-processing\nde-biasing approaches and with models without any type of bias correction,\nalbeit at the cost of reducing accuracy.\n", "date": "2024-06-06", "categories": []}
{"id": "2404.00022", "title": "Analysing and Organising Human Communications for AI Fairness-Related\n  Decisions: Use Cases from the Public Sector", "abstract": "  AI algorithms used in the public sector, e.g., for allocating social benefits\nor predicting fraud, often involve multiple public and private stakeholders at\nvarious phases of the algorithm's life-cycle. Communication issues between\nthese diverse stakeholders can lead to misinterpretation and misuse of\nalgorithms. We investigate the communication processes for AI fairness-related\ndecisions by conducting interviews with practitioners working on algorithmic\nsystems in the public sector. By applying qualitative coding analysis, we\nidentify key elements of communication processes that underlie fairness-related\nhuman decisions. We analyze the division of roles, tasks, skills, and\nchallenges perceived by stakeholders. We formalize the underlying communication\nissues within a conceptual framework that i. represents the communication\npatterns ii. outlines missing elements, such as actors who miss skills for\ntheir tasks. The framework is used for describing and analyzing key\norganizational issues for fairness-related decisions. Three general patterns\nemerge from the analysis: 1. Policy-makers, civil servants, and domain experts\nare less involved compared to developers throughout a system's life-cycle. This\nleads to developers taking on extra roles such as advisor, while they\npotentially miss the required skills and guidance from domain experts. 2.\nEnd-users and policy-makers often lack the technical skills to interpret a\nsystem's limitations, and rely on developer roles for making decisions\nconcerning fairness issues. 3. Citizens are structurally absent throughout a\nsystem's life-cycle, which may lead to decisions that do not include relevant\nconsiderations from impacted stakeholders.\n", "date": "2024-03-20", "categories": []}
{"id": "2303.01928", "title": "Towards Algorithmic Fairness by means of Instance-level Data\n  Re-weighting based on Shapley Values", "abstract": "  Algorithmic fairness is of utmost societal importance, yet state-of-the-art\nlarge-scale machine learning models require training with massive datasets that\nare frequently biased. In this context, pre-processing methods that focus on\nmodeling and correcting bias in the data emerge as valuable approaches. In this\npaper, we propose FairShap, a novel instance-level data re-weighting method for\nfair algorithmic decision-making through data valuation by means of Shapley\nValues. FairShap is model-agnostic and easily interpretable. It measures the\ncontribution of each training data point to a predefined fairness metric. We\nempirically validate FairShap on several state-of-the-art datasets of different\nnature, with a variety of training scenarios and machine learning models and\nshow how it yields fairer models with similar levels of accuracy than the\nbaselines. We illustrate FairShap's interpretability by means of histograms and\nlatent space visualizations. Moreover, we perform a utility-fairness study and\nanalyze FairShap's computational cost depending on the size of the dataset and\nthe number of features. We believe that FairShap represents a novel\ncontribution in interpretable and model-agnostic approaches to algorithmic\nfairness that yields competitive accuracy even when only biased training\ndatasets are available.\n", "date": "2023-03-03", "categories": []}
{"id": "2306.09468", "title": "FFB: A Fair Fairness Benchmark for In-Processing Group Fairness Methods", "abstract": "  This paper introduces the Fair Fairness Benchmark (\\textsf{FFB}), a\nbenchmarking framework for in-processing group fairness methods. Ensuring\nfairness in machine learning is important for ethical compliance. However,\nthere exist challenges in comparing and developing fairness methods due to\ninconsistencies in experimental settings, lack of accessible algorithmic\nimplementations, and limited extensibility of current fairness packages and\ntools. To address these issues, we introduce an open-source standardized\nbenchmark for evaluating in-processing group fairness methods and provide a\ncomprehensive analysis of state-of-the-art methods to ensure different notions\nof group fairness. This work offers the following key contributions: the\nprovision of flexible, extensible, minimalistic, and research-oriented\nopen-source code; the establishment of unified fairness method benchmarking\npipelines; and extensive benchmarking, which yields key insights from\n$\\mathbf{45,079}$ experiments, $\\mathbf{14,428}$ GPU hours. We believe that our\nwork will significantly facilitate the growth and development of the fairness\nresearch community.\n", "date": "2023-06-15", "categories": []}
{"id": "2311.01349", "title": "Post-hoc Orthogonalization for Mitigation of Protected Feature Bias in\n  CXR Embeddings", "abstract": "  Purpose: To analyze and remove protected feature effects in chest radiograph\nembeddings of deep learning models. Methods: An orthogonalization is utilized\nto remove the influence of protected features (e.g., age, sex, race) in CXR\nembeddings, ensuring feature-independent results. To validate the efficacy of\nthe approach, we retrospectively study the MIMIC and CheXpert datasets using\nthree pre-trained models, namely a supervised contrastive, a self-supervised\ncontrastive, and a baseline classifier model. Our statistical analysis involves\ncomparing the original versus the orthogonalized embeddings by estimating\nprotected feature influences and evaluating the ability to predict race, age,\nor sex using the two types of embeddings. Results: Our experiments reveal a\nsignificant influence of protected features on predictions of pathologies.\nApplying orthogonalization removes these feature effects. Apart from removing\nany influence on pathology classification, while maintaining competitive\npredictive performance, orthogonalized embeddings further make it infeasible to\ndirectly predict protected attributes and mitigate subgroup disparities.\nConclusion: The presented work demonstrates the successful application and\nevaluation of the orthogonalization technique in the domain of chest X-ray\nimage classification.\n", "date": "2023-11-02", "categories": []}
{"id": "2406.00046", "title": "Hate Speech Detection with Generalizable Target-aware Fairness", "abstract": "  To counter the side effect brought by the proliferation of social media\nplatforms, hate speech detection (HSD) plays a vital role in halting the\ndissemination of toxic online posts at an early stage. However, given the\nubiquitous topical communities on social media, a trained HSD classifier easily\nbecomes biased towards specific targeted groups (e.g., female and black\npeople), where a high rate of false positive/negative results can significantly\nimpair public trust in the fairness of content moderation mechanisms, and\neventually harm the diversity of online society. Although existing\nfairness-aware HSD methods can smooth out some discrepancies across targeted\ngroups, they are mostly specific to a narrow selection of targets that are\nassumed to be known and fixed. This inevitably prevents those methods from\ngeneralizing to real-world use cases where new targeted groups constantly\nemerge over time. To tackle this defect, we propose Generalizable target-aware\nFairness (GetFair), a new method for fairly classifying each post that contains\ndiverse and even unseen targets during inference. To remove the HSD\nclassifier's spurious dependence on target-related features, GetFair trains a\nseries of filter functions in an adversarial pipeline, so as to deceive the\ndiscriminator that recovers the targeted group from filtered post embeddings.\nTo maintain scalability and generalizability, we innovatively parameterize all\nfilter functions via a hypernetwork that is regularized by the semantic\naffinity among targets. Taking a target's pretrained word embedding as input,\nthe hypernetwork generates the weights used by each target-specific filter\non-the-fly without storing dedicated filter parameters. Finally, comparative\nexperiments on two HSD datasets have shown advantageous performance of GetFair\non out-of-sample targets.\n", "date": "2024-05-28", "categories": []}
{"id": "2406.07029", "title": "Fairness-Aware Meta-Learning via Nash Bargaining", "abstract": "  To address issues of group-level fairness in machine learning, it is natural\nto adjust model parameters based on specific fairness objectives over a\nsensitive-attributed validation set. Such an adjustment procedure can be cast\nwithin a meta-learning framework. However, naive integration of fairness goals\nvia meta-learning can cause hypergradient conflicts for subgroups, resulting in\nunstable convergence and compromising model performance and fairness. To\nnavigate this issue, we frame the resolution of hypergradient conflicts as a\nmulti-player cooperative bargaining game. We introduce a two-stage\nmeta-learning framework in which the first stage involves the use of a Nash\nBargaining Solution (NBS) to resolve hypergradient conflicts and steer the\nmodel toward the Pareto front, and the second stage optimizes with respect to\nspecific fairness goals. Our method is supported by theoretical results,\nnotably a proof of the NBS for gradient aggregation free from linear\nindependence assumptions, a proof of Pareto improvement, and a proof of\nmonotonic improvement in validation loss. We also show empirical effects across\nvarious fairness objectives in six key fairness datasets and two image\nclassification tasks.\n", "date": "2024-06-11", "categories": []}
{"id": "2406.03314", "title": "Reproducibility study of FairAC", "abstract": "  This work aims to reproduce the findings of the paper \"Fair Attribute\nCompletion on Graph with Missing Attributes\" written by Guo, Chu, and Li\narXiv:2302.12977 by investigating the claims made in the paper. This paper\nsuggests that the results of the original paper are reproducible and thus, the\nclaims hold. However, the claim that FairAC is a generic framework for many\ndownstream tasks is very broad and could therefore only be partially tested.\nMoreover, we show that FairAC is generalizable to various datasets and\nsensitive attributes and show evidence that the improvement in group fairness\nof the FairAC framework does not come at the expense of individual fairness.\nLastly, the codebase of FairAC has been refactored and is now easily applicable\nfor various datasets and models.\n", "date": "2024-06-05", "categories": []}
{"id": "2406.05902", "title": "Whose Preferences? Differences in Fairness Preferences and Their Impact\n  on the Fairness of AI Utilizing Human Feedback", "abstract": "  There is a growing body of work on learning from human feedback to align\nvarious aspects of machine learning systems with human values and preferences.\nWe consider the setting of fairness in content moderation, in which human\nfeedback is used to determine how two comments -- referencing different\nsensitive attribute groups -- should be treated in comparison to one another.\nWith a novel dataset collected from Prolific and MTurk, we find significant\ngaps in fairness preferences depending on the race, age, political stance,\neducational level, and LGBTQ+ identity of annotators. We also demonstrate that\ndemographics mentioned in text have a strong influence on how users perceive\nindividual fairness in moderation. Further, we find that differences also exist\nin downstream classifiers trained to predict human preferences. Finally, we\nobserve that an ensemble, giving equal weight to classifiers trained on\nannotations from different demographics, performs better for different\ndemographic intersections; compared to a single classifier that gives equal\nweight to each annotation.\n", "date": "2024-06-09", "categories": []}
{"id": "2406.05918", "title": "Why Don't Prompt-Based Fairness Metrics Correlate?", "abstract": "  The widespread use of large language models has brought up essential\nquestions about the potential biases these models might learn. This led to the\ndevelopment of several metrics aimed at evaluating and mitigating these biases.\nIn this paper, we first demonstrate that prompt-based fairness metrics exhibit\npoor agreement, as measured by correlation, raising important questions about\nthe reliability of fairness assessment using prompts. Then, we outline six\nrelevant reasons why such a low correlation is observed across existing\nmetrics. Based on these insights, we propose a method called Correlated\nFairness Output (CAIRO) to enhance the correlation between fairness metrics.\nCAIRO augments the original prompts of a given fairness metric by using several\npre-trained language models and then selects the combination of the augmented\nprompts that achieves the highest correlation across metrics. We show a\nsignificant improvement in Pearson correlation from 0.3 and 0.18 to 0.90 and\n0.98 across metrics for gender and religion biases, respectively. Our code is\navailable at https://github.com/chandar-lab/CAIRO.\n", "date": "2024-06-09", "categories": []}
{"id": "2406.06134", "title": "DiffInject: Revisiting Debias via Synthetic Data Generation using\n  Diffusion-based Style Injection", "abstract": "  Dataset bias is a significant challenge in machine learning, where specific\nattributes, such as texture or color of the images are unintentionally learned\nresulting in detrimental performance. To address this, previous efforts have\nfocused on debiasing models either by developing novel debiasing algorithms or\nby generating synthetic data to mitigate the prevalent dataset biases. However,\ngenerative approaches to date have largely relied on using bias-specific\nsamples from the dataset, which are typically too scarce. In this work, we\npropose, DiffInject, a straightforward yet powerful method to augment synthetic\nbias-conflict samples using a pretrained diffusion model. This approach\nsignificantly advances the use of diffusion models for debiasing purposes by\nmanipulating the latent space. Our framework does not require any explicit\nknowledge of the bias types or labelling, making it a fully unsupervised\nsetting for debiasing. Our methodology demonstrates substantial result in\neffectively reducing dataset bias.\n", "date": "2024-06-10", "categories": []}
{"id": "2406.06352", "title": "Latent Directions: A Simple Pathway to Bias Mitigation in Generative AI", "abstract": "  Mitigating biases in generative AI and, particularly in text-to-image models,\nis of high importance given their growing implications in society. The biased\ndatasets used for training pose challenges in ensuring the responsible\ndevelopment of these models, and mitigation through hard prompting or embedding\nalteration, are the most common present solutions. Our work introduces a novel\napproach to achieve diverse and inclusive synthetic images by learning a\ndirection in the latent space and solely modifying the initial Gaussian noise\nprovided for the diffusion process. Maintaining a neutral prompt and untouched\nembeddings, this approach successfully adapts to diverse debiasing scenarios,\nsuch as geographical biases. Moreover, our work proves it is possible to\nlinearly combine these learned latent directions to introduce new mitigations,\nand if desired, integrate it with text embedding adjustments. Furthermore,\ntext-to-image models lack transparency for assessing bias in outputs, unless\nvisually inspected. Thus, we provide a tool to empower developers to select\ntheir desired concepts to mitigate. The project page with code is available\nonline.\n", "date": "2024-06-10", "categories": []}
{"id": "2405.20790", "title": "Intersectional Unfairness Discovery", "abstract": "  AI systems have been shown to produce unfair results for certain subgroups of\npopulation, highlighting the need to understand bias on certain sensitive\nattributes. Current research often falls short, primarily focusing on the\nsubgroups characterized by a single sensitive attribute, while neglecting the\nnature of intersectional fairness of multiple sensitive attributes. This paper\nfocuses on its one fundamental aspect by discovering diverse high-bias\nsubgroups under intersectional sensitive attributes. Specifically, we propose a\nBias-Guided Generative Network (BGGN). By treating each bias value as a reward,\nBGGN efficiently generates high-bias intersectional sensitive attributes.\nExperiments on real-world text and image datasets demonstrate a diverse and\nefficient discovery of BGGN. To further evaluate the generated unseen but\npossible unfair intersectional sensitive attributes, we formulate them as\nprompts and use modern generative AI to produce new texts and images. The\nresults of frequently generating biased data provides new insights of\ndiscovering potential unfairness in popular modern generative AI systems.\nWarning: This paper contains generative examples that are offensive in nature.\n", "date": "2024-05-31", "categories": []}
{"id": "2309.17417", "title": "Networked Inequality: Preferential Attachment Bias in Graph Neural\n  Network Link Prediction", "abstract": "  Graph neural network (GNN) link prediction is increasingly deployed in\ncitation, collaboration, and online social networks to recommend academic\nliterature, collaborators, and friends. While prior research has investigated\nthe dyadic fairness of GNN link prediction, the within-group (e.g., queer\nwomen) fairness and \"rich get richer\" dynamics of link prediction remain\nunderexplored. However, these aspects have significant consequences for degree\nand power imbalances in networks. In this paper, we shed light on how degree\nbias in networks affects Graph Convolutional Network (GCN) link prediction. In\nparticular, we theoretically uncover that GCNs with a symmetric normalized\ngraph filter have a within-group preferential attachment bias. We validate our\ntheoretical analysis on real-world citation, collaboration, and online social\nnetworks. We further bridge GCN's preferential attachment bias with unfairness\nin link prediction and propose a new within-group fairness metric. This metric\nquantifies disparities in link prediction scores within social groups, towards\ncombating the amplification of degree and power disparities. Finally, we\npropose a simple training-time strategy to alleviate within-group unfairness,\nand we show that it is effective on citation, social, and credit networks.\n", "date": "2023-09-29", "categories": []}
{"id": "2406.02889", "title": "Language-guided Detection and Mitigation of Unknown Dataset Bias", "abstract": "  Dataset bias is a significant problem in training fair classifiers. When\nattributes unrelated to classification exhibit strong biases towards certain\nclasses, classifiers trained on such dataset may overfit to these bias\nattributes, substantially reducing the accuracy for minority groups. Mitigation\ntechniques can be categorized according to the availability of bias information\n(\\ie, prior knowledge). Although scenarios with unknown biases are better\nsuited for real-world settings, previous work in this field often suffers from\na lack of interpretability regarding biases and lower performance. In this\nstudy, we propose a framework to identify potential biases as keywords without\nprior knowledge based on the partial occurrence in the captions. We further\npropose two debiasing methods: (a) handing over to an existing debiasing\napproach which requires prior knowledge by assigning pseudo-labels, and (b)\nemploying data augmentation via text-to-image generative models, using acquired\nbias keywords as prompts. Despite its simplicity, experimental results show\nthat our framework not only outperforms existing methods without prior\nknowledge, but also is even comparable with a method that assumes prior\nknowledge.\n", "date": "2024-06-05", "categories": []}
{"id": "2406.03064", "title": "Path-Specific Causal Reasoning for Fairness-aware Cognitive Diagnosis", "abstract": "  Cognitive Diagnosis~(CD), which leverages students and exercise data to\npredict students' proficiency levels on different knowledge concepts, is one of\nfundamental components in Intelligent Education. Due to the scarcity of\nstudent-exercise interaction data, most existing methods focus on making the\nbest use of available data, such as exercise content and student\ninformation~(e.g., educational context). Despite the great progress, the abuse\nof student sensitive information has not been paid enough attention. Due to the\nimportant position of CD in Intelligent Education, employing sensitive\ninformation when making diagnosis predictions will cause serious social issues.\nMoreover, data-driven neural networks are easily misled by the shortcut between\ninput data and output prediction, exacerbating this problem. Therefore, it is\ncrucial to eliminate the negative impact of sensitive information in CD models.\nIn response, we argue that sensitive attributes of students can also provide\nuseful information, and only the shortcuts directly related to the sensitive\ninformation should be eliminated from the diagnosis process. Thus, we employ\ncausal reasoning and design a novel Path-Specific Causal Reasoning Framework\n(PSCRF) to achieve this goal. Specifically, we first leverage an encoder to\nextract features and generate embeddings for general information and sensitive\ninformation of students. Then, we design a novel attribute-oriented predictor\nto decouple the sensitive attributes, in which fairness-related sensitive\nfeatures will be eliminated and other useful information will be retained.\nFinally, we designed a multi-factor constraint to ensure the performance of\nfairness and diagnosis performance simultaneously. Extensive experiments over\nreal-world datasets (e.g., PISA dataset) demonstrate the effectiveness of our\nproposed PSCRF.\n", "date": "2024-06-05", "categories": []}
{"id": "2406.03255", "title": "On the Maximal Local Disparity of Fairness-Aware Classifiers", "abstract": "  Fairness has become a crucial aspect in the development of trustworthy\nmachine learning algorithms. Current fairness metrics to measure the violation\nof demographic parity have the following drawbacks: (i) the average difference\nof model predictions on two groups cannot reflect their distribution disparity,\nand (ii) the overall calculation along all possible predictions conceals the\nextreme local disparity at or around certain predictions. In this work, we\npropose a novel fairness metric called Maximal Cumulative ratio Disparity along\nvarying Predictions' neighborhood (MCDP), for measuring the maximal local\ndisparity of the fairness-aware classifiers. To accurately and efficiently\ncalculate the MCDP, we develop a provably exact and an approximate calculation\nalgorithm that greatly reduces the computational complexity with low estimation\nerror. We further propose a bi-level optimization algorithm using a\ndifferentiable approximation of the MCDP for improving the algorithmic\nfairness. Extensive experiments on both tabular and image datasets validate\nthat our fair training algorithm can achieve superior fairness-accuracy\ntrade-offs.\n", "date": "2024-06-05", "categories": []}
{"id": "2406.03292", "title": "Evaluating AI fairness in credit scoring with the BRIO tool", "abstract": "  We present a method for quantitative, in-depth analyses of fairness issues in\nAI systems with an application to credit scoring. To this aim we use BRIO, a\ntool for the evaluation of AI systems with respect to social unfairness and,\nmore in general, ethically undesirable behaviours. It features a model-agnostic\nbias detection module, presented in \\cite{DBLP:conf/beware/CoragliaDGGPPQ23},\nto which a full-fledged unfairness risk evaluation module is added. As a case\nstudy, we focus on the context of credit scoring, analysing the UCI German\nCredit Dataset \\cite{misc_statlog_(german_credit_data)_144}. We apply the BRIO\nfairness metrics to several, socially sensitive attributes featured in the\nGerman Credit Dataset, quantifying fairness across various demographic\nsegments, with the aim of identifying potential sources of bias and\ndiscrimination in a credit scoring model. We conclude by combining our results\nwith a revenue analysis.\n", "date": "2024-06-05", "categories": []}
{"id": "2306.11112", "title": "Correcting Underrepresentation and Intersectional Bias for\n  Classification", "abstract": "  We consider the problem of learning from data corrupted by\nunderrepresentation bias, where positive examples are filtered from the data at\ndifferent, unknown rates for a fixed number of sensitive groups. We show that\nwith a small amount of unbiased data, we can efficiently estimate the\ngroup-wise drop-out rates, even in settings where intersectional group\nmembership makes learning each intersectional rate computationally infeasible.\nUsing these estimates, we construct a reweighting scheme that allows us to\napproximate the loss of any hypothesis on the true distribution, even if we\nonly observe the empirical error on a biased sample. From this, we present an\nalgorithm encapsulating this learning and reweighting process along with a\nthorough empirical investigation. Finally, we define a bespoke notion of PAC\nlearnability for the underrepresentation and intersectional bias setting and\nshow that our algorithm permits efficient learning for model classes of finite\nVC dimension.\n", "date": "2023-06-19", "categories": []}
{"id": "2405.06582", "title": "The Role of Learning Algorithms in Collective Action", "abstract": "  Collective action in machine learning is the study of the control that a\ncoordinated group can have over machine learning algorithms. While previous\nresearch has concentrated on assessing the impact of collectives against Bayes\n(sub-)optimal classifiers, this perspective is limited in that it does not\naccount for the choice of learning algorithm. Since classifiers seldom behave\nlike Bayes classifiers and are influenced by the choice of learning algorithms\nalong with their inherent biases, in this work we initiate the study of how the\nchoice of the learning algorithm plays a role in the success of a collective in\npractical settings. Specifically, we focus on distributionally robust\noptimization (DRO), popular for improving a worst group error, and on the\nubiquitous stochastic gradient descent (SGD), due to its inductive bias for\n\"simpler\" functions. Our empirical results, supported by a theoretical\nfoundation, show that the effective size and success of the collective are\nhighly dependent on properties of the learning algorithm. This highlights the\nnecessity of taking the learning algorithm into account when studying the\nimpact of collective action in machine learning.\n", "date": "2024-05-10", "categories": []}
{"id": "2406.01757", "title": "Position: Cracking the Code of Cascading Disparity Towards Marginalized\n  Communities", "abstract": "  The rise of foundation models holds immense promise for advancing AI, but\nthis progress may amplify existing risks and inequalities, leaving marginalized\ncommunities behind. In this position paper, we discuss that disparities towards\nmarginalized communities - performance, representation, privacy, robustness,\ninterpretability and safety - are not isolated concerns but rather\ninterconnected elements of a cascading disparity phenomenon. We contrast\nfoundation models with traditional models and highlight the potential for\nexacerbated disparity against marginalized communities. Moreover, we emphasize\nthe unique threat of cascading impacts in foundation models, where\ninterconnected disparities can trigger long-lasting negative consequences,\nspecifically to the people on the margin. We define marginalized communities\nwithin the machine learning context and explore the multifaceted nature of\ndisparities. We analyze the sources of these disparities, tracing them from\ndata creation, training and deployment procedures to highlight the complex\ntechnical and socio-technical landscape. To mitigate the pressing crisis, we\nconclude with a set of calls to action to mitigate disparity at its source.\n", "date": "2024-06-03", "categories": []}
{"id": "2406.02361", "title": "Using Self-supervised Learning Can Improve Model Fairness", "abstract": "  Self-supervised learning (SSL) has become the de facto training paradigm of\nlarge models, where pre-training is followed by supervised fine-tuning using\ndomain-specific data and labels. Despite demonstrating comparable performance\nwith supervised methods, comprehensive efforts to assess SSL's impact on\nmachine learning fairness (i.e., performing equally on different demographic\nbreakdowns) are lacking. Hypothesizing that SSL models would learn more\ngeneric, hence less biased representations, this study explores the impact of\npre-training and fine-tuning strategies on fairness. We introduce a fairness\nassessment framework for SSL, comprising five stages: defining dataset\nrequirements, pre-training, fine-tuning with gradual unfreezing, assessing\nrepresentation similarity conditioned on demographics, and establishing\ndomain-specific evaluation processes. We evaluate our method's generalizability\non three real-world human-centric datasets (i.e., MIMIC, MESA, and GLOBEM) by\nsystematically comparing hundreds of SSL and fine-tuned models on various\ndimensions spanning from the intermediate representations to appropriate\nevaluation metrics. Our findings demonstrate that SSL can significantly improve\nmodel fairness, while maintaining performance on par with supervised\nmethods-exhibiting up to a 30% increase in fairness with minimal loss in\nperformance through self-supervision. We posit that such differences can be\nattributed to representation dissimilarities found between the best- and the\nworst-performing demographics across models-up to x13 greater for protected\nattributes with larger performance discrepancies between segments.\n", "date": "2024-06-04", "categories": []}
{"id": "2207.04581", "title": "How Robust is your Fair Model? Exploring the Robustness of Diverse\n  Fairness Strategies", "abstract": "  With the introduction of machine learning in high-stakes decision making,\nensuring algorithmic fairness has become an increasingly important problem to\nsolve. In response to this, many mathematical definitions of fairness have been\nproposed, and a variety of optimisation techniques have been developed, all\ndesigned to maximise a defined notion of fairness. However, fair solutions are\nreliant on the quality of the training data, and can be highly sensitive to\nnoise. Recent studies have shown that robustness (the ability for a model to\nperform well on unseen data) plays a significant role in the type of strategy\nthat should be used when approaching a new problem and, hence, measuring the\nrobustness of these strategies has become a fundamental problem. In this work,\nwe therefore propose a new criterion to measure the robustness of various\nfairness optimisation strategies - the robustness ratio. We conduct multiple\nextensive experiments on five bench mark fairness data sets using three of the\nmost popular fairness strategies with respect to four of the most popular\ndefinitions of fairness. Our experiments empirically show that fairness methods\nthat rely on threshold optimisation are very sensitive to noise in all the\nevaluated data sets, despite mostly outperforming other methods. This is in\ncontrast to the other two methods, which are less fair for low noise scenarios\nbut fairer for high noise ones. To the best of our knowledge, we are the first\nto quantitatively evaluate the robustness of fairness optimisation strategies.\nThis can potentially can serve as a guideline in choosing the most suitable\nfairness strategy for various data sets.\n", "date": "2022-07-11", "categories": []}
{"id": "2303.01704", "title": "Feature Importance Disparities for Data Bias Investigations", "abstract": "  It is widely held that one cause of downstream bias in classifiers is bias\npresent in the training data. Rectifying such biases may involve\ncontext-dependent interventions such as training separate models on subgroups,\nremoving features with bias in the collection process, or even conducting\nreal-world experiments to ascertain sources of bias. Despite the need for such\ndata bias investigations, few automated methods exist to assist practitioners\nin these efforts. In this paper, we present one such method that given a\ndataset $X$ consisting of protected and unprotected features, outcomes $y$, and\na regressor $h$ that predicts $y$ given $X$, outputs a tuple $(f_j, g)$, with\nthe following property: $g$ corresponds to a subset of the training dataset\n$(X, y)$, such that the $j^{th}$ feature $f_j$ has much larger (or smaller)\ninfluence in the subgroup $g$, than on the dataset overall, which we call\nfeature importance disparity (FID). We show across $4$ datasets and $4$ common\nfeature importance methods of broad interest to the machine learning community\nthat we can efficiently find subgroups with large FID values even over\nexponentially large subgroup classes and in practice these groups correspond to\nsubgroups with potentially serious bias issues as measured by standard fairness\nmetrics.\n", "date": "2023-03-03", "categories": []}
{"id": "2312.10396", "title": "How Far Can Fairness Constraints Help Recover From Biased Data?", "abstract": "  A general belief in fair classification is that fairness constraints incur a\ntrade-off with accuracy, which biased data may worsen. Contrary to this belief,\nBlum & Stangl (2019) show that fair classification with equal opportunity\nconstraints even on extremely biased data can recover optimally accurate and\nfair classifiers on the original data distribution. Their result is interesting\nbecause it demonstrates that fairness constraints can implicitly rectify data\nbias and simultaneously overcome a perceived fairness-accuracy trade-off. Their\ndata bias model simulates under-representation and label bias in\nunderprivileged population, and they show the above result on a stylized data\ndistribution with i.i.d. label noise, under simple conditions on the data\ndistribution and bias parameters. We propose a general approach to extend the\nresult of Blum & Stangl (2019) to different fairness constraints, data bias\nmodels, data distributions, and hypothesis classes. We strengthen their result,\nand extend it to the case when their stylized distribution has labels with\nMassart noise instead of i.i.d. noise. We prove a similar recovery result for\narbitrary data distributions using fair reject option classifiers. We further\ngeneralize it to arbitrary data distributions and arbitrary hypothesis classes,\ni.e., we prove that for any data distribution, if the optimally accurate\nclassifier in a given hypothesis class is fair and robust, then it can be\nrecovered through fair classification with equal opportunity constraints on the\nbiased distribution whenever the bias parameters satisfy certain simple\nconditions. Finally, we show applications of our technique to time-varying data\nbias in classification and fair machine learning pipelines.\n", "date": "2023-12-16", "categories": []}
{"id": "2402.11338", "title": "Fair Classification with Partial Feedback: An Exploration-Based Data\n  Collection Approach", "abstract": "  In many predictive contexts (e.g., credit lending), true outcomes are only\nobserved for samples that were positively classified in the past. These past\nobservations, in turn, form training datasets for classifiers that make future\npredictions. However, such training datasets lack information about the\noutcomes of samples that were (incorrectly) negatively classified in the past\nand can lead to erroneous classifiers. We present an approach that trains a\nclassifier using available data and comes with a family of exploration\nstrategies to collect outcome data about subpopulations that otherwise would\nhave been ignored. For any exploration strategy, the approach comes with\nguarantees that (1) all sub-populations are explored, (2) the fraction of false\npositives is bounded, and (3) the trained classifier converges to a ``desired''\nclassifier. The right exploration strategy is context-dependent; it can be\nchosen to improve learning guarantees and encode context-specific group\nfairness properties. Evaluation on real-world datasets shows that this approach\nconsistently boosts the quality of collected outcome data and improves the\nfraction of true positives for all groups, with only a small reduction in\npredictive utility.\n", "date": "2024-02-17", "categories": []}
{"id": "2405.02475", "title": "Generalizing Orthogonalization for Models with Non-Linearities", "abstract": "  The complexity of black-box algorithms can lead to various challenges,\nincluding the introduction of biases. These biases present immediate risks in\nthe algorithms' application. It was, for instance, shown that neural networks\ncan deduce racial information solely from a patient's X-ray scan, a task beyond\nthe capability of medical experts. If this fact is not known to the medical\nexpert, automatic decision-making based on this algorithm could lead to\nprescribing a treatment (purely) based on racial information. While current\nmethodologies allow for the \"orthogonalization\" or \"normalization\" of neural\nnetworks with respect to such information, existing approaches are grounded in\nlinear models. Our paper advances the discourse by introducing corrections for\nnon-linearities such as ReLU activations. Our approach also encompasses scalar\nand tensor-valued predictions, facilitating its integration into neural network\narchitectures. Through extensive experiments, we validate our method's\neffectiveness in safeguarding sensitive data in generalized linear models,\nnormalizing convolutional neural networks for metadata, and rectifying\npre-existing embeddings for undesired attributes.\n", "date": "2024-05-03", "categories": []}
{"id": "2406.00219", "title": "Fairness in Autonomous Driving: Towards Understanding Confounding\n  Factors in Object Detection under Challenging Weather", "abstract": "  The deployment of autonomous vehicles (AVs) is rapidly expanding to numerous\ncities. At the heart of AVs, the object detection module assumes a paramount\nrole, directly influencing all downstream decision-making tasks by considering\nthe presence of nearby pedestrians, vehicles, and more. Despite high accuracy\nof pedestrians detected on held-out datasets, the potential presence of\nalgorithmic bias in such object detectors, particularly in challenging weather\nconditions, remains unclear. This study provides a comprehensive empirical\nanalysis of fairness in detecting pedestrians in a state-of-the-art\ntransformer-based object detector. In addition to classical metrics, we\nintroduce novel probability-based metrics to measure various intricate\nproperties of object detection. Leveraging the state-of-the-art FACET dataset\nand the Carla high-fidelity vehicle simulator, our analysis explores the effect\nof protected attributes such as gender, skin tone, and body size on object\ndetection performance in varying environmental conditions such as ambient\ndarkness and fog. Our quantitative analysis reveals how the previously\noverlooked yet intuitive factors, such as the distribution of demographic\ngroups in the scene, the severity of weather, the pedestrians' proximity to the\nAV, among others, affect object detection performance. Our code is available at\nhttps://github.com/bimsarapathiraja/fair-AV.\n", "date": "2024-05-31", "categories": []}
{"id": "2406.00787", "title": "Applying Intrinsic Debiasing on Downstream Tasks: Challenges and\n  Considerations for Machine Translation", "abstract": "  Most works on gender bias focus on intrinsic bias -- removing traces of\ninformation about a protected group from the model's internal representation.\nHowever, these works are often disconnected from the impact of such debiasing\non downstream applications, which is the main motivation for debiasing in the\nfirst place. In this work, we systematically test how methods for intrinsic\ndebiasing affect neural machine translation models, by measuring the extrinsic\nbias of such systems under different design choices. We highlight three\nchallenges and mismatches between the debiasing techniques and their end-goal\nusage, including the choice of embeddings to debias, the mismatch between words\nand sub-word tokens debiasing, and the effect on different target languages. We\nfind that these considerations have a significant impact on downstream\nperformance and the success of debiasing.\n", "date": "2024-06-02", "categories": []}
{"id": "2405.20447", "title": "Algorithmic Fairness in Performative Policy Learning: Escaping the\n  Impossibility of Group Fairness", "abstract": "  In many prediction problems, the predictive model affects the distribution of\nthe prediction target. This phenomenon is known as performativity and is often\ncaused by the behavior of individuals with vested interests in the outcome of\nthe predictive model. Although performativity is generally problematic because\nit manifests as distribution shifts, we develop algorithmic fairness practices\nthat leverage performativity to achieve stronger group fairness guarantees in\nsocial classification problems (compared to what is achievable in\nnon-performative settings). In particular, we leverage the policymaker's\nability to steer the population to remedy inequities in the long term. A\ncrucial benefit of this approach is that it is possible to resolve the\nincompatibilities between conflicting group fairness definitions.\n", "date": "2024-05-30", "categories": []}
{"id": "2210.14562", "title": "FairCLIP: Social Bias Elimination based on Attribute Prototype Learning\n  and Representation Neutralization", "abstract": "  The Vision-Language Pre-training (VLP) models like CLIP have gained\npopularity in recent years. However, many works found that the social biases\nhidden in CLIP easily manifest in downstream tasks, especially in image\nretrieval, which can have harmful effects on human society. In this work, we\npropose FairCLIP to eliminate the social bias in CLIP-based image retrieval\nwithout damaging the retrieval performance achieving the compatibility between\nthe debiasing effect and the retrieval performance. FairCLIP is divided into\ntwo steps: Attribute Prototype Learning (APL) and Representation Neutralization\n(RN). In the first step, we extract the concepts needed for debiasing in CLIP.\nWe use the query with learnable word vector prefixes as the extraction\nstructure. In the second step, we first divide the attributes into target and\nbias attributes. By analysis, we find that both attributes have an impact on\nthe bias. Therefore, we try to eliminate the bias by using Re-Representation\nMatrix (RRM) to achieve the neutralization of the representation. We compare\nthe debiasing effect and retrieval performance with other methods, and\nexperiments demonstrate that FairCLIP can achieve the best compatibility.\nAlthough FairCLIP is used to eliminate bias in image retrieval, it achieves the\nneutralization of the representation which is common to all CLIP downstream\ntasks. This means that FairCLIP can be applied as a general debiasing method\nfor other fairness issues related to CLIP.\n", "date": "2022-10-26", "categories": []}
{"id": "2405.19667", "title": "Reconciling Model Multiplicity for Downstream Decision Making", "abstract": "  We consider the problem of model multiplicity in downstream decision-making,\na setting where two predictive models of equivalent accuracy cannot agree on\nthe best-response action for a downstream loss function. We show that even when\nthe two predictive models approximately agree on their individual predictions\nalmost everywhere, it is still possible for their induced best-response actions\nto differ on a substantial portion of the population. We address this issue by\nproposing a framework that calibrates the predictive models with regard to both\nthe downstream decision-making problem and the individual probability\nprediction. Specifically, leveraging tools from multi-calibration, we provide\nan algorithm that, at each time-step, first reconciles the differences in\nindividual probability prediction, then calibrates the updated models such that\nthey are indistinguishable from the true probability distribution to the\ndecision-maker. We extend our results to the setting where one does not have\ndirect access to the true probability distribution and instead relies on a set\nof i.i.d data to be the empirical distribution. Finally, we provide a set of\nexperiments to empirically evaluate our methods: compared to existing work, our\nproposed algorithm creates a pair of predictive models with both improved\ndownstream decision-making losses and agrees on their best-response actions\nalmost everywhere.\n", "date": "2024-05-30", "categories": []}
{"id": "2405.20051", "title": "Threshold-Independent Fair Matching through Score Calibration", "abstract": "  Entity Matching (EM) is a critical task in numerous fields, such as\nhealthcare, finance, and public administration, as it identifies records that\nrefer to the same entity within or across different databases. EM faces\nconsiderable challenges, particularly with false positives and negatives. These\nare typically addressed by generating matching scores and apply thresholds to\nbalance false positives and negatives in various contexts. However, adjusting\nthese thresholds can affect the fairness of the outcomes, a critical factor\nthat remains largely overlooked in current fair EM research. The existing body\nof research on fair EM tends to concentrate on static thresholds, neglecting\ntheir critical impact on fairness. To address this, we introduce a new approach\nin EM using recent metrics for evaluating biases in score based binary\nclassification, particularly through the lens of distributional parity. This\napproach enables the application of various bias metrics like equalized odds,\nequal opportunity, and demographic parity without depending on threshold\nsettings. Our experiments with leading matching methods reveal potential\nbiases, and by applying a calibration technique for EM scores using Wasserstein\nbarycenters, we not only mitigate these biases but also preserve accuracy\nacross real world datasets. This paper contributes to the field of fairness in\ndata cleaning, especially within EM, which is a central task in data cleaning,\nby promoting a method for generating matching scores that reduce biases across\ndifferent thresholds.\n", "date": "2024-05-30", "categories": []}
{"id": "2305.12090", "title": "UP5: Unbiased Foundation Model for Fairness-aware Recommendation", "abstract": "  Recent advances in Foundation Models such as Large Language Models (LLMs)\nhave propelled them to the forefront of Recommender Systems (RS). Despite their\nutility, there is a growing concern that LLMs might inadvertently perpetuate\nsocietal stereotypes, resulting in unfair recommendations. Since fairness is\ncritical for RS as many users take it for decision-making and demand\nfulfillment, this paper focuses on user-side fairness for LLM-based\nrecommendation where the users may require a recommender system to be fair on\nspecific sensitive features such as gender or age. In this paper, we dive into\nthe extent of unfairness exhibited by LLM-based recommender models based on\nboth T5 and LLaMA backbones, and discuss appropriate methods for promoting\nequitable treatment of users in LLM-based recommendation models. We introduce a\nnovel Counterfactually-Fair-Prompt (CFP) method towards Unbiased Foundation\nmOdels (UFO) for fairness-aware LLM-based recommendation. Experiments are\nconducted on two real-world datasets, MovieLens-1M and Insurance, and compared\nwith both matching-based and sequential-based fairness-aware recommendation\nmodels. Results show that CFP achieves better recommendation performance with a\nhigh level of fairness. Data and code are open-sourced at\nhttps://github.com/agiresearch/UP5.\n", "date": "2023-05-20", "categories": []}
{"id": "2405.13166", "title": "FairLENS: Assessing Fairness in Law Enforcement Speech Recognition", "abstract": "  Automatic speech recognition (ASR) techniques have become powerful tools,\nenhancing efficiency in law enforcement scenarios. To ensure fairness for\ndemographic groups in different acoustic environments, ASR engines must be\ntested across a variety of speakers in realistic settings. However, describing\nthe fairness discrepancies between models with confidence remains a challenge.\nMeanwhile, most public ASR datasets are insufficient to perform a satisfying\nfairness evaluation. To address the limitations, we built FairLENS - a\nsystematic fairness evaluation framework. We propose a novel and adaptable\nevaluation method to examine the fairness disparity between different models.\nWe also collected a fairness evaluation dataset covering multiple scenarios and\ndemographic dimensions. Leveraging this framework, we conducted fairness\nassessments on 1 open-source and 11 commercially available state-of-the-art ASR\nmodels. Our results reveal that certain models exhibit more biases than others,\nserving as a fairness guideline for users to make informed choices when\nselecting ASR models for a given real-world scenario. We further explored model\nbiases towards specific demographic groups and observed that shifts in the\nacoustic domain can lead to the emergence of new biases.\n", "date": "2024-05-21", "categories": []}
{"id": "2405.18662", "title": "Understanding Intrinsic Socioeconomic Biases in Large Language Models", "abstract": "  Large Language Models (LLMs) are increasingly integrated into critical\ndecision-making processes, such as loan approvals and visa applications, where\ninherent biases can lead to discriminatory outcomes. In this paper, we examine\nthe nuanced relationship between demographic attributes and socioeconomic\nbiases in LLMs, a crucial yet understudied area of fairness in LLMs. We\nintroduce a novel dataset of one million English sentences to systematically\nquantify socioeconomic biases across various demographic groups. Our findings\nreveal pervasive socioeconomic biases in both established models such as GPT-2\nand state-of-the-art models like Llama 2 and Falcon. We demonstrate that these\nbiases are significantly amplified when considering intersectionality, with\nLLMs exhibiting a remarkable capacity to extract multiple demographic\nattributes from names and then correlate them with specific socioeconomic\nbiases. This research highlights the urgent necessity for proactive and robust\nbias mitigation techniques to safeguard against discriminatory outcomes when\ndeploying these powerful models in critical real-world applications.\n", "date": "2024-05-28", "categories": []}
{"id": "2405.19022", "title": "Towards Standardizing AI Bias Exploration", "abstract": "  Creating fair AI systems is a complex problem that involves the assessment of\ncontext-dependent bias concerns. Existing research and programming libraries\nexpress specific concerns as measures of bias that they aim to constrain or\nmitigate. In practice, one should explore a wide variety of (sometimes\nincompatible) measures before deciding which ones warrant corrective action,\nbut their narrow scope means that most new situations can only be examined\nafter devising new measures. In this work, we present a mathematical framework\nthat distils literature measures of bias into building blocks, hereby\nfacilitating new combinations to cover a wide range of fairness concerns, such\nas classification or recommendation differences across multiple multi-value\nsensitive attributes (e.g., many genders and races, and their intersections).\nWe show how this framework generalizes existing concepts and present frequently\nused blocks. We provide an open-source implementation of our framework as a\nPython library, called FairBench, that facilitates systematic and extensible\nexploration of potential bias concerns.\n", "date": "2024-05-29", "categories": []}
{"id": "2405.19299", "title": "Expert-Guided Extinction of Toxic Tokens for Debiased Generation", "abstract": "  Large language models (LLMs) can elicit social bias during generations,\nespecially when inference with toxic prompts. Controlling the sensitive\nattributes in generation encounters challenges in data distribution,\ngeneralizability, and efficiency. Specifically, fine-tuning and retrieval\ndemand extensive unbiased corpus, while direct prompting requires meticulously\ncurated instructions for correcting the output in multiple rounds of thoughts\nbut poses challenges on memory and inference latency. In this work, we propose\nthe Expert-Guided Extinction of Toxic Tokens for Debiased Generation (EXPOSED)\nto eliminate the undesired harmful outputs for LLMs without the aforementioned\nrequirements. EXPOSED constructs a debiasing expert based on the abundant toxic\ncorpus to expose and elicit the potentially dangerous tokens. It then processes\nthe output to the LLMs and constructs a fair distribution by suppressing and\nattenuating the toxic tokens. EXPOSED is evaluated on fairness benchmarks over\nthree LLM families. Extensive experiments demonstrate that compared with other\nbaselines, the proposed EXPOSED significantly reduces the potential social bias\nwhile balancing fairness and generation performance.\n", "date": "2024-05-29", "categories": []}
{"id": "2309.03426", "title": "Adapting Static Fairness to Sequential Decision-Making: Bias Mitigation\n  Strategies towards Equal Long-term Benefit Rate", "abstract": "  Decisions made by machine learning models can have lasting impacts, making\nlong-term fairness a critical consideration. It has been observed that ignoring\nthe long-term effect and directly applying fairness criterion in static\nsettings can actually worsen bias over time. To address biases in sequential\ndecision-making, we introduce a long-term fairness concept named Equal\nLong-term Benefit Rate (ELBERT). This concept is seamlessly integrated into a\nMarkov Decision Process (MDP) to consider the future effects of actions on\nlong-term fairness, thus providing a unified framework for fair sequential\ndecision-making problems. ELBERT effectively addresses the temporal\ndiscrimination issues found in previous long-term fairness notions.\nAdditionally, we demonstrate that the policy gradient of Long-term Benefit Rate\ncan be analytically simplified to standard policy gradients. This\nsimplification makes conventional policy optimization methods viable for\nreducing bias, leading to our bias mitigation approach ELBERT-PO. Extensive\nexperiments across various diverse sequential decision-making environments\nconsistently reveal that ELBERT-PO significantly diminishes bias while\nmaintaining high utility. Code is available at\nhttps://github.com/umd-huang-lab/ELBERT.\n", "date": "2023-09-07", "categories": []}
{"id": "2405.17782", "title": "Post-Fair Federated Learning: Achieving Group and Community Fairness in\n  Federated Learning via Post-processing", "abstract": "  Federated Learning (FL) is a distributed machine learning framework in which\na set of local communities collaboratively learn a shared global model while\nretaining all training data locally within each community. Two notions of\nfairness have recently emerged as important issues for federated learning:\ngroup fairness and community fairness. Group fairness requires that a model's\ndecisions do not favor any particular group based on a set of legally protected\nattributes such as race or gender. Community fairness requires that global\nmodels exhibit similar levels of performance (accuracy) across all\ncollaborating communities. Both fairness concepts can coexist within an FL\nframework, but the existing literature has focused on either one concept or the\nother. This paper proposes and analyzes a post-processing fair federated\nlearning (FFL) framework called post-FFL. Post-FFL uses a linear program to\nsimultaneously enforce group and community fairness while maximizing the\nutility of the global model. Because Post-FFL is a post-processing approach, it\ncan be used with existing FL training pipelines whose convergence properties\nare well understood. This paper uses post-FFL on real-world datasets to mimic\nhow hospital networks, for example, use federated learning to deliver community\nhealth care. Theoretical results bound the accuracy lost when post-FFL enforces\nboth notion of fairness. Experimental results illustrate that post-FFL\nsimultaneously improves both group and community fairness in FL. Moreover,\npost-FFL outperforms the existing in-processing fair federated learning in\nterms of improving both notions of fairness, communication efficiency and\ncomputation cost.\n", "date": "2024-05-28", "categories": []}
{"id": "2405.17921", "title": "Towards Clinical AI Fairness: Filling Gaps in the Puzzle", "abstract": "  The ethical integration of Artificial Intelligence (AI) in healthcare\nnecessitates addressing fairness-a concept that is highly context-specific\nacross medical fields. Extensive studies have been conducted to expand the\ntechnical components of AI fairness, while tremendous calls for AI fairness\nhave been raised from healthcare. Despite this, a significant disconnect\npersists between technical advancements and their practical clinical\napplications, resulting in a lack of contextualized discussion of AI fairness\nin clinical settings. Through a detailed evidence gap analysis, our review\nsystematically pinpoints several deficiencies concerning both healthcare data\nand the provided AI fairness solutions. We highlight the scarcity of research\non AI fairness in many medical domains where AI technology is increasingly\nutilized. Additionally, our analysis highlights a substantial reliance on group\nfairness, aiming to ensure equality among demographic groups from a macro\nhealthcare system perspective; in contrast, individual fairness, focusing on\nequity at a more granular level, is frequently overlooked. To bridge these\ngaps, our review advances actionable strategies for both the healthcare and AI\nresearch communities. Beyond applying existing AI fairness methods in\nhealthcare, we further emphasize the importance of involving healthcare\nprofessionals to refine AI fairness concepts and methods to ensure contextually\nrelevant and ethically sound AI applications in healthcare.\n", "date": "2024-05-28", "categories": []}
{"id": "2306.10050", "title": "Interpolating Item and User Fairness in Multi-Sided Recommendations", "abstract": "  Today's online platforms heavily lean on algorithmic recommendations for\nbolstering user engagement and driving revenue. However, these recommendations\ncan impact multiple stakeholders simultaneously -- the platform, items\n(sellers), and users (customers) -- each with their unique objectives, making\nit difficult to find the right middle ground that accommodates all\nstakeholders. To address this, we introduce a novel fair recommendation\nframework, Problem (FAIR), that flexibly balances multi-stakeholder interests\nvia a constrained optimization formulation. We next explore Problem (FAIR) in a\ndynamic online setting where data uncertainty further adds complexity, and\npropose a low-regret algorithm FORM that concurrently performs real-time\nlearning and fair recommendations, two tasks that are often at odds. Via both\ntheoretical analysis and a numerical case study on real-world data, we\ndemonstrate the efficacy of our framework and method in maintaining platform\nrevenue while ensuring desired levels of fairness for both items and users.\n", "date": "2023-06-12", "categories": []}
{"id": "2312.03867", "title": "Multi-Group Fairness Evaluation via Conditional Value-at-Risk Testing", "abstract": "  Machine learning (ML) models used in prediction and classification tasks may\ndisplay performance disparities across population groups determined by\nsensitive attributes (e.g., race, sex, age). We consider the problem of\nevaluating the performance of a fixed ML model across population groups defined\nby multiple sensitive attributes (e.g., race and sex and age). Here, the sample\ncomplexity for estimating the worst-case performance gap across groups (e.g.,\nthe largest difference in error rates) increases exponentially with the number\nof group-denoting sensitive attributes. To address this issue, we propose an\napproach to test for performance disparities based on Conditional Value-at-Risk\n(CVaR). By allowing a small probabilistic slack on the groups over which a\nmodel has approximately equal performance, we show that the sample complexity\nrequired for discovering performance violations is reduced exponentially to be\nat most upper bounded by the square root of the number of groups. As a\nbyproduct of our analysis, when the groups are weighted by a specific prior\ndistribution, we show that R\\'enyi entropy of order 2/3 of the prior\ndistribution captures the sample complexity of the proposed CVaR test\nalgorithm. Finally, we also show that there exists a non-i.i.d. data collection\nstrategy that results in a sample complexity independent of the number of\ngroups.\n", "date": "2023-12-06", "categories": []}
{"id": "2405.16762", "title": "Addressing Discretization-Induced Bias in Demographic Prediction", "abstract": "  Racial and other demographic imputation is necessary for many applications,\nespecially in auditing disparities and outreach targeting in political\ncampaigns. The canonical approach is to construct continuous predictions --\ne.g., based on name and geography -- and then to $\\textit{discretize}$ the\npredictions by selecting the most likely class (argmax). We study how this\npractice produces $\\textit{discretization bias}$. In particular, we show that\nargmax labeling, as used by a prominent commercial voter file vendor to impute\nrace/ethnicity, results in a substantial under-count of African-American\nvoters, e.g., by 28.2% points in North Carolina. This bias can have substantial\nimplications in downstream tasks that use such labels.\n  We then introduce a $\\textit{joint optimization}$ approach -- and a tractable\n$\\textit{data-driven thresholding}$ heuristic -- that can eliminate this bias,\nwith negligible individual-level accuracy loss. Finally, we theoretically\nanalyze discretization bias, show that calibrated continuous models are\ninsufficient to eliminate it, and that an approach such as ours is necessary.\nBroadly, we warn researchers and practitioners against discretizing continuous\ndemographic predictions without considering downstream consequences.\n", "date": "2024-05-27", "categories": []}
{"id": "2403.18803", "title": "Projective Methods for Mitigating Gender Bias in Pre-trained Language\n  Models", "abstract": "  Mitigation of gender bias in NLP has a long history tied to debiasing static\nword embeddings. More recently, attention has shifted to debiasing pre-trained\nlanguage models. We study to what extent the simplest projective debiasing\nmethods, developed for word embeddings, can help when applied to BERT's\ninternal representations. Projective methods are fast to implement, use a small\nnumber of saved parameters, and make no updates to the existing model\nparameters. We evaluate the efficacy of the methods in reducing both intrinsic\nbias, as measured by BERT's next sentence prediction task, and in mitigating\nobserved bias in a downstream setting when fine-tuned. To this end, we also\nprovide a critical analysis of a popular gender-bias assessment test for\nquantifying intrinsic bias, resulting in an enhanced test set and new bias\nmeasures. We find that projective methods can be effective at both intrinsic\nbias and downstream bias mitigation, but that the two outcomes are not\nnecessarily correlated. This finding serves as a warning that intrinsic bias\ntest sets, based either on language modeling tasks or next sentence prediction,\nshould not be the only benchmark in developing a debiased language model.\n", "date": "2024-03-27", "categories": []}
{"id": "2405.15446", "title": "Mind the Gap: A Causal Perspective on Bias Amplification in Prediction &\n  Decision-Making", "abstract": "  Investigating fairness and equity of automated systems has become a critical\nfield of inquiry. Most of the literature in fair machine learning focuses on\ndefining and achieving fairness criteria in the context of prediction, while\nnot explicitly focusing on how these predictions may be used later on in the\npipeline. For instance, if commonly used criteria, such as independence or\nsufficiency, are satisfied for a prediction score $S$ used for binary\nclassification, they need not be satisfied after an application of a simple\nthresholding operation on $S$ (as commonly used in practice). In this paper, we\ntake an important step to address this issue in numerous statistical and causal\nnotions of fairness. We introduce the notion of a margin complement, which\nmeasures how much a prediction score $S$ changes due to a thresholding\noperation. We then demonstrate that the marginal difference in the optimal 0/1\npredictor $\\widehat Y$ between groups, written $P(\\hat y \\mid x_1) - P(\\hat y\n\\mid x_0)$, can be causally decomposed into the influences of $X$ on the\n$L_2$-optimal prediction score $S$ and the influences of $X$ on the margin\ncomplement $M$, along different causal pathways (direct, indirect, spurious).\nWe then show that under suitable causal assumptions, the influences of $X$ on\nthe prediction score $S$ are equal to the influences of $X$ on the true outcome\n$Y$. This yields a new decomposition of the disparity in the predictor\n$\\widehat Y$ that allows us to disentangle causal differences inherited from\nthe true outcome $Y$ that exists in the real world vs. those coming from the\noptimization procedure itself. This observation highlights the need for more\nregulatory oversight due to the potential for bias amplification, and to\naddress this issue we introduce new notions of weak and strong business\nnecessity, together with an algorithm for assessing whether these notions are\nsatisfied.\n", "date": "2024-05-24", "categories": []}
{"id": "2302.07185", "title": "When mitigating bias is unfair: multiplicity and arbitrariness in\n  algorithmic group fairness", "abstract": "  Most research on fair machine learning has prioritized optimizing criteria\nsuch as Demographic Parity and Equalized Odds. Despite these efforts, there\nremains a limited understanding of how different bias mitigation strategies\naffect individual predictions and whether they introduce arbitrariness into the\ndebiasing process. This paper addresses these gaps by exploring whether models\nthat achieve comparable fairness and accuracy metrics impact the same\nindividuals and mitigate bias in a consistent manner. We introduce the FRAME\n(FaiRness Arbitrariness and Multiplicity Evaluation) framework, which evaluates\nbias mitigation through five dimensions: Impact Size (how many people were\naffected), Change Direction (positive versus negative changes), Decision Rates\n(impact on models' acceptance rates), Affected Subpopulations (who was\naffected), and Neglected Subpopulations (where unfairness persists). This\nframework is intended to help practitioners understand the impacts of debiasing\nprocesses and make better-informed decisions regarding model selection.\nApplying FRAME to various bias mitigation approaches across key datasets allows\nus to exhibit significant differences in the behaviors of debiasing methods.\nThese findings highlight the limitations of current fairness criteria and the\ninherent arbitrariness in the debiasing process.\n", "date": "2023-02-14", "categories": []}
{"id": "2402.01732", "title": "Identifying and Improving Disability Bias in GPT-Based Resume Screening", "abstract": "  As Generative AI rises in adoption, its use has expanded to include domains\nsuch as hiring and recruiting. However, without examining the potential of\nbias, this may negatively impact marginalized populations, including people\nwith disabilities. To address this important concern, we present a resume audit\nstudy, in which we ask ChatGPT (specifically, GPT-4) to rank a resume against\nthe same resume enhanced with an additional leadership award, scholarship,\npanel presentation, and membership that are disability related. We find that\nGPT-4 exhibits prejudice towards these enhanced CVs. Further, we show that this\nprejudice can be quantifiably reduced by training a custom GPTs on principles\nof DEI and disability justice. Our study also includes a unique qualitative\nanalysis of the types of direct and indirect ableism GPT-4 uses to justify its\nbiased decisions and suggest directions for additional bias mitigation work.\nAdditionally, since these justifications are presumably drawn from training\ndata containing real-world biased statements made by humans, our analysis\nsuggests additional avenues for understanding and addressing human bias.\n", "date": "2024-01-28", "categories": []}
{"id": "2402.04105", "title": "Measuring Implicit Bias in Explicitly Unbiased Large Language Models", "abstract": "  Large language models (LLMs) can pass explicit social bias tests but still\nharbor implicit biases, similar to humans who endorse egalitarian beliefs yet\nexhibit subtle biases. Measuring such implicit biases can be a challenge: as\nLLMs become increasingly proprietary, it may not be possible to access their\nembeddings and apply existing bias measures; furthermore, implicit biases are\nprimarily a concern if they affect the actual decisions that these systems\nmake. We address both challenges by introducing two new measures of bias: LLM\nImplicit Bias, a prompt-based method for revealing implicit bias; and LLM\nDecision Bias, a strategy to detect subtle discrimination in decision-making\ntasks. Both measures are based on psychological research: LLM Implicit Bias\nadapts the Implicit Association Test, widely used to study the automatic\nassociations between concepts held in human minds; and LLM Decision Bias\noperationalizes psychological results indicating that relative evaluations\nbetween two candidates, not absolute evaluations assessing each independently,\nare more diagnostic of implicit biases. Using these measures, we found\npervasive stereotype biases mirroring those in society in 8 value-aligned\nmodels across 4 social categories (race, gender, religion, health) in 21\nstereotypes (such as race and criminality, race and weapons, gender and\nscience, age and negativity). Our prompt-based LLM Implicit Bias measure\ncorrelates with existing language model embedding-based bias methods, but\nbetter predicts downstream behaviors measured by LLM Decision Bias. These new\nprompt-based measures draw from psychology's long history of research into\nmeasuring stereotype biases based on purely observable behavior; they expose\nnuanced biases in proprietary value-aligned LLMs that appear unbiased according\nto standard benchmarks.\n", "date": "2024-02-06", "categories": []}
{"id": "2405.14521", "title": "Synthetic Data Generation for Intersectional Fairness by Leveraging\n  Hierarchical Group Structure", "abstract": "  In this paper, we introduce a data augmentation approach specifically\ntailored to enhance intersectional fairness in classification tasks. Our method\ncapitalizes on the hierarchical structure inherent to intersectionality, by\nviewing groups as intersections of their parent categories. This perspective\nallows us to augment data for smaller groups by learning a transformation\nfunction that combines data from these parent groups. Our empirical analysis,\nconducted on four diverse datasets including both text and images, reveals that\nclassifiers trained with this data augmentation approach achieve superior\nintersectional fairness and are more robust to ``leveling down'' when compared\nto methods optimizing traditional group fairness metrics.\n", "date": "2024-05-23", "categories": []}
{"id": "2311.13816", "title": "Algorithmic Fairness Generalization under Covariate and Dependence\n  Shifts Simultaneously", "abstract": "  The endeavor to preserve the generalization of a fair and invariant\nclassifier across domains, especially in the presence of distribution shifts,\nbecomes a significant and intricate challenge in machine learning. In response\nto this challenge, numerous effective algorithms have been developed with a\nfocus on addressing the problem of fairness-aware domain generalization. These\nalgorithms are designed to navigate various types of distribution shifts, with\na particular emphasis on covariate and dependence shifts. In this context,\ncovariate shift pertains to changes in the marginal distribution of input\nfeatures, while dependence shift involves alterations in the joint distribution\nof the label variable and sensitive attributes. In this paper, we introduce a\nsimple but effective approach that aims to learn a fair and invariant\nclassifier by simultaneously addressing both covariate and dependence shifts\nacross domains. We assert the existence of an underlying transformation model\ncan transform data from one domain to another, while preserving the semantics\nrelated to non-sensitive attributes and classes. By augmenting various\nsynthetic data domains through the model, we learn a fair and invariant\nclassifier in source domains. This classifier can then be generalized to\nunknown target domains, maintaining both model prediction and fairness\nconcerns. Extensive empirical studies on four benchmark datasets demonstrate\nthat our approach surpasses state-of-the-art methods.\n", "date": "2023-11-23", "categories": []}
{"id": "2309.08375", "title": "Boosting Fair Classifier Generalization through Adaptive Priority\n  Reweighing", "abstract": "  With the increasing penetration of machine learning applications in critical\ndecision-making areas, calls for algorithmic fairness are more prominent.\nAlthough there have been various modalities to improve algorithmic fairness\nthrough learning with fairness constraints, their performance does not\ngeneralize well in the test set. A performance-promising fair algorithm with\nbetter generalizability is needed. This paper proposes a novel adaptive\nreweighing method to eliminate the impact of the distribution shifts between\ntraining and test data on model generalizability. Most previous reweighing\nmethods propose to assign a unified weight for each (sub)group. Rather, our\nmethod granularly models the distance from the sample predictions to the\ndecision boundary. Our adaptive reweighing method prioritizes samples closer to\nthe decision boundary and assigns a higher weight to improve the\ngeneralizability of fair classifiers. Extensive experiments are performed to\nvalidate the generalizability of our adaptive priority reweighing method for\naccuracy and fairness measures (i.e., equal opportunity, equalized odds, and\ndemographic parity) in tabular benchmarks. We also highlight the performance of\nour method in improving the fairness of language and vision models. The code is\navailable at https://github.com/che2198/APW.\n", "date": "2023-09-15", "categories": []}
{"id": "2309.09697", "title": "Evaluating Gender Bias of Pre-trained Language Models in Natural\n  Language Inference by Considering All Labels", "abstract": "  Discriminatory gender biases have been found in Pre-trained Language Models\n(PLMs) for multiple languages. In Natural Language Inference (NLI), existing\nbias evaluation methods have focused on the prediction results of one specific\nlabel out of three labels, such as neutral. However, such evaluation methods\ncan be inaccurate since unique biased inferences are associated with unique\nprediction labels. Addressing this limitation, we propose a bias evaluation\nmethod for PLMs, called NLI-CoAL, which considers all the three labels of NLI\ntask. First, we create three evaluation data groups that represent different\ntypes of biases. Then, we define a bias measure based on the corresponding\nlabel output of each data group. In the experiments, we introduce a\nmeta-evaluation technique for NLI bias measures and use it to confirm that our\nbias measure can distinguish biased, incorrect inferences from non-biased\nincorrect inferences better than the baseline, resulting in a more accurate\nbias evaluation. We create the datasets in English, Japanese, and Chinese, and\nsuccessfully validate the compatibility of our bias measure across multiple\nlanguages. Lastly, we observe the bias tendencies in PLMs of different\nlanguages. To our knowledge, we are the first to construct evaluation datasets\nand measure PLMs' bias from NLI in Japanese and Chinese.\n", "date": "2023-09-18", "categories": []}
{"id": "2402.15603", "title": "Differentially Private Fair Binary Classifications", "abstract": "  In this work, we investigate binary classification under the constraints of\nboth differential privacy and fairness. We first propose an algorithm based on\nthe decoupling technique for learning a classifier with only fairness\nguarantee. This algorithm takes in classifiers trained on different demographic\ngroups and generates a single classifier satisfying statistical parity. We then\nrefine this algorithm to incorporate differential privacy. The performance of\nthe final algorithm is rigorously examined in terms of privacy, fairness, and\nutility guarantees. Empirical evaluations conducted on the Adult and Credit\nCard datasets illustrate that our algorithm outperforms the state-of-the-art in\nterms of fairness guarantees, while maintaining the same level of privacy and\nutility.\n", "date": "2024-02-23", "categories": []}
{"id": "2404.04534", "title": "Impact of Fairness Regulations on Institutions' Policies and Population\n  Qualifications", "abstract": "  The proliferation of algorithmic systems has fueled discussions surrounding\nthe regulation and control of their social impact. Herein, we consider a system\nwhose primary objective is to maximize utility by selecting the most qualified\nindividuals. To promote demographic parity in the selection algorithm, we\nconsider penalizing discrimination across social groups. We examine conditions\nunder which a discrimination penalty can effectively reduce disparity in the\nselection. Additionally, we explore the implications of such a penalty when\nindividual qualifications may evolve over time in response to the imposed\npenalizing policy. We identify scenarios where the penalty could hinder the\nnatural attainment of equity within the population. Moreover, we propose\ncertain conditions that can counteract this undesirable outcome, thus ensuring\nfairness.\n", "date": "2024-04-06", "categories": []}
{"id": "2405.09483", "title": "DemOpts: Fairness corrections in COVID-19 case prediction models", "abstract": "  COVID-19 forecasting models have been used to inform decision making around\nresource allocation and intervention decisions e.g., hospital beds or\nstay-at-home orders. State of the art deep learning models often use multimodal\ndata such as mobility or socio-demographic data to enhance COVID-19 case\nprediction models. Nevertheless, related work has revealed under-reporting bias\nin COVID-19 cases as well as sampling bias in mobility data for certain\nminority racial and ethnic groups, which could in turn affect the fairness of\nthe COVID-19 predictions along race labels. In this paper, we show that state\nof the art deep learning models output mean prediction errors that are\nsignificantly different across racial and ethnic groups; and which could, in\nturn, support unfair policy decisions. We also propose a novel de-biasing\nmethod, DemOpts, to increase the fairness of deep learning based forecasting\nmodels trained on potentially biased datasets. Our results show that DemOpts\ncan achieve better error parity that other state of the art de-biasing\napproaches, thus effectively reducing the differences in the mean error\ndistributions across more racial and ethnic groups.\n", "date": "2024-05-15", "categories": []}
{"id": "2405.11421", "title": "Assessing Group Fairness with Social Welfare Optimization", "abstract": "  Statistical parity metrics have been widely studied and endorsed in the AI\ncommunity as a means of achieving fairness, but they suffer from at least two\nweaknesses. They disregard the actual welfare consequences of decisions and may\ntherefore fail to achieve the kind of fairness that is desired for\ndisadvantaged groups. In addition, they are often incompatible with each other,\nand there is no convincing justification for selecting one rather than another.\nThis paper explores whether a broader conception of social justice, based on\noptimizing a social welfare function (SWF), can be useful for assessing various\ndefinitions of parity. We focus on the well-known alpha fairness SWF, which has\nbeen defended by axiomatic and bargaining arguments over a period of 70 years.\nWe analyze the optimal solution and show that it can justify demographic parity\nor equalized odds under certain conditions, but frequently requires a departure\nfrom these types of parity. In addition, we find that predictive rate parity is\nof limited usefulness. These results suggest that optimization theory can shed\nlight on the intensely discussed question of how to achieve group fairness in\nAI.\n", "date": "2024-05-19", "categories": []}
{"id": "2403.15593", "title": "FairerCLIP: Debiasing CLIP's Zero-Shot Predictions using Functions in\n  RKHSs", "abstract": "  Large pre-trained vision-language models such as CLIP provide compact and\ngeneral-purpose representations of text and images that are demonstrably\neffective across multiple downstream zero-shot prediction tasks. However, owing\nto the nature of their training process, these models have the potential to 1)\npropagate or amplify societal biases in the training data and 2) learn to rely\non spurious features. This paper proposes FairerCLIP, a general approach for\nmaking zero-shot predictions of CLIP more fair and robust to spurious\ncorrelations. We formulate the problem of jointly debiasing CLIP's image and\ntext representations in reproducing kernel Hilbert spaces (RKHSs), which\naffords multiple benefits: 1) Flexibility: Unlike existing approaches, which\nare specialized to either learn with or without ground-truth labels, FairerCLIP\nis adaptable to learning in both scenarios. 2) Ease of Optimization: FairerCLIP\nlends itself to an iterative optimization involving closed-form solvers, which\nleads to $4\\times$-$10\\times$ faster training than the existing methods. 3)\nSample Efficiency: Under sample-limited conditions, FairerCLIP significantly\noutperforms baselines when they fail entirely. And, 4) Performance:\nEmpirically, FairerCLIP achieves appreciable accuracy gains on benchmark\nfairness and spurious correlation datasets over their respective baselines.\n", "date": "2024-03-22", "categories": []}
{"id": "2405.10431", "title": "Thinking Fair and Slow: On the Efficacy of Structured Prompts for\n  Debiasing Language Models", "abstract": "  Existing debiasing techniques are typically training-based or require access\nto the model's internals and output distributions, so they are inaccessible to\nend-users looking to adapt LLM outputs for their particular needs. In this\nstudy, we examine whether structured prompting techniques can offer\nopportunities for fair text generation. We evaluate a comprehensive\nend-user-focused iterative framework of debiasing that applies System 2\nthinking processes for prompts to induce logical, reflective, and critical text\ngeneration, with single, multi-step, instruction, and role-based variants. By\nsystematically evaluating many LLMs across many datasets and different\nprompting strategies, we show that the more complex System 2-based Implicative\nPrompts significantly improve over other techniques demonstrating lower mean\nbias in the outputs with competitive performance on the downstream tasks. Our\nwork offers research directions for the design and the potential of\nend-user-focused evaluative frameworks for LLM use.\n", "date": "2024-05-16", "categories": []}
{"id": "2405.06841", "title": "Bridging the Gap: Protocol Towards Fair and Consistent Affect Analysis", "abstract": "  The increasing integration of machine learning algorithms in daily life\nunderscores the critical need for fairness and equity in their deployment. As\nthese technologies play a pivotal role in decision-making, addressing biases\nacross diverse subpopulation groups, including age, gender, and race, becomes\nparamount. Automatic affect analysis, at the intersection of physiology,\npsychology, and machine learning, has seen significant development. However,\nexisting databases and methodologies lack uniformity, leading to biased\nevaluations. This work addresses these issues by analyzing six affective\ndatabases, annotating demographic attributes, and proposing a common protocol\nfor database partitioning. Emphasis is placed on fairness in evaluations.\nExtensive experiments with baseline and state-of-the-art methods demonstrate\nthe impact of these changes, revealing the inadequacy of prior assessments. The\nfindings underscore the importance of considering demographic attributes in\naffect analysis research and provide a foundation for more equitable\nmethodologies. Our annotations, code and pre-trained models are available at:\nhttps://github.com/dkollias/Fair-Consistent-Affect-Analysis\n", "date": "2024-05-10", "categories": []}
{"id": "2405.07393", "title": "Intrinsic Fairness-Accuracy Tradeoffs under Equalized Odds", "abstract": "  With the growing adoption of machine learning (ML) systems in areas like law\nenforcement, criminal justice, finance, hiring, and admissions, it is\nincreasingly critical to guarantee the fairness of decisions assisted by ML. In\nthis paper, we study the tradeoff between fairness and accuracy under the\nstatistical notion of equalized odds. We present a new upper bound on the\naccuracy (that holds for any classifier), as a function of the fairness budget.\nIn addition, our bounds also exhibit dependence on the underlying statistics of\nthe data, labels and the sensitive group attributes. We validate our\ntheoretical upper bounds through empirical analysis on three real-world\ndatasets: COMPAS, Adult, and Law School. Specifically, we compare our upper\nbound to the tradeoffs that are achieved by various existing fair classifiers\nin the literature. Our results show that achieving high accuracy subject to a\nlow-bias could be fundamentally limited based on the statistical disparity\nacross the groups.\n", "date": "2024-05-12", "categories": []}
{"id": "2405.10256", "title": "Biasing & Debiasing based Approach Towards Fair Knowledge Transfer for\n  Equitable Skin Analysis", "abstract": "  Deep learning models, particularly Convolutional Neural Networks (CNNs), have\ndemonstrated exceptional performance in diagnosing skin diseases, often\noutperforming dermatologists. However, they have also unveiled biases linked to\nspecific demographic traits, notably concerning diverse skin tones or gender,\nprompting concerns regarding fairness and limiting their widespread deployment.\nResearchers are actively working to ensure fairness in AI-based solutions, but\nexisting methods incur an accuracy loss when striving for fairness. To solve\nthis issue, we propose a `two-biased teachers' (i.e., biased on different\nsensitive attributes) based approach to transfer fair knowledge into the\nstudent network. Our approach mitigates biases present in the student network\nwithout harming its predictive accuracy. In fact, in most cases, our approach\nimproves the accuracy of the baseline model. To achieve this goal, we developed\na weighted loss function comprising biasing and debiasing loss terms. We\nsurpassed available state-of-the-art approaches to attain fairness and also\nimproved the accuracy at the same time. The proposed approach has been\nevaluated and validated on two dermatology datasets using standard accuracy and\nfairness evaluation measures. We will make source code publicly available to\nfoster reproducibility and future research.\n", "date": "2024-05-16", "categories": []}
{"id": "2405.09251", "title": "Does Machine Bring in Extra Bias in Learning? Approximating Fairness in\n  Models Promptly", "abstract": "  Providing various machine learning (ML) applications in the real world,\nconcerns about discrimination hidden in ML models are growing, particularly in\nhigh-stakes domains. Existing techniques for assessing the discrimination level\nof ML models include commonly used group and individual fairness measures.\nHowever, these two types of fairness measures are usually hard to be compatible\nwith each other, and even two different group fairness measures might be\nincompatible as well. To address this issue, we investigate to evaluate the\ndiscrimination level of classifiers from a manifold perspective and propose a\n\"harmonic fairness measure via manifolds (HFM)\" based on distances between\nsets. Yet the direct calculation of distances might be too expensive to afford,\nreducing its practical applicability. Therefore, we devise an approximation\nalgorithm named \"Approximation of distance between sets (ApproxDist)\" to\nfacilitate accurate estimation of distances, and we further demonstrate its\nalgorithmic effectiveness under certain reasonable assumptions. Empirical\nresults indicate that the proposed fairness measure HFM is valid and that the\nproposed ApproxDist is effective and efficient.\n", "date": "2024-05-15", "categories": []}
{"id": "2405.09543", "title": "Algorithmic Fairness: A Tolerance Perspective", "abstract": "  Recent advancements in machine learning and deep learning have brought\nalgorithmic fairness into sharp focus, illuminating concerns over\ndiscriminatory decision making that negatively impacts certain individuals or\ngroups. These concerns have manifested in legal, ethical, and societal\nchallenges, including the erosion of trust in intelligent systems. In response,\nthis survey delves into the existing literature on algorithmic fairness,\nspecifically highlighting its multifaceted social consequences. We introduce a\nnovel taxonomy based on 'tolerance', a term we define as the degree to which\nvariations in fairness outcomes are acceptable, providing a structured approach\nto understanding the subtleties of fairness within algorithmic decisions. Our\nsystematic review covers diverse industries, revealing critical insights into\nthe balance between algorithmic decision making and social equity. By\nsynthesizing these insights, we outline a series of emerging challenges and\npropose strategic directions for future research and policy making, with the\ngoal of advancing the field towards more equitable algorithmic systems.\n", "date": "2024-04-26", "categories": []}
{"id": "2303.07247", "title": "Are Models Trained on Indian Legal Data Fair?", "abstract": "  Recent advances and applications of language technology and artificial\nintelligence have enabled much success across multiple domains like law,\nmedical and mental health. AI-based Language Models, like Judgement Prediction,\nhave recently been proposed for the legal sector. However, these models are\nstrife with encoded social biases picked up from the training data. While bias\nand fairness have been studied across NLP, most studies primarily locate\nthemselves within a Western context. In this work, we present an initial\ninvestigation of fairness from the Indian perspective in the legal domain. We\nhighlight the propagation of learnt algorithmic biases in the bail prediction\ntask for models trained on Hindi legal documents. We evaluate the fairness gap\nusing demographic parity and show that a decision tree model trained for the\nbail prediction task has an overall fairness disparity of 0.237 between input\nfeatures associated with Hindus and Muslims. Additionally, we highlight the\nneed for further research and studies in the avenues of fairness/bias in\napplying AI in the legal sector with a specific focus on the Indian context.\n", "date": "2023-03-13", "categories": []}
{"id": "2401.14893", "title": "A structured regression approach for evaluating model performance across\n  intersectional subgroups", "abstract": "  Disaggregated evaluation is a central task in AI fairness assessment, where\nthe goal is to measure an AI system's performance across different subgroups\ndefined by combinations of demographic or other sensitive attributes. The\nstandard approach is to stratify the evaluation data across subgroups and\ncompute performance metrics separately for each group. However, even for\nmoderately-sized evaluation datasets, sample sizes quickly get small once\nconsidering intersectional subgroups, which greatly limits the extent to which\nintersectional groups are included in analysis. In this work, we introduce a\nstructured regression approach to disaggregated evaluation that we demonstrate\ncan yield reliable system performance estimates even for very small subgroups.\nWe provide corresponding inference strategies for constructing confidence\nintervals and explore how goodness-of-fit testing can yield insight into the\nstructure of fairness-related harms experienced by intersectional groups. We\nevaluate our approach on two publicly available datasets, and several variants\nof semi-synthetic data. The results show that our method is considerably more\naccurate than the standard approach, especially for small subgroups, and\ndemonstrate how goodness-of-fit testing helps identify the key factors that\ndrive differences in performance.\n", "date": "2024-01-26", "categories": []}
{"id": "2405.08681", "title": "Achieving Fairness Through Channel Pruning for Dermatological Disease\n  Diagnosis", "abstract": "  Numerous studies have revealed that deep learning-based medical image\nclassification models may exhibit bias towards specific demographic attributes,\nsuch as race, gender, and age. Existing bias mitigation methods often achieve\nhigh level of fairness at the cost of significant accuracy degradation. In\nresponse to this challenge, we propose an innovative and adaptable Soft Nearest\nNeighbor Loss-based channel pruning framework, which achieves fairness through\nchannel pruning. Traditionally, channel pruning is utilized to accelerate\nneural network inference. However, our work demonstrates that pruning can also\nbe a potent tool for achieving fairness. Our key insight is that different\nchannels in a layer contribute differently to the accuracy of different groups.\nBy selectively pruning critical channels that lead to the accuracy difference\nbetween the privileged and unprivileged groups, we can effectively improve\nfairness without sacrificing accuracy significantly. Experiments conducted on\ntwo skin lesion diagnosis datasets across multiple sensitive attributes\nvalidate the effectiveness of our method in achieving state-of-the-art\ntrade-off between accuracy and fairness. Our code is available at\nhttps://github.com/Kqp1227/Sensitive-Channel-Pruning.\n", "date": "2024-05-14", "categories": []}
{"id": "2311.08605", "title": "Exploring the Jungle of Bias: Political Bias Attribution in Language\n  Models via Dependency Analysis", "abstract": "  The rapid advancement of Large Language Models (LLMs) has sparked intense\ndebate regarding the prevalence of bias in these models and its mitigation.\nYet, as exemplified by both results on debiasing methods in the literature and\nreports of alignment-related defects from the wider community, bias remains a\npoorly understood topic despite its practical relevance. To enhance the\nunderstanding of the internal causes of bias, we analyse LLM bias through the\nlens of causal fairness analysis, which enables us to both comprehend the\norigins of bias and reason about its downstream consequences and mitigation. To\noperationalize this framework, we propose a prompt-based method for the\nextraction of confounding and mediating attributes which contribute to the LLM\ndecision process. By applying Activity Dependency Networks (ADNs), we then\nanalyse how these attributes influence an LLM's decision process. We apply our\nmethod to LLM ratings of argument quality in political debates. We find that\nthe observed disparate treatment can at least in part be attributed to\nconfounding and mitigating attributes and model misalignment, and discuss the\nconsequences of our findings for human-AI alignment and bias mitigation. Our\ncode and data are at https://github.com/david-jenny/LLM-Political-Study.\n", "date": "2023-11-15", "categories": []}
{"id": "2405.06909", "title": "Fairness in Reinforcement Learning: A Survey", "abstract": "  While our understanding of fairness in machine learning has significantly\nprogressed, our understanding of fairness in reinforcement learning (RL)\nremains nascent. Most of the attention has been on fairness in one-shot\nclassification tasks; however, real-world, RL-enabled systems (e.g., autonomous\nvehicles) are much more complicated in that agents operate in dynamic\nenvironments over a long period of time. To ensure the responsible development\nand deployment of these systems, we must better understand fairness in RL. In\nthis paper, we survey the literature to provide the most up-to-date snapshot of\nthe frontiers of fairness in RL. We start by reviewing where fairness\nconsiderations can arise in RL, then discuss the various definitions of\nfairness in RL that have been put forth thus far. We continue to highlight the\nmethodologies researchers used to implement fairness in single- and multi-agent\nRL systems before showcasing the distinct application domains that fair RL has\nbeen investigated in. Finally, we critically examine gaps in the literature,\nsuch as understanding fairness in the context of RLHF, that still need to be\naddressed in future work to truly operationalize fair RL in real-world systems.\n", "date": "2024-05-11", "categories": []}
{"id": "2405.07011", "title": "Fair Graph Representation Learning via Sensitive Attribute\n  Disentanglement", "abstract": "  Group fairness for Graph Neural Networks (GNNs), which emphasizes algorithmic\ndecisions neither favoring nor harming certain groups defined by sensitive\nattributes (e.g., race and gender), has gained considerable attention. In\nparticular, the objective of group fairness is to ensure that the decisions\nmade by GNNs are independent of the sensitive attribute. To achieve this\nobjective, most existing approaches involve eliminating sensitive attribute\ninformation in node representations or algorithmic decisions. However, such\nways may also eliminate task-related information due to its inherent\ncorrelation with the sensitive attribute, leading to a sacrifice in utility. In\nthis work, we focus on improving the fairness of GNNs while preserving\ntask-related information and propose a fair GNN framework named FairSAD.\nInstead of eliminating sensitive attribute information, FairSAD enhances the\nfairness of GNNs via Sensitive Attribute Disentanglement (SAD), which separates\nthe sensitive attribute-related information into an independent component to\nmitigate its impact. Additionally, FairSAD utilizes a channel masking mechanism\nto adaptively identify the sensitive attribute-related component and\nsubsequently decorrelates it. Overall, FairSAD minimizes the impact of the\nsensitive attribute on GNN outcomes rather than eliminating sensitive\nattributes, thereby preserving task-related information associated with the\nsensitive attribute. Furthermore, experiments conducted on several real-world\ndatasets demonstrate that FairSAD outperforms other state-of-the-art methods by\na significant margin in terms of both fairness and utility performance. Our\nsource code is available at https://github.com/ZzoomD/FairSAD.\n", "date": "2024-05-11", "categories": []}
{"id": "2208.09951", "title": "Individual Fairness under Varied Notions of Group Fairness in Bipartite\n  Matching - One Framework to Approximate Them All", "abstract": "  We study the probabilistic assignment of items to platforms that satisfies\nboth group and individual fairness constraints. Each item belongs to specific\ngroups and has a preference ordering over platforms. Each platform enforces\ngroup fairness by limiting the number of items per group that can be assigned\nto it. There could be multiple optimal solutions that satisfy the group\nfairness constraints, but this alone ignores item preferences. Our approach\nexplores a `best of both worlds fairness' solution to get a randomized\nmatching, which is ex-ante individually fair and ex-post group-fair. Thus, we\nseek a `probabilistic individually fair' distribution over `group-fair'\nmatchings where each item has a `high' probability of matching to one of its\ntop choices. This distribution is also ex-ante group-fair. Users can customize\nfairness constraints to suit their requirements. Our first result is a\npolynomial-time algorithm that computes a distribution over `group-fair'\nmatchings such that the individual fairness constraints are approximately\nsatisfied and the expected size of a matching is close to OPT. We empirically\ntest this on real-world datasets. We present two additional polynomial-time\nbi-criteria approximation algorithms that users can choose from to balance\ngroup fairness and individual fairness trade-offs.\n  For disjoint groups, we provide an exact polynomial-time algorithm adaptable\nto additional lower `group fairness' bounds. Extending our model, we encompass\n`maxmin group fairness,' amplifying underrepresented groups, and `mindom group\nfairness,' reducing the representation of dominant groups.'\n", "date": "2022-08-21", "categories": []}
{"id": "2311.14688", "title": "Procedural Fairness Through Decoupling Objectionable Data Generating\n  Components", "abstract": "  We reveal and address the frequently overlooked yet important issue of\ndisguised procedural unfairness, namely, the potentially inadvertent\nalterations on the behavior of neutral (i.e., not problematic) aspects of data\ngenerating process, and/or the lack of procedural assurance of the greatest\nbenefit of the least advantaged individuals. Inspired by John Rawls's advocacy\nfor pure procedural justice, we view automated decision-making as a microcosm\nof social institutions, and consider how the data generating process itself can\nsatisfy the requirements of procedural fairness. We propose a framework that\ndecouples the objectionable data generating components from the neutral ones by\nutilizing reference points and the associated value instantiation rule. Our\nfindings highlight the necessity of preventing disguised procedural unfairness,\ndrawing attention not only to the objectionable data generating components that\nwe aim to mitigate, but also more importantly, to the neutral components that\nwe intend to keep unaffected.\n", "date": "2023-11-05", "categories": []}
{"id": "2405.05934", "title": "Theoretical Guarantees of Data Augmented Last Layer Retraining Methods", "abstract": "  Ensuring fair predictions across many distinct subpopulations in the training\ndata can be prohibitive for large models. Recently, simple linear last layer\nretraining strategies, in combination with data augmentation methods such as\nupweighting, downsampling and mixup, have been shown to achieve\nstate-of-the-art performance for worst-group accuracy, which quantifies\naccuracy for the least prevalent subpopulation. For linear last layer\nretraining and the abovementioned augmentations, we present the optimal\nworst-group accuracy when modeling the distribution of the latent\nrepresentations (input to the last layer) as Gaussian for each subpopulation.\nWe evaluate and verify our results for both synthetic and large publicly\navailable datasets.\n", "date": "2024-05-09", "categories": []}
{"id": "2405.01711", "title": "Individual Fairness Through Reweighting and Tuning", "abstract": "  Inherent bias within society can be amplified and perpetuated by artificial\nintelligence (AI) systems. To address this issue, a wide range of solutions\nhave been proposed to identify and mitigate bias and enforce fairness for\nindividuals and groups. Recently, Graph Laplacian Regularizer (GLR), a\nregularization technique from the semi-supervised learning literature has been\nused as a substitute for the common Lipschitz condition to enhance individual\nfairness. Notable prior work has shown that enforcing individual fairness\nthrough a GLR can improve the transfer learning accuracy of AI models under\ncovariate shifts. However, the prior work defines a GLR on the source and\ntarget data combined, implicitly assuming that the target data are available at\ntrain time, which might not hold in practice. In this work, we investigated\nwhether defining a GLR independently on the train and target data could\nmaintain similar accuracy. Furthermore, we introduced the Normalized Fairness\nGain score (NFG) to measure individual fairness by measuring the amount of\ngained fairness when a GLR is used versus not. We evaluated the new and\noriginal methods under NFG, the Prediction Consistency (PC), and traditional\nclassification metrics on the German Credit Approval dataset. The results\nshowed that the two models achieved similar statistical mean performances over\nfive-fold cross-validation. Furthermore, the proposed metric showed that PC\nscores can be misleading as the scores can be high and statistically similar to\nfairness-enhanced models while NFG scores are small. This work therefore\nprovides new insights into when a GLR effectively enhances individual fairness\nand the pitfalls of PC.\n", "date": "2024-05-02", "categories": []}
{"id": "2310.13007", "title": "A Critical Survey on Fairness Benefits of Explainable AI", "abstract": "  In this critical survey, we analyze typical claims on the relationship\nbetween explainable AI (XAI) and fairness to disentangle the multidimensional\nrelationship between these two concepts. Based on a systematic literature\nreview and a subsequent qualitative content analysis, we identify seven\narchetypal claims from 175 scientific articles on the alleged fairness benefits\nof XAI. We present crucial caveats with respect to these claims and provide an\nentry point for future discussions around the potentials and limitations of XAI\nfor specific fairness desiderata. Importantly, we notice that claims are often\n(i) vague and simplistic, (ii) lacking normative grounding, or (iii) poorly\naligned with the actual capabilities of XAI. We suggest to conceive XAI not as\nan ethical panacea but as one of many tools to approach the multidimensional,\nsociotechnical challenge of algorithmic fairness. Moreover, when making a claim\nabout XAI and fairness, we emphasize the need to be more specific about what\nkind of XAI method is used, which fairness desideratum it refers to, how\nexactly it enables fairness, and who is the stakeholder that benefits from XAI.\n", "date": "2023-10-15", "categories": []}
{"id": "2312.05429", "title": "Mitigating Nonlinear Algorithmic Bias in Binary Classification", "abstract": "  This paper proposes the use of causal modeling to detect and mitigate\nalgorithmic bias that is nonlinear in the protected attribute. We provide a\ngeneral overview of our approach. We use the German Credit data set, which is\navailable for download from the UC Irvine Machine Learning Repository, to\ndevelop (1) a prediction model, which is treated as a black box, and (2) a\ncausal model for bias mitigation. In this paper, we focus on age bias and the\nproblem of binary classification. We show that the probability of getting\ncorrectly classified as \"low risk\" is lowest among young people. The\nprobability increases with age nonlinearly. To incorporate the nonlinearity\ninto the causal model, we introduce a higher order polynomial term. Based on\nthe fitted causal model, the de-biased probability estimates are computed,\nshowing improved fairness with little impact on overall classification\naccuracy. Causal modeling is intuitive and, hence, its use can enhance\nexplicability and promotes trust among different stakeholders of AI.\n", "date": "2023-12-09", "categories": []}
{"id": "2402.01705", "title": "Beyond Behaviorist Representational Harms: A Plan for Measurement and\n  Mitigation", "abstract": "  Algorithmic harms are commonly categorized as either allocative or\nrepresentational. This study specifically addresses the latter, focusing on an\nexamination of current definitions of representational harms to discern what is\nincluded and what is not. This analysis motivates our expansion beyond\nbehavioral definitions to encompass harms to cognitive and affective states.\nThe paper outlines high-level requirements for measurement: identifying the\nnecessary expertise to implement this approach and illustrating it through a\ncase study. Our work highlights the unique vulnerabilities of large language\nmodels to perpetrating representational harms, particularly when these harms go\nunmeasured and unmitigated. The work concludes by presenting proposed\nmitigations and delineating when to employ them. The overarching aim of this\nresearch is to establish a framework for broadening the definition of\nrepresentational harms and to translate insights from fairness research into\npractical measurement and mitigation praxis.\n", "date": "2024-01-25", "categories": []}
{"id": "2405.04034", "title": "Differentially Private Post-Processing for Fair Regression", "abstract": "  This paper describes a differentially private post-processing algorithm for\nlearning fair regressors satisfying statistical parity, addressing privacy\nconcerns of machine learning models trained on sensitive data, as well as\nfairness concerns of their potential to propagate historical biases. Our\nalgorithm can be applied to post-process any given regressor to improve\nfairness by remapping its outputs. It consists of three steps: first, the\noutput distributions are estimated privately via histogram density estimation\nand the Laplace mechanism, then their Wasserstein barycenter is computed, and\nthe optimal transports to the barycenter are used for post-processing to\nsatisfy fairness. We analyze the sample complexity of our algorithm and provide\nfairness guarantee, revealing a trade-off between the statistical bias and\nvariance induced from the choice of the number of bins in the histogram, in\nwhich using less bins always favors fairness at the expense of error.\n", "date": "2024-05-07", "categories": []}
{"id": "2305.18569", "title": "Fairness of ChatGPT", "abstract": "  Understanding and addressing unfairness in LLMs are crucial for responsible\nAI deployment. However, there is a limited number of quantitative analyses and\nin-depth studies regarding fairness evaluations in LLMs, especially when\napplying LLMs to high-stakes fields. This work aims to fill this gap by\nproviding a systematic evaluation of the effectiveness and fairness of LLMs\nusing ChatGPT as a study case. We focus on assessing ChatGPT's performance in\nhigh-takes fields including education, criminology, finance and healthcare. To\nconduct a thorough evaluation, we consider both group fairness and individual\nfairness metrics. We also observe the disparities in ChatGPT's outputs under a\nset of biased or unbiased prompts. This work contributes to a deeper\nunderstanding of LLMs' fairness performance, facilitates bias mitigation and\nfosters the development of responsible AI systems.\n", "date": "2023-05-22", "categories": []}
{"id": "2307.02891", "title": "BaBE: Enhancing Fairness via Estimation of Latent Explaining Variables", "abstract": "  We consider the problem of unfair discrimination between two groups and\npropose a pre-processing method to achieve fairness. Corrective methods like\nstatistical parity usually lead to bad accuracy and do not really achieve\nfairness in situations where there is a correlation between the sensitive\nattribute S and the legitimate attribute E (explanatory variable) that should\ndetermine the decision. To overcome these drawbacks, other notions of fairness\nhave been proposed, in particular, conditional statistical parity and equal\nopportunity. However, E is often not directly observable in the data, i.e., it\nis a latent variable. We may observe some other variable Z representing E, but\nthe problem is that Z may also be affected by S, hence Z itself can be biased.\nTo deal with this problem, we propose BaBE (Bayesian Bias Elimination), an\napproach based on a combination of Bayes inference and the\nExpectation-Maximization method, to estimate the most likely value of E for a\ngiven Z for each group. The decision can then be based directly on the\nestimated E. We show, by experiments on synthetic and real data sets, that our\napproach provides a good level of fairness as well as high accuracy.\n", "date": "2023-07-06", "categories": []}
{"id": "2309.13005", "title": "Towards Counterfactual Fairness-aware Domain Generalization in Changing\n  Environments", "abstract": "  Recognizing the prevalence of domain shift as a common challenge in machine\nlearning, various domain generalization (DG) techniques have been developed to\nenhance the performance of machine learning systems when dealing with\nout-of-distribution (OOD) data. Furthermore, in real-world scenarios, data\ndistributions can gradually change across a sequence of sequential domains.\nWhile current methodologies primarily focus on improving model effectiveness\nwithin these new domains, they often overlook fairness issues throughout the\nlearning process. In response, we introduce an innovative framework called\nCounterfactual Fairness-Aware Domain Generalization with Sequential Autoencoder\n(CDSAE). This approach effectively separates environmental information and\nsensitive attributes from the embedded representation of classification\nfeatures. This concurrent separation not only greatly improves model\ngeneralization across diverse and unfamiliar domains but also effectively\naddresses challenges related to unfair classification. Our strategy is rooted\nin the principles of causal inference to tackle these dual issues. To examine\nthe intricate relationship between semantic information, sensitive attributes,\nand environmental cues, we systematically categorize exogenous uncertainty\nfactors into four latent variables: 1) semantic information influenced by\nsensitive attributes, 2) semantic information unaffected by sensitive\nattributes, 3) environmental cues influenced by sensitive attributes, and 4)\nenvironmental cues unaffected by sensitive attributes. By incorporating\nfairness regularization, we exclusively employ semantic information for\nclassification purposes. Empirical validation on synthetic and real-world\ndatasets substantiates the effectiveness of our approach, demonstrating\nimproved accuracy levels while ensuring the preservation of fairness in the\nevolving landscape of continuous domains.\n", "date": "2023-09-22", "categories": []}
{"id": "2402.01327", "title": "Supervised Algorithmic Fairness in Distribution Shifts: A Survey", "abstract": "  Supervised fairness-aware machine learning under distribution shifts is an\nemerging field that addresses the challenge of maintaining equitable and\nunbiased predictions when faced with changes in data distributions from source\nto target domains. In real-world applications, machine learning models are\noften trained on a specific dataset but deployed in environments where the data\ndistribution may shift over time due to various factors. This shift can lead to\nunfair predictions, disproportionately affecting certain groups characterized\nby sensitive attributes, such as race and gender. In this survey, we provide a\nsummary of various types of distribution shifts and comprehensively investigate\nexisting methods based on these shifts, highlighting six commonly used\napproaches in the literature. Additionally, this survey lists publicly\navailable datasets and evaluation metrics for empirical studies. We further\nexplore the interconnection with related research fields, discuss the\nsignificant challenges, and identify potential directions for future studies.\n", "date": "2024-02-02", "categories": []}
{"id": "2405.02743", "title": "Beyond Performance: Quantifying and Mitigating Label Bias in LLMs", "abstract": "  Large language models (LLMs) have shown remarkable adaptability to diverse\ntasks, by leveraging context prompts containing instructions, or minimal\ninput-output examples. However, recent work revealed they also exhibit label\nbias -- an undesirable preference toward predicting certain answers over\nothers. Still, detecting and measuring this bias reliably and at scale has\nremained relatively unexplored. In this study, we evaluate different approaches\nto quantifying label bias in a model's predictions, conducting a comprehensive\ninvestigation across 279 classification tasks and ten LLMs. Our investigation\nreveals substantial label bias in models both before and after debiasing\nattempts, as well as highlights the importance of outcomes-based evaluation\nmetrics, which were not previously used in this regard. We further propose a\nnovel label bias calibration method tailored for few-shot prompting, which\noutperforms recent calibration approaches for both improving performance and\nmitigating label bias. Our results emphasize that label bias in the predictions\nof LLMs remains a barrier to their reliability.\n", "date": "2024-05-04", "categories": []}
{"id": "2405.03098", "title": "FairMonitor: A Dual-framework for Detecting Stereotypes and Biases in\n  Large Language Models", "abstract": "  Detecting stereotypes and biases in Large Language Models (LLMs) is crucial\nfor enhancing fairness and reducing adverse impacts on individuals or groups\nwhen these models are applied. Traditional methods, which rely on embedding\nspaces or are based on probability metrics, fall short in revealing the nuanced\nand implicit biases present in various contexts. To address this challenge, we\npropose the FairMonitor framework and adopt a static-dynamic detection method\nfor a comprehensive evaluation of stereotypes and biases in LLMs. The static\ncomponent consists of a direct inquiry test, an implicit association test, and\nan unknown situation test, including 10,262 open-ended questions with 9\nsensitive factors and 26 educational scenarios. And it is effective for\nevaluating both explicit and implicit biases. Moreover, we utilize the\nmulti-agent system to construst the dynamic scenarios for detecting subtle\nbiases in more complex and realistic setting. This component detects the biases\nbased on the interaction behaviors of LLMs across 600 varied educational\nscenarios. The experimental results show that the cooperation of static and\ndynamic methods can detect more stereotypes and biased in LLMs.\n", "date": "2024-05-06", "categories": []}
{"id": "2401.08066", "title": "Achieve Fairness without Demographics for Dermatological Disease\n  Diagnosis", "abstract": "  In medical image diagnosis, fairness has become increasingly crucial. Without\nbias mitigation, deploying unfair AI would harm the interests of the\nunderprivileged population and potentially tear society apart. Recent research\naddresses prediction biases in deep learning models concerning demographic\ngroups (e.g., gender, age, and race) by utilizing demographic (sensitive\nattribute) information during training. However, many sensitive attributes\nnaturally exist in dermatological disease images. If the trained model only\ntargets fairness for a specific attribute, it remains unfair for other\nattributes. Moreover, training a model that can accommodate multiple sensitive\nattributes is impractical due to privacy concerns. To overcome this, we propose\na method enabling fair predictions for sensitive attributes during the testing\nphase without using such information during training. Inspired by prior work\nhighlighting the impact of feature entanglement on fairness, we enhance the\nmodel features by capturing the features related to the sensitive and target\nattributes and regularizing the feature entanglement between corresponding\nclasses. This ensures that the model can only classify based on the features\nrelated to the target attribute without relying on features associated with\nsensitive attributes, thereby improving fairness and accuracy. Additionally, we\nuse disease masks from the Segment Anything Model (SAM) to enhance the quality\nof the learned feature. Experimental results demonstrate that the proposed\nmethod can improve fairness in classification compared to state-of-the-art\nmethods in two dermatological disease datasets.\n", "date": "2024-01-16", "categories": []}
{"id": "2401.08788", "title": "The Impact of Differential Feature Under-reporting on Algorithmic\n  Fairness", "abstract": "  Predictive risk models in the public sector are commonly developed using\nadministrative data that is more complete for subpopulations that more greatly\nrely on public services. In the United States, for instance, information on\nhealth care utilization is routinely available to government agencies for\nindividuals supported by Medicaid and Medicare, but not for the privately\ninsured. Critiques of public sector algorithms have identified such\ndifferential feature under-reporting as a driver of disparities in algorithmic\ndecision-making. Yet this form of data bias remains understudied from a\ntechnical viewpoint. While prior work has examined the fairness impacts of\nadditive feature noise and features that are clearly marked as missing, the\nsetting of data missingness absent indicators (i.e. differential feature\nunder-reporting) has been lacking in research attention. In this work, we\npresent an analytically tractable model of differential feature under-reporting\nwhich we then use to characterize the impact of this kind of data bias on\nalgorithmic fairness. We demonstrate how standard missing data methods\ntypically fail to mitigate bias in this setting, and propose a new set of\nmethods specifically tailored to differential feature under-reporting. Our\nresults show that, in real world data settings, under-reporting typically leads\nto increasing disparities. The proposed solution methods show success in\nmitigating increases in unfairness.\n", "date": "2024-01-16", "categories": []}
{"id": "2405.01790", "title": "Understanding Position Bias Effects on Fairness in Social Multi-Document\n  Summarization", "abstract": "  Text summarization models have typically focused on optimizing aspects of\nquality such as fluency, relevance, and coherence, particularly in the context\nof news articles. However, summarization models are increasingly being used to\nsummarize diverse sources of text, such as social media data, that encompass a\nwide demographic user base. It is thus crucial to assess not only the quality\nof the generated summaries, but also the extent to which they can fairly\nrepresent the opinions of diverse social groups. Position bias, a long-known\nissue in news summarization, has received limited attention in the context of\nsocial multi-document summarization. We deeply investigate this phenomenon by\nanalyzing the effect of group ordering in input documents when summarizing\ntweets from three distinct linguistic communities: African-American English,\nHispanic-aligned Language, and White-aligned Language. Our empirical analysis\nshows that although the textual quality of the summaries remains consistent\nregardless of the input document order, in terms of fairness, the results vary\nsignificantly depending on how the dialect groups are presented in the input\ndata. Our results suggest that position bias manifests differently in social\nmulti-document summarization, severely impacting the fairness of summarization\nmodels.\n", "date": "2024-05-03", "categories": []}
{"id": "2405.02225", "title": "Fair Risk Control: A Generalized Framework for Calibrating Multi-group\n  Fairness Risks", "abstract": "  This paper introduces a framework for post-processing machine learning models\nso that their predictions satisfy multi-group fairness guarantees. Based on the\ncelebrated notion of multicalibration, we introduce $(\\mathbf{s},\\mathcal{G},\n\\alpha)-$GMC (Generalized Multi-Dimensional Multicalibration) for\nmulti-dimensional mappings $\\mathbf{s}$, constraint set $\\mathcal{G}$, and a\npre-specified threshold level $\\alpha$. We propose associated algorithms to\nachieve this notion in general settings. This framework is then applied to\ndiverse scenarios encompassing different fairness concerns, including false\nnegative rate control in image segmentation, prediction set conditional\nuncertainty quantification in hierarchical classification, and de-biased text\ngeneration in language models. We conduct numerical studies on several datasets\nand tasks.\n", "date": "2024-05-03", "categories": []}
{"id": "2404.01030", "title": "Survey of Bias In Text-to-Image Generation: Definition, Evaluation, and\n  Mitigation", "abstract": "  The recent advancement of large and powerful models with Text-to-Image (T2I)\ngeneration abilities -- such as OpenAI's DALLE-3 and Google's Gemini -- enables\nusers to generate high-quality images from textual prompts. However, it has\nbecome increasingly evident that even simple prompts could cause T2I models to\nexhibit conspicuous social bias in generated images. Such bias might lead to\nboth allocational and representational harms in society, further marginalizing\nminority groups. Noting this problem, a large body of recent works has been\ndedicated to investigating different dimensions of bias in T2I systems.\nHowever, an extensive review of these studies is lacking, hindering a\nsystematic understanding of current progress and research gaps. We present the\nfirst extensive survey on bias in T2I generative models. In this survey, we\nreview prior studies on dimensions of bias: Gender, Skintone, and Geo-Culture.\nSpecifically, we discuss how these works define, evaluate, and mitigate\ndifferent aspects of bias. We found that: (1) while gender and skintone biases\nare widely studied, geo-cultural bias remains under-explored; (2) most works on\ngender and skintone bias investigated occupational association, while other\naspects are less frequently studied; (3) almost all gender bias works overlook\nnon-binary identities in their studies; (4) evaluation datasets and metrics are\nscattered, with no unified framework for measuring biases; and (5) current\nmitigation methods fail to resolve biases comprehensively. Based on current\nlimitations, we point out future research directions that contribute to\nhuman-centric definitions, evaluations, and mitigation of biases. We hope to\nhighlight the importance of studying biases in T2I systems, as well as\nencourage future efforts to holistically understand and tackle biases, building\nfair and trustworthy T2I technologies for everyone.\n", "date": "2024-04-01", "categories": []}
{"id": "2405.00910", "title": "De-Biasing Models of Biased Decisions: A Comparison of Methods Using\n  Mortgage Application Data", "abstract": "  Prediction models can improve efficiency by automating decisions such as the\napproval of loan applications. However, they may inherit bias against protected\ngroups from the data they are trained on. This paper adds counterfactual\n(simulated) ethnic bias to real data on mortgage application decisions, and\nshows that this bias is replicated by a machine learning model (XGBoost) even\nwhen ethnicity is not used as a predictive variable. Next, several other\nde-biasing methods are compared: averaging over prohibited variables, taking\nthe most favorable prediction over prohibited variables (a novel method), and\njointly minimizing errors as well as the association between predictions and\nprohibited variables. De-biasing can recover some of the original decisions,\nbut the results are sensitive to whether the bias is effected through a proxy.\n", "date": "2024-05-01", "categories": []}
{"id": "2308.12820", "title": "Prediction without Preclusion: Recourse Verification with Reachable Sets", "abstract": "  Machine learning models are often used to decide who receives a loan, a job\ninterview, or a public benefit. Models in such settings use features without\nconsidering their actionability. As a result, they can assign predictions that\nare fixed $-$ meaning that individuals who are denied loans and interviews are,\nin fact, precluded from access to credit and employment. In this work, we\nintroduce a procedure called recourse verification to test if a model assigns\nfixed predictions to its decision subjects. We propose a model-agnostic\napproach for recourse verification with reachable sets $-$ i.e., the set of all\npoints that a person can reach through their actions in feature space. We\ndevelop methods to construct reachable sets for discrete feature spaces, which\ncan certify the responsiveness of any model by simply querying its predictions.\nWe conduct a comprehensive empirical study on the infeasibility of recourse on\ndatasets from consumer finance. Our results highlight how models can\ninadvertently preclude access by assigning fixed predictions and underscore the\nneed to account for actionability in model development.\n", "date": "2023-08-24", "categories": []}
{"id": "2405.00588", "title": "Are Models Biased on Text without Gender-related Language?", "abstract": "  Gender bias research has been pivotal in revealing undesirable behaviors in\nlarge language models, exposing serious gender stereotypes associated with\noccupations, and emotions. A key observation in prior work is that models\nreinforce stereotypes as a consequence of the gendered correlations that are\npresent in the training data. In this paper, we focus on bias where the effect\nfrom training data is unclear, and instead address the question: Do language\nmodels still exhibit gender bias in non-stereotypical settings? To do so, we\nintroduce UnStereoEval (USE), a novel framework tailored for investigating\ngender bias in stereotype-free scenarios. USE defines a sentence-level score\nbased on pretraining data statistics to determine if the sentence contain\nminimal word-gender associations. To systematically benchmark the fairness of\npopular language models in stereotype-free scenarios, we utilize USE to\nautomatically generate benchmarks without any gender-related language. By\nleveraging USE's sentence-level score, we also repurpose prior gender bias\nbenchmarks (Winobias and Winogender) for non-stereotypical evaluation.\nSurprisingly, we find low fairness across all 28 tested models. Concretely,\nmodels demonstrate fair behavior in only 9%-41% of stereotype-free sentences,\nsuggesting that bias does not solely stem from the presence of gender-related\nwords. These results raise important questions about where underlying model\nbiases come from and highlight the need for more systematic and comprehensive\nbias evaluation. We release the full dataset and code at\nhttps://ucinlp.github.io/unstereo-eval.\n", "date": "2024-05-01", "categories": []}
{"id": "2401.13555", "title": "Benchmarking the Fairness of Image Upsampling Methods", "abstract": "  Recent years have witnessed a rapid development of deep generative models for\ncreating synthetic media, such as images and videos. While the practical\napplications of these models in everyday tasks are enticing, it is crucial to\nassess the inherent risks regarding their fairness. In this work, we introduce\na comprehensive framework for benchmarking the performance and fairness of\nconditional generative models. We develop a set of\nmetrics$\\unicode{x2013}$inspired by their supervised fairness\ncounterparts$\\unicode{x2013}$to evaluate the models on their fairness and\ndiversity. Focusing on the specific application of image upsampling, we create\na benchmark covering a wide variety of modern upsampling methods. As part of\nthe benchmark, we introduce UnfairFace, a subset of FairFace that replicates\nthe racial distribution of common large-scale face datasets. Our empirical\nstudy highlights the importance of using an unbiased training set and reveals\nvariations in how the algorithms respond to dataset imbalances. Alarmingly, we\nfind that none of the considered methods produces statistically fair and\ndiverse results. All experiments can be reproduced using our provided\nrepository.\n", "date": "2024-01-24", "categories": []}
{"id": "2310.11401", "title": "Enhancing Group Fairness in Online Settings Using Oblique Decision\n  Forests", "abstract": "  Fairness, especially group fairness, is an important consideration in the\ncontext of machine learning systems. The most commonly adopted group\nfairness-enhancing techniques are in-processing methods that rely on a mixture\nof a fairness objective (e.g., demographic parity) and a task-specific\nobjective (e.g., cross-entropy) during the training process. However, when data\narrives in an online fashion -- one instance at a time -- optimizing such\nfairness objectives poses several challenges. In particular, group fairness\nobjectives are defined using expectations of predictions across different\ndemographic groups. In the online setting, where the algorithm has access to a\nsingle instance at a time, estimating the group fairness objective requires\nadditional storage and significantly more computation (e.g., forward/backward\npasses) than the task-specific objective at every time step. In this paper, we\npropose Aranyani, an ensemble of oblique decision trees, to make fair decisions\nin online settings. The hierarchical tree structure of Aranyani enables\nparameter isolation and allows us to efficiently compute the fairness gradients\nusing aggregate statistics of previous decisions, eliminating the need for\nadditional storage and forward/backward passes. We also present an efficient\nframework to train Aranyani and theoretically analyze several of its\nproperties. We conduct empirical evaluations on 5 publicly available benchmarks\n(including vision and language datasets) to show that Aranyani achieves a\nbetter accuracy-fairness trade-off compared to baseline approaches.\n", "date": "2023-10-17", "categories": []}
{"id": "2404.10942", "title": "What Hides behind Unfairness? Exploring Dynamics Fairness in\n  Reinforcement Learning", "abstract": "  In sequential decision-making problems involving sensitive attributes like\nrace and gender, reinforcement learning (RL) agents must carefully consider\nlong-term fairness while maximizing returns. Recent works have proposed many\ndifferent types of fairness notions, but how unfairness arises in RL problems\nremains unclear. In this paper, we address this gap in the literature by\ninvestigating the sources of inequality through a causal lens. We first analyse\nthe causal relationships governing the data generation process and decompose\nthe effect of sensitive attributes on long-term well-being into distinct\ncomponents. We then introduce a novel notion called dynamics fairness, which\nexplicitly captures the inequality stemming from environmental dynamics,\ndistinguishing it from those induced by decision-making or inherited from the\npast. This notion requires evaluating the expected changes in the next state\nand the reward induced by changing the value of the sensitive attribute while\nholding everything else constant. To quantitatively evaluate this\ncounterfactual concept, we derive identification formulas that allow us to\nobtain reliable estimations from data. Extensive experiments demonstrate the\neffectiveness of the proposed techniques in explaining, detecting, and reducing\ninequality in reinforcement learning. We publicly release code at\nhttps://github.com/familyld/InsightFair.\n", "date": "2024-04-16", "categories": []}
{"id": "2404.18276", "title": "Bias Neutralization Framework: Measuring Fairness in Large Language\n  Models with Bias Intelligence Quotient (BiQ)", "abstract": "  The burgeoning influence of Large Language Models (LLMs) in shaping public\ndiscourse and decision-making underscores the imperative to address inherent\nbiases within these AI systems. In the wake of AI's expansive integration\nacross sectors, addressing racial bias in LLMs has never been more critical.\nThis paper introduces a novel framework called Comprehensive Bias\nNeutralization Framework (CBNF) which embodies an innovative approach to\nquantifying and mitigating biases within LLMs. Our framework combines the Large\nLanguage Model Bias Index (LLMBI) [Oketunji, A., Anas, M., Saina, D., (2023)]\nand Bias removaL with No Demographics (BLIND) [Orgad, H., Belinkov, Y. (2023)]\nmethodologies to create a new metric called Bias Intelligence Quotient\n(BiQ)which detects, measures, and mitigates racial bias in LLMs without\nreliance on demographic annotations.\n  By introducing a new metric called BiQ that enhances LLMBI with additional\nfairness metrics, CBNF offers a multi-dimensional metric for bias assessment,\nunderscoring the necessity of a nuanced approach to fairness in AI [Mehrabi et\nal., 2021]. This paper presents a detailed analysis of Latimer AI (a language\nmodel incrementally trained on black history and culture) in comparison to\nChatGPT 3.5, illustrating Latimer AI's efficacy in detecting racial, cultural,\nand gender biases through targeted training and refined bias mitigation\nstrategies [Latimer & Bender, 2023].\n", "date": "2024-04-28", "categories": []}
{"id": "2404.18685", "title": "FALE: Fairness-Aware ALE Plots for Auditing Bias in Subgroups", "abstract": "  Fairness is steadily becoming a crucial requirement of Machine Learning (ML)\nsystems. A particularly important notion is subgroup fairness, i.e., fairness\nin subgroups of individuals that are defined by more than one attributes.\nIdentifying bias in subgroups can become both computationally challenging, as\nwell as problematic with respect to comprehensibility and intuitiveness of the\nfinding to end users. In this work we focus on the latter aspects; we propose\nan explainability method tailored to identifying potential bias in subgroups\nand visualizing the findings in a user friendly manner to end users. In\nparticular, we extend the ALE plots explainability method, proposing FALE\n(Fairness aware Accumulated Local Effects) plots, a method for measuring the\nchange in fairness for an affected population corresponding to different values\nof a feature (attribute). We envision FALE to function as an efficient, user\nfriendly, comprehensible and reliable first-stage tool for identifying\nsubgroups with potential bias issues.\n", "date": "2024-04-29", "categories": []}
{"id": "2305.12829", "title": "On Bias and Fairness in NLP: Investigating the Impact of Bias and\n  Debiasing in Language Models on the Fairness of Toxicity Detection", "abstract": "  Language models are the new state-of-the-art natural language processing\n(NLP) models and they are being increasingly used in many NLP tasks. Even\nthough there is evidence that language models are biased, the impact of that\nbias on the fairness of downstream NLP tasks is still understudied.\nFurthermore, despite that numerous debiasing methods have been proposed in the\nliterature, the impact of bias removal methods on the fairness of NLP tasks is\nalso understudied. In this work, we investigate three different sources of bias\nin NLP models, i.e. representation bias, selection bias and overamplification\nbias, and examine how they impact the fairness of the downstream task of\ntoxicity detection. Moreover, we investigate the impact of removing these\nbiases using different bias removal techniques on the fairness of toxicity\ndetection. Results show strong evidence that downstream sources of bias,\nespecially overamplification bias, are the most impactful types of bias on the\nfairness of the task of toxicity detection. We also found strong evidence that\nremoving overamplification bias by fine-tuning the language models on a dataset\nwith balanced contextual representations and ratios of positive examples\nbetween different identity groups can improve the fairness of the task of\ntoxicity detection. Finally, we build on our findings and introduce a list of\nguidelines to ensure the fairness of the task of toxicity detection.\n", "date": "2023-05-22", "categories": []}
{"id": "2308.10684", "title": "Systematic Offensive Stereotyping (SOS) Bias in Language Models", "abstract": "  In this paper, we propose a new metric to measure the SOS bias in language\nmodels (LMs). Then, we validate the SOS bias and investigate the effectiveness\nof removing it. Finally, we investigate the impact of the SOS bias in LMs on\ntheir performance and fairness on hate speech detection. Our results suggest\nthat all the inspected LMs are SOS biased. And that the SOS bias is reflective\nof the online hate experienced by marginalized identities. The results indicate\nthat using debias methods from the literature worsens the SOS bias in LMs for\nsome sensitive attributes and improves it for others. Finally, Our results\nsuggest that the SOS bias in the inspected LMs has an impact on their fairness\nof hate speech detection. However, there is no strong evidence that the SOS\nbias has an impact on the performance of hate speech detection.\n", "date": "2023-08-21", "categories": []}
{"id": "2404.17169", "title": "FairGT: A Fairness-aware Graph Transformer", "abstract": "  The design of Graph Transformers (GTs) generally neglects considerations for\nfairness, resulting in biased outcomes against certain sensitive subgroups.\nSince GTs encode graph information without relying on message-passing\nmechanisms, conventional fairness-aware graph learning methods cannot be\ndirectly applicable to address these issues. To tackle this challenge, we\npropose FairGT, a Fairness-aware Graph Transformer explicitly crafted to\nmitigate fairness concerns inherent in GTs. FairGT incorporates a meticulous\nstructural feature selection strategy and a multi-hop node feature integration\nmethod, ensuring independence of sensitive features and bolstering fairness\nconsiderations. These fairness-aware graph information encodings seamlessly\nintegrate into the Transformer framework for downstream tasks. We also prove\nthat the proposed fair structural topology encoding with adjacency matrix\neigenvector selection and multi-hop integration are theoretically effective.\nEmpirical evaluations conducted across five real-world datasets demonstrate\nFairGT's superiority in fairness metrics over existing graph transformers,\ngraph neural networks, and state-of-the-art fairness-aware graph learning\napproaches.\n", "date": "2024-04-26", "categories": []}
{"id": "2404.17511", "title": "Bridging the Fairness Divide: Achieving Group and Individual Fairness in\n  Graph Neural Networks", "abstract": "  Graph neural networks (GNNs) have emerged as a powerful tool for analyzing\nand learning from complex data structured as graphs, demonstrating remarkable\neffectiveness in various applications, such as social network analysis,\nrecommendation systems, and drug discovery. However, despite their impressive\nperformance, the fairness problem has increasingly gained attention as a\ncrucial aspect to consider. Existing research in graph learning focuses on\neither group fairness or individual fairness. However, since each concept\nprovides unique insights into fairness from distinct perspectives, integrating\nthem into a fair graph neural network system is crucial. To the best of our\nknowledge, no study has yet to comprehensively tackle both individual and group\nfairness simultaneously. In this paper, we propose a new concept of individual\nfairness within groups and a novel framework named Fairness for Group and\nIndividual (FairGI), which considers both group fairness and individual\nfairness within groups in the context of graph learning. FairGI employs the\nsimilarity matrix of individuals to achieve individual fairness within groups,\nwhile leveraging adversarial learning to address group fairness in terms of\nboth Equal Opportunity and Statistical Parity. The experimental results\ndemonstrate that our approach not only outperforms other state-of-the-art\nmodels in terms of group fairness and individual fairness within groups, but\nalso exhibits excellent performance in population-level individual fairness,\nwhile maintaining comparable prediction accuracy.\n", "date": "2024-04-26", "categories": []}
{"id": "2404.09454", "title": "Utility-Fairness Trade-Offs and How to Find Them", "abstract": "  When building classification systems with demographic fairness\nconsiderations, there are two objectives to satisfy: 1) maximizing utility for\nthe specific task and 2) ensuring fairness w.r.t. a known demographic\nattribute. These objectives often compete, so optimizing both can lead to a\ntrade-off between utility and fairness. While existing works acknowledge the\ntrade-offs and study their limits, two questions remain unanswered: 1) What are\nthe optimal trade-offs between utility and fairness? and 2) How can we\nnumerically quantify these trade-offs from data for a desired prediction task\nand demographic attribute of interest? This paper addresses these questions. We\nintroduce two utility-fairness trade-offs: the Data-Space and Label-Space\nTrade-off. The trade-offs reveal three regions within the utility-fairness\nplane, delineating what is fully and partially possible and impossible. We\npropose U-FaTE, a method to numerically quantify the trade-offs for a given\nprediction task and group fairness definition from data samples. Based on the\ntrade-offs, we introduce a new scheme for evaluating representations. An\nextensive evaluation of fair representation learning methods and\nrepresentations from over 1000 pre-trained models revealed that most current\napproaches are far from the estimated and achievable fairness-utility\ntrade-offs across multiple datasets and prediction tasks.\n", "date": "2024-04-15", "categories": []}
{"id": "2404.15385", "title": "Sum of Group Error Differences: A Critical Examination of Bias\n  Evaluation in Biometric Verification and a Dual-Metric Measure", "abstract": "  Biometric Verification (BV) systems often exhibit accuracy disparities across\ndifferent demographic groups, leading to biases in BV applications. Assessing\nand quantifying these biases is essential for ensuring the fairness of BV\nsystems. However, existing bias evaluation metrics in BV have limitations, such\nas focusing exclusively on match or non-match error rates, overlooking bias on\ndemographic groups with performance levels falling between the best and worst\nperformance levels, and neglecting the magnitude of the bias present.\n  This paper presents an in-depth analysis of the limitations of current bias\nevaluation metrics in BV and, through experimental analysis, demonstrates their\ncontextual suitability, merits, and limitations. Additionally, it introduces a\nnovel general-purpose bias evaluation measure for BV, the ``Sum of Group Error\nDifferences (SEDG)''. Our experimental results on controlled synthetic datasets\ndemonstrate the effectiveness of demographic bias quantification when using\nexisting metrics and our own proposed measure. We discuss the applicability of\nthe bias evaluation metrics in a set of simulated demographic bias scenarios\nand provide scenario-based metric recommendations. Our code is publicly\navailable under \\url{https://github.com/alaaobeid/SEDG}.\n", "date": "2024-04-23", "categories": []}
{"id": "2404.15418", "title": "Machine Learning Techniques with Fairness for Prediction of Completion\n  of Drug and Alcohol Rehabilitation", "abstract": "  The aim of this study is to look at predicting whether a person will complete\na drug and alcohol rehabilitation program and the number of times a person\nattends. The study is based on demographic data obtained from Substance Abuse\nand Mental Health Services Administration (SAMHSA) from both admissions and\ndischarge data from drug and alcohol rehabilitation centers in Oklahoma.\nDemographic data is highly categorical which led to binary encoding being used\nand various fairness measures being utilized to mitigate bias of nine\ndemographic variables. Kernel methods such as linear, polynomial, sigmoid, and\nradial basis functions were compared using support vector machines at various\nparameter ranges to find the optimal values. These were then compared to\nmethods such as decision trees, random forests, and neural networks. Synthetic\nMinority Oversampling Technique Nominal (SMOTEN) for categorical data was used\nto balance the data with imputation for missing data. The nine bias variables\nwere then intersectionalized to mitigate bias and the dual and triple\ninteractions were integrated to use the probabilities to look at worst case\nratio fairness mitigation. Disparate Impact, Statistical Parity difference,\nConditional Statistical Parity Ratio, Demographic Parity, Demographic Parity\nRatio, Equalized Odds, Equalized Odds Ratio, Equal Opportunity, and Equalized\nOpportunity Ratio were all explored at both the binary and multiclass\nscenarios.\n", "date": "2024-04-23", "categories": []}
{"id": "2404.14620", "title": "Fairness Incentives in Response to Unfair Dynamic Pricing", "abstract": "  The use of dynamic pricing by profit-maximizing firms gives rise to demand\nfairness concerns, measured by discrepancies in consumer groups' demand\nresponses to a given pricing strategy. Notably, dynamic pricing may result in\nbuyer distributions unreflective of those of the underlying population, which\ncan be problematic in markets where fair representation is socially desirable.\nTo address this, policy makers might leverage tools such as taxation and\nsubsidy to adapt policy mechanisms dependent upon their social objective. In\nthis paper, we explore the potential for AI methods to assist such intervention\nstrategies. To this end, we design a basic simulated economy, wherein we\nintroduce a dynamic social planner (SP) to generate corporate taxation\nschedules geared to incentivizing firms towards adopting fair pricing\nbehaviours, and to use the collected tax budget to subsidize consumption among\nunderrepresented groups. To cover a range of possible policy scenarios, we\nformulate our social planner's learning problem as a multi-armed bandit, a\ncontextual bandit and finally as a full reinforcement learning (RL) problem,\nevaluating welfare outcomes from each case. To alleviate the difficulty in\nretaining meaningful tax rates that apply to less frequently occurring\nbrackets, we introduce FairReplayBuffer, which ensures that our RL agent\nsamples experiences uniformly across a discretized fairness space. We find\nthat, upon deploying a learned tax and redistribution policy, social welfare\nimproves on that of the fairness-agnostic baseline, and approaches that of the\nanalytically optimal fairness-aware baseline for the multi-armed and contextual\nbandit settings, and surpassing it by 13.19% in the full RL setting.\n", "date": "2024-04-22", "categories": []}
{"id": "2307.14366", "title": "Explainable Disparity Compensation for Efficient Fair Ranking", "abstract": "  Ranking functions that are used in decision systems often produce disparate\nresults for different populations because of bias in the underlying data.\nAddressing, and compensating for, these disparate outcomes is a critical\nproblem for fair decision-making. Recent compensatory measures have mostly\nfocused on opaque transformations of the ranking functions to satisfy fairness\nguarantees or on the use of quotas or set-asides to guarantee a minimum number\nof positive outcomes to members of underrepresented groups. In this paper we\npropose easily explainable data-driven compensatory measures for ranking\nfunctions. Our measures rely on the generation of bonus points given to members\nof underrepresented groups to address disparity in the ranking function. The\nbonus points can be set in advance, and can be combined, allowing for\nconsidering the intersections of representations and giving better transparency\nto stakeholders. We propose efficient sampling-based algorithms to calculate\nthe number of bonus points to minimize disparity. We validate our algorithms\nusing real-world school admissions and recidivism datasets, and compare our\nresults with that of existing fair ranking algorithms.\n", "date": "2023-07-25", "categories": []}
{"id": "2404.14050", "title": "Unlawful Proxy Discrimination: A Framework for Challenging Inherently\n  Discriminatory Algorithms", "abstract": "  Emerging scholarship suggests that the EU legal concept of direct\ndiscrimination - where a person is given different treatment on grounds of a\nprotected characteristic - may apply to various algorithmic decision-making\ncontexts. This has important implications: unlike indirect discrimination,\nthere is generally no 'objective justification' stage in the direct\ndiscrimination framework, which means that the deployment of directly\ndiscriminatory algorithms will usually be unlawful per se. In this paper, we\nfocus on the most likely candidate for direct discrimination in the algorithmic\ncontext, termed inherent direct discrimination, where a proxy is inextricably\nlinked to a protected characteristic. We draw on computer science literature to\nsuggest that, in the algorithmic context, 'treatment on the grounds of' needs\nto be understood in terms of two steps: proxy capacity and proxy use. Only\nwhere both elements can be made out can direct discrimination be said to be `on\ngrounds of' a protected characteristic. We analyse the legal conditions of our\nproposed proxy capacity and proxy use tests. Based on this analysis, we discuss\ntechnical approaches and metrics that could be developed or applied to identify\ninherent direct discrimination in algorithmic decision-making.\n", "date": "2024-04-22", "categories": []}
{"id": "2304.09779", "title": "Equalised Odds is not Equal Individual Odds: Post-processing for Group\n  and Individual Fairness", "abstract": "  Group fairness is achieved by equalising prediction distributions between\nprotected sub-populations; individual fairness requires treating similar\nindividuals alike. These two objectives, however, are incompatible when a\nscoring model is calibrated through discontinuous probability functions, where\nindividuals can be randomly assigned an outcome determined by a fixed\nprobability. This procedure may provide two similar individuals from the same\nprotected group with classification odds that are disparately different -- a\nclear violation of individual fairness. Assigning unique odds to each protected\nsub-population may also prevent members of one sub-population from ever\nreceiving equal chances of a positive outcome to another, which we argue is\nanother type of unfairness called individual odds. We reconcile all this by\nconstructing continuous probability functions between group thresholds that are\nconstrained by their Lipschitz constant. Our solution preserves the model's\npredictive power, individual fairness and robustness while ensuring group\nfairness.\n", "date": "2023-04-19", "categories": []}
{"id": "2404.11782", "title": "REQUAL-LM: Reliability and Equity through Aggregation in Large Language\n  Models", "abstract": "  The extensive scope of large language models (LLMs) across various domains\nunderscores the critical importance of responsibility in their application,\nbeyond natural language processing. In particular, the randomized nature of\nLLMs, coupled with inherent biases and historical stereotypes in data, raises\ncritical concerns regarding reliability and equity. Addressing these challenges\nare necessary before using LLMs for applications with societal impact. Towards\naddressing this gap, we introduce REQUAL-LM, a novel method for finding\nreliable and equitable LLM outputs through aggregation. Specifically, we\ndevelop a Monte Carlo method based on repeated sampling to find a reliable\noutput close to the mean of the underlying distribution of possible outputs. We\nformally define the terms such as reliability and bias, and design an\nequity-aware aggregation to minimize harmful bias while finding a highly\nreliable output. REQUAL-LM does not require specialized hardware, does not\nimpose a significant computing load, and uses LLMs as a blackbox. This design\nchoice enables seamless scalability alongside the rapid advancement of LLM\ntechnologies. Our system does not require retraining the LLMs, which makes it\ndeployment ready and easy to adapt. Our comprehensive experiments using various\ntasks and datasets demonstrate that REQUAL- LM effectively mitigates bias and\nselects a more equitable response, specifically the outputs that properly\nrepresents minority groups.\n", "date": "2024-04-17", "categories": []}
{"id": "2404.12076", "title": "Evolutionary Multi-Objective Optimisation for Fairness-Aware Self\n  Adjusting Memory Classifiers in Data Streams", "abstract": "  This paper introduces a novel approach, evolutionary multi-objective\noptimisation for fairness-aware self-adjusting memory classifiers, designed to\nenhance fairness in machine learning algorithms applied to data stream\nclassification. With the growing concern over discrimination in algorithmic\ndecision-making, particularly in dynamic data stream environments, there is a\nneed for methods that ensure fair treatment of individuals across sensitive\nattributes like race or gender. The proposed approach addresses this challenge\nby integrating the strengths of the self-adjusting memory K-Nearest-Neighbour\nalgorithm with evolutionary multi-objective optimisation. This combination\nallows the new approach to efficiently manage concept drift in streaming data\nand leverage the flexibility of evolutionary multi-objective optimisation to\nmaximise accuracy and minimise discrimination simultaneously. We demonstrate\nthe effectiveness of the proposed approach through extensive experiments on\nvarious datasets, comparing its performance against several baseline methods in\nterms of accuracy and fairness metrics. Our results show that the proposed\napproach maintains competitive accuracy and significantly reduces\ndiscrimination, highlighting its potential as a robust solution for\nfairness-aware data stream classification. Further analyses also confirm the\neffectiveness of the strategies to trigger evolutionary multi-objective\noptimisation and adapt classifiers in the proposed approach.\n", "date": "2024-04-18", "categories": []}
{"id": "2404.12143", "title": "The Neutrality Fallacy: When Algorithmic Fairness Interventions are\n  (Not) Positive Action", "abstract": "  Various metrics and interventions have been developed to identify and\nmitigate unfair outputs of machine learning systems. While individuals and\norganizations have an obligation to avoid discrimination, the use of\nfairness-aware machine learning interventions has also been described as\namounting to 'algorithmic positive action' under European Union (EU)\nnon-discrimination law. As the Court of Justice of the European Union has been\nstrict when it comes to assessing the lawfulness of positive action, this would\nimpose a significant legal burden on those wishing to implement fair-ml\ninterventions. In this paper, we propose that algorithmic fairness\ninterventions often should be interpreted as a means to prevent discrimination,\nrather than a measure of positive action. Specifically, we suggest that this\ncategory mistake can often be attributed to neutrality fallacies: faulty\nassumptions regarding the neutrality of fairness-aware algorithmic\ndecision-making. Our findings raise the question of whether a negative\nobligation to refrain from discrimination is sufficient in the context of\nalgorithmic decision-making. Consequently, we suggest moving away from a duty\nto 'not do harm' towards a positive obligation to actively 'do no harm' as a\nmore adequate framework for algorithmic decision-making and fair\nml-interventions.\n", "date": "2024-04-18", "categories": []}
{"id": "2209.04589", "title": "Causal Intervention for Fairness in Multi-behavior Recommendation", "abstract": "  Recommender systems usually learn user interests from various user behaviors,\nincluding clicks and post-click behaviors (e.g., like and favorite). However,\nthese behaviors inevitably exhibit popularity bias, leading to some unfairness\nissues: 1) for items with similar quality, more popular ones get more exposure;\nand 2) even worse the popular items with lower popularity might receive more\nexposure. Existing work on mitigating popularity bias blindly eliminates the\nbias and usually ignores the effect of item quality. We argue that the\nrelationships between different user behaviors (e.g., conversion rate) actually\nreflect the item quality. Therefore, to handle the unfairness issues, we\npropose to mitigate the popularity bias by considering multiple user behaviors.\n  In this work, we examine causal relationships behind the interaction\ngeneration procedure in multi-behavior recommendation. Specifically, we find\nthat: 1) item popularity is a confounder between the exposed items and users'\npost-click interactions, leading to the first unfairness; and 2) some hidden\nconfounders (e.g., the reputation of item producers) affect both item\npopularity and quality, resulting in the second unfairness. To alleviate these\nconfounding issues, we propose a causal framework to estimate the causal\neffect, which leverages backdoor adjustment to block the backdoor paths caused\nby the confounders. In the inference stage, we remove the negative effect of\npopularity and utilize the good effect of quality for recommendation.\nExperiments on two real-world datasets validate the effectiveness of our\nproposed framework, which enhances fairness without sacrificing recommendation\naccuracy.\n", "date": "2022-09-10", "categories": []}
{"id": "2301.11781", "title": "Aleatoric and Epistemic Discrimination: Fundamental Limits of Fairness\n  Interventions", "abstract": "  Machine learning (ML) models can underperform on certain population groups\ndue to choices made during model development and bias inherent in the data. We\ncategorize sources of discrimination in the ML pipeline into two classes:\naleatoric discrimination, which is inherent in the data distribution, and\nepistemic discrimination, which is due to decisions made during model\ndevelopment. We quantify aleatoric discrimination by determining the\nperformance limits of a model under fairness constraints, assuming perfect\nknowledge of the data distribution. We demonstrate how to characterize\naleatoric discrimination by applying Blackwell's results on comparing\nstatistical experiments. We then quantify epistemic discrimination as the gap\nbetween a model's accuracy when fairness constraints are applied and the limit\nposed by aleatoric discrimination. We apply this approach to benchmark existing\nfairness interventions and investigate fairness risks in data with missing\nvalues. Our results indicate that state-of-the-art fairness interventions are\neffective at removing epistemic discrimination on standard (overused) tabular\ndatasets. However, when data has missing values, there is still significant\nroom for improvement in handling aleatoric discrimination.\n", "date": "2023-01-27", "categories": []}
{"id": "2305.12671", "title": "Transferring Fairness using Multi-Task Learning with Limited Demographic\n  Information", "abstract": "  Training supervised machine learning systems with a fairness loss can improve\nprediction fairness across different demographic groups. However, doing so\nrequires demographic annotations for training data, without which we cannot\nproduce debiased classifiers for most tasks. Drawing inspiration from transfer\nlearning methods, we investigate whether we can utilize demographic data from a\nrelated task to improve the fairness of a target task. We adapt a single-task\nfairness loss to a multi-task setting to exploit demographic labels from a\nrelated task in debiasing a target task and demonstrate that demographic\nfairness objectives transfer fairness within a multi-task framework.\nAdditionally, we show that this approach enables intersectional fairness by\ntransferring between two datasets with different single-axis demographics. We\nexplore different data domains to show how our loss can improve fairness\ndomains and tasks.\n", "date": "2023-05-22", "categories": []}
{"id": "2402.07329", "title": "The Bias of Harmful Label Associations in Vision-Language Models", "abstract": "  Despite the remarkable performance of foundation vision-language models, the\nshared representation space for text and vision can also encode harmful label\nassociations detrimental to fairness. While prior work has uncovered bias in\nvision-language models' (VLMs) classification performance across geography,\nwork has been limited along the important axis of harmful label associations\ndue to a lack of rich, labeled data. In this work, we investigate harmful label\nassociations in the recently released Casual Conversations datasets containing\nmore than 70,000 videos. We study bias in the frequency of harmful label\nassociations across self-provided labels for age, gender, apparent skin tone,\nand physical adornments across several leading VLMs. We find that VLMs are\n$4-7$x more likely to harmfully classify individuals with darker skin tones. We\nalso find scaling transformer encoder model size leads to higher confidence in\nharmful predictions. Finally, we find improvements on standard vision tasks\nacross VLMs does not address disparities in harmful label associations.\n", "date": "2024-02-11", "categories": []}
{"id": "2307.08199", "title": "Unbiased Image Synthesis via Manifold Guidance in Diffusion Models", "abstract": "  Diffusion Models are a potent class of generative models capable of producing\nhigh-quality images. However, they often inadvertently favor certain data\nattributes, undermining the diversity of generated images. This issue is\nstarkly apparent in skewed datasets like CelebA, where the initial dataset\ndisproportionately favors females over males by 57.9%, this bias amplified in\ngenerated data where female representation outstrips males by 148%. In\nresponse, we propose a plug-and-play method named Manifold Guidance Sampling,\nwhich is also the first unsupervised method to mitigate bias issue in DDPMs.\nLeveraging the inherent structure of the data manifold, this method steers the\nsampling process towards a more uniform distribution, effectively dispersing\nthe clustering of biased data. Without the need for modifying the existing\nmodel or additional training, it significantly mitigates data bias and enhances\nthe quality and unbiasedness of the generated images.\n", "date": "2023-07-17", "categories": []}
{"id": "2404.09391", "title": "Privacy at a Price: Exploring its Dual Impact on AI Fairness", "abstract": "  The worldwide adoption of machine learning (ML) and deep learning models,\nparticularly in critical sectors, such as healthcare and finance, presents\nsubstantial challenges in maintaining individual privacy and fairness. These\ntwo elements are vital to a trustworthy environment for learning systems. While\nnumerous studies have concentrated on protecting individual privacy through\ndifferential privacy (DP) mechanisms, emerging research indicates that\ndifferential privacy in machine learning models can unequally impact separate\ndemographic subgroups regarding prediction accuracy. This leads to a fairness\nconcern, and manifests as biased performance. Although the prevailing view is\nthat enhancing privacy intensifies fairness disparities, a smaller, yet\nsignificant, subset of research suggests the opposite view. In this article,\nwith extensive evaluation results, we demonstrate that the impact of\ndifferential privacy on fairness is not monotonous. Instead, we observe that\nthe accuracy disparity initially grows as more DP noise (enhanced privacy) is\nadded to the ML process, but subsequently diminishes at higher privacy levels\nwith even more noise. Moreover, implementing gradient clipping in the\ndifferentially private stochastic gradient descent ML method can mitigate the\nnegative impact of DP noise on fairness. This mitigation is achieved by\nmoderating the disparity growth through a lower clipping threshold.\n", "date": "2024-04-15", "categories": []}
{"id": "2404.09664", "title": "Closing the Gap in the Trade-off between Fair Representations and\n  Accuracy", "abstract": "  The rapid developments of various machine learning models and their\ndeployments in several applications has led to discussions around the\nimportance of looking beyond the accuracies of these models. Fairness of such\nmodels is one such aspect that is deservedly gaining more attention. In this\nwork, we analyse the natural language representations of documents and\nsentences (i.e., encodings) for any embedding-level bias that could potentially\nalso affect the fairness of the downstream tasks that rely on them. We identify\nbias in these encodings either towards or against different sub-groups based on\nthe difference in their reconstruction errors along various subsets of\nprincipal components. We explore and recommend ways to mitigate such bias in\nthe encodings while also maintaining a decent accuracy in classification models\nthat use them.\n", "date": "2024-04-15", "categories": []}
{"id": "2310.02492", "title": "FairVision: Equitable Deep Learning for Eye Disease Screening via Fair\n  Identity Scaling", "abstract": "  Equity in AI for healthcare is crucial due to its direct impact on human\nwell-being. Despite advancements in 2D medical imaging fairness, the fairness\nof 3D models remains underexplored, hindered by the small sizes of 3D fairness\ndatasets. Since 3D imaging surpasses 2D imaging in SOTA clinical care, it is\ncritical to understand the fairness of these 3D models. To address this\nresearch gap, we conduct the first comprehensive study on the fairness of 3D\nmedical imaging models across multiple protected attributes. Our investigation\nspans both 2D and 3D models and evaluates fairness across five architectures on\nthree common eye diseases, revealing significant biases across race, gender,\nand ethnicity. To alleviate these biases, we propose a novel fair identity\nscaling (FIS) method that improves both overall performance and fairness,\noutperforming various SOTA fairness methods. Moreover, we release\nHarvard-FairVision, the first large-scale medical fairness dataset with 30,000\nsubjects featuring both 2D and 3D imaging data and six demographic identity\nattributes. Harvard-FairVision provides labels for three major eye disorders\naffecting about 380 million people worldwide, serving as a valuable resource\nfor both 2D and 3D fairness learning. Our code and dataset are publicly\naccessible at\n\\url{https://ophai.hms.harvard.edu/datasets/harvard-fairvision30k}.\n", "date": "2023-10-03", "categories": []}
{"id": "2310.17256", "title": "fairret: a Framework for Differentiable Fairness Regularization Terms", "abstract": "  Current fairness toolkits in machine learning only admit a limited range of\nfairness definitions and have seen little integration with automatic\ndifferentiation libraries, despite the central role these libraries play in\nmodern machine learning pipelines.\n  We introduce a framework of fairness regularization terms (fairrets) which\nquantify bias as modular, flexible objectives that are easily integrated in\nautomatic differentiation pipelines. By employing a general definition of\nfairness in terms of linear-fractional statistics, a wide class of fairrets can\nbe computed efficiently. Experiments show the behavior of their gradients and\ntheir utility in enforcing fairness with minimal loss of predictive power\ncompared to baselines. Our contribution includes a PyTorch implementation of\nthe fairret framework.\n", "date": "2023-10-26", "categories": []}
{"id": "2404.06619", "title": "FairPair: A Robust Evaluation of Biases in Language Models through\n  Paired Perturbations", "abstract": "  The accurate evaluation of differential treatment in language models to\nspecific groups is critical to ensuring a positive and safe user experience. An\nideal evaluation should have the properties of being robust, extendable to new\ngroups or attributes, and being able to capture biases that appear in typical\nusage (rather than just extreme, rare cases). Relatedly, bias evaluation should\nsurface not only egregious biases but also ones that are subtle and\ncommonplace, such as a likelihood for talking about appearances with regard to\nwomen. We present FairPair, an evaluation framework for assessing differential\ntreatment that occurs during ordinary usage. FairPair operates through\ncounterfactual pairs, but crucially, the paired continuations are grounded in\nthe same demographic group, which ensures equivalent comparison. Additionally,\nunlike prior work, our method factors in the inherent variability that comes\nfrom the generation process itself by measuring the sampling variability. We\npresent an evaluation of several commonly used generative models and a\nqualitative analysis that indicates a preference for discussing family and\nhobbies with regard to women.\n", "date": "2024-04-09", "categories": []}
{"id": "2404.06717", "title": "Racial/Ethnic Categories in AI and Algorithmic Fairness: Why They Matter\n  and What They Represent", "abstract": "  Racial diversity has become increasingly discussed within the AI and\nalgorithmic fairness literature, yet little attention is focused on justifying\nthe choices of racial categories and understanding how people are racialized\ninto these chosen racial categories. Even less attention is given to how racial\ncategories shift and how the racialization process changes depending on the\ncontext of a dataset or model. An unclear understanding of \\textit{who}\ncomprises the racial categories chosen and \\textit{how} people are racialized\ninto these categories can lead to varying interpretations of these categories.\nThese varying interpretations can lead to harm when the understanding of racial\ncategories and the racialization process is misaligned from the actual\nracialization process and racial categories used. Harm can also arise if the\nracialization process and racial categories used are irrelevant or do not exist\nin the context they are applied.\n  In this paper, we make two contributions. First, we demonstrate how racial\ncategories with unclear assumptions and little justification can lead to\nvarying datasets that poorly represent groups obfuscated or unrepresented by\nthe given racial categories and models that perform poorly on these groups.\nSecond, we develop a framework, CIRCSheets, for documenting the choices and\nassumptions in choosing racial categories and the process of racialization into\nthese categories to facilitate transparency in understanding the processes and\nassumptions made by dataset or model developers when selecting or using these\nracial categories.\n", "date": "2024-04-10", "categories": []}
{"id": "2309.06717", "title": "Bias Amplification Enhances Minority Group Performance", "abstract": "  Neural networks produced by standard training are known to suffer from poor\naccuracy on rare subgroups despite achieving high accuracy on average, due to\nthe correlations between certain spurious features and labels. Previous\napproaches based on worst-group loss minimization (e.g. Group-DRO) are\neffective in improving worse-group accuracy but require expensive group\nannotations for all the training samples. In this paper, we focus on the more\nchallenging and realistic setting where group annotations are only available on\na small validation set or are not available at all. We propose BAM, a novel\ntwo-stage training algorithm: in the first stage, the model is trained using a\nbias amplification scheme via introducing a learnable auxiliary variable for\neach training sample; in the second stage, we upweight the samples that the\nbias-amplified model misclassifies, and then continue training the same model\non the reweighted dataset. Empirically, BAM achieves competitive performance\ncompared with existing methods evaluated on spurious correlation benchmarks in\ncomputer vision and natural language processing. Moreover, we find a simple\nstopping criterion based on minimum class accuracy difference that can remove\nthe need for group annotations, with little or no loss in worst-group accuracy.\nWe perform extensive analyses and ablations to verify the effectiveness and\nrobustness of our algorithm in varying class and group imbalance ratios.\n", "date": "2023-09-13", "categories": []}
{"id": "2310.08617", "title": "The Impact of Explanations on Fairness in Human-AI Decision-Making:\n  Protected vs Proxy Features", "abstract": "  AI systems have been known to amplify biases in real-world data. Explanations\nmay help human-AI teams address these biases for fairer decision-making.\nTypically, explanations focus on salient input features. If a model is biased\nagainst some protected group, explanations may include features that\ndemonstrate this bias, but when biases are realized through proxy features, the\nrelationship between this proxy feature and the protected one may be less clear\nto a human. In this work, we study the effect of the presence of protected and\nproxy features on participants' perception of model fairness and their ability\nto improve demographic parity over an AI alone. Further, we examine how\ndifferent treatments -- explanations, model bias disclosure and proxy\ncorrelation disclosure -- affect fairness perception and parity. We find that\nexplanations help people detect direct but not indirect biases. Additionally,\nregardless of bias type, explanations tend to increase agreement with model\nbiases. Disclosures can help mitigate this effect for indirect biases,\nimproving both unfairness recognition and decision-making fairness. We hope\nthat our findings can help guide further research into advancing explanations\nin support of fair human-AI decision-making.\n", "date": "2023-10-12", "categories": []}
{"id": "2310.11211", "title": "Understanding Fairness Surrogate Functions in Algorithmic Fairness", "abstract": "  It has been observed that machine learning algorithms exhibit biased\npredictions against certain population groups. To mitigate such bias while\nachieving comparable accuracy, a promising approach is to introduce surrogate\nfunctions of the concerned fairness definition and solve a constrained\noptimization problem. However, it is intriguing in previous work that such\nfairness surrogate functions may yield unfair results and high instability. In\nthis work, in order to deeply understand them, taking a widely used fairness\ndefinition--demographic parity as an example, we show that there is a\nsurrogate-fairness gap between the fairness definition and the fairness\nsurrogate function. Also, the theoretical analysis and experimental results\nabout the gap motivate us that the fairness and stability will be affected by\nthe points far from the decision boundary, which is the large margin points\nissue investigated in this paper. To address it, we propose the general sigmoid\nsurrogate to simultaneously reduce both the surrogate-fairness gap and the\nvariance, and offer a rigorous fairness and stability upper bound.\nInterestingly, the theory also provides insights into two important issues that\ndeal with the large margin points as well as obtaining a more balanced dataset\nare beneficial to fairness and stability. Furthermore, we elaborate a novel and\ngeneral algorithm called Balanced Surrogate, which iteratively reduces the gap\nto mitigate unfairness. Finally, we provide empirical evidence showing that our\nmethods consistently improve fairness and stability while maintaining accuracy\ncomparable to the baselines in three real-world datasets.\n", "date": "2023-10-17", "categories": []}
{"id": "2404.06090", "title": "Fair Graph Neural Network with Supervised Contrastive Regularization", "abstract": "  In recent years, Graph Neural Networks (GNNs) have made significant\nadvancements, particularly in tasks such as node classification, link\nprediction, and graph representation. However, challenges arise from biases\nthat can be hidden not only in the node attributes but also in the connections\nbetween entities. Therefore, ensuring fairness in graph neural network learning\nhas become a critical problem. To address this issue, we propose a novel model\nfor training fairness-aware GNN, which enhances the Counterfactual Augmented\nFair Graph Neural Network Framework (CAF). Our approach integrates Supervised\nContrastive Loss and Environmental Loss to enhance both accuracy and fairness.\nExperimental validation on three real datasets demonstrates the superiority of\nour proposed model over CAF and several other existing graph-based learning\nmethods.\n", "date": "2024-04-09", "categories": []}
{"id": "2312.03259", "title": "f-FERM: A Scalable Framework for Robust Fair Empirical Risk Minimization", "abstract": "  Training and deploying machine learning models that meet fairness criteria\nfor protected groups are fundamental in modern artificial intelligence. While\nnumerous constraints and regularization terms have been proposed in the\nliterature to promote fairness in machine learning tasks, most of these methods\nare not amenable to stochastic optimization due to the complex and nonlinear\nstructure of constraints and regularizers. Here, the term \"stochastic\" refers\nto the ability of the algorithm to work with small mini-batches of data.\nMotivated by the limitation of existing literature, this paper presents a\nunified stochastic optimization framework for fair empirical risk minimization\nbased on f-divergence measures (f-FERM). The proposed stochastic algorithm\nenjoys theoretical convergence guarantees. In addition, our experiments\ndemonstrate the superiority of fairness-accuracy tradeoffs offered by f-FERM\nfor almost all batch sizes (ranging from full-batch to batch size of one).\nMoreover, we show that our framework can be extended to the case where there is\na distribution shift from training to the test data. Our extension is based on\na distributionally robust optimization reformulation of f-FERM objective under\n$L_p$ norms as uncertainty sets. Again, in this distributionally robust\nsetting, f-FERM not only enjoys theoretical convergence guarantees but also\noutperforms other baselines in the literature in the tasks involving\ndistribution shifts. An efficient stochastic implementation of $f$-FERM is\npublicly available.\n", "date": "2023-12-06", "categories": []}
{"id": "2403.09516", "title": "Leveraging Prototypical Representations for Mitigating Social Bias\n  without Demographic Information", "abstract": "  Mitigating social biases typically requires identifying the social groups\nassociated with each data sample. In this paper, we present DAFair, a novel\napproach to address social bias in language models. Unlike traditional methods\nthat rely on explicit demographic labels, our approach does not require any\nsuch information. Instead, we leverage predefined prototypical demographic\ntexts and incorporate a regularization term during the fine-tuning process to\nmitigate bias in the model's representations. Our empirical results across two\ntasks and two models demonstrate the effectiveness of our method compared to\nprevious approaches that do not rely on labeled data. Moreover, with limited\ndemographic-annotated data, our approach outperforms common debiasing\napproaches.\n", "date": "2024-03-14", "categories": []}
{"id": "2404.03800", "title": "Learning Social Fairness Preferences from Non-Expert Stakeholder\n  Opinions in Kidney Placement", "abstract": "  Modern kidney placement incorporates several intelligent recommendation\nsystems which exhibit social discrimination due to biases inherited from\ntraining data. Although initial attempts were made in the literature to study\nalgorithmic fairness in kidney placement, these methods replace true outcomes\nwith surgeons' decisions due to the long delays involved in recording such\noutcomes reliably. However, the replacement of true outcomes with surgeons'\ndecisions disregards expert stakeholders' biases as well as social opinions of\nother stakeholders who do not possess medical expertise. This paper alleviates\nthe latter concern and designs a novel fairness feedback survey to evaluate an\nacceptance rate predictor (ARP) that predicts a kidney's acceptance rate in a\ngiven kidney-match pair. The survey is launched on Prolific, a crowdsourcing\nplatform, and public opinions are collected from 85 anonymous crowd\nparticipants. A novel social fairness preference learning algorithm is proposed\nbased on minimizing social feedback regret computed using a novel logit-based\nfairness feedback model. The proposed model and learning algorithm are both\nvalidated using simulation experiments as well as Prolific data. Public\npreferences towards group fairness notions in the context of kidney placement\nhave been estimated and discussed in detail. The specific ARP tested in the\nProlific survey has been deemed fair by the participants.\n", "date": "2024-04-04", "categories": []}
{"id": "2308.01923", "title": "Fairness Improvement with Multiple Protected Attributes: How Far Are We?", "abstract": "  Existing research mostly improves the fairness of Machine Learning (ML)\nsoftware regarding a single protected attribute at a time, but this is\nunrealistic given that many users have multiple protected attributes. This\npaper conducts an extensive study of fairness improvement regarding multiple\nprotected attributes, covering 11 state-of-the-art fairness improvement\nmethods. We analyze the effectiveness of these methods with different datasets,\nmetrics, and ML models when considering multiple protected attributes. The\nresults reveal that improving fairness for a single protected attribute can\nlargely decrease fairness regarding unconsidered protected attributes. This\ndecrease is observed in up to 88.3% of scenarios (57.5% on average). More\nsurprisingly, we find little difference in accuracy loss when considering\nsingle and multiple protected attributes, indicating that accuracy can be\nmaintained in the multiple-attribute paradigm. However, the effect on F1-score\nwhen handling two protected attributes is about twice that of a single\nattribute. This has important implications for future fairness research:\nreporting only accuracy as the ML performance metric, which is currently common\nin the literature, is inadequate.\n", "date": "2023-07-25", "categories": []}
{"id": "2404.03086", "title": "Auditing the Use of Language Models to Guide Hiring Decisions", "abstract": "  Regulatory efforts to protect against algorithmic bias have taken on\nincreased urgency with rapid advances in large language models (LLMs), which\nare machine learning models that can achieve performance rivaling human experts\non a wide array of tasks. A key theme of these initiatives is algorithmic\n\"auditing,\" but current regulations -- as well as the scientific literature --\nprovide little guidance on how to conduct these assessments. Here we propose\nand investigate one approach for auditing algorithms: correspondence\nexperiments, a widely applied tool for detecting bias in human judgements. In\nthe employment context, correspondence experiments aim to measure the extent to\nwhich race and gender impact decisions by experimentally manipulating elements\nof submitted application materials that suggest an applicant's demographic\ntraits, such as their listed name. We apply this method to audit candidate\nassessments produced by several state-of-the-art LLMs, using a novel corpus of\napplications to K-12 teaching positions in a large public school district. We\nfind evidence of moderate race and gender disparities, a pattern largely robust\nto varying the types of application material input to the models, as well as\nthe framing of the task to the LLMs. We conclude by discussing some important\nlimitations of correspondence experiments for auditing algorithms.\n", "date": "2024-04-03", "categories": []}
{"id": "2307.05862", "title": "Ecosystem-level Analysis of Deployed Machine Learning Reveals\n  Homogeneous Outcomes", "abstract": "  Machine learning is traditionally studied at the model level: researchers\nmeasure and improve the accuracy, robustness, bias, efficiency, and other\ndimensions of specific models. In practice, the societal impact of machine\nlearning is determined by the surrounding context of machine learning\ndeployments. To capture this, we introduce ecosystem-level analysis: rather\nthan analyzing a single model, we consider the collection of models that are\ndeployed in a given context. For example, ecosystem-level analysis in hiring\nrecognizes that a job candidate's outcomes are not only determined by a single\nhiring algorithm or firm but instead by the collective decisions of all the\nfirms they applied to. Across three modalities (text, images, speech) and 11\ndatasets, we establish a clear trend: deployed machine learning is prone to\nsystemic failure, meaning some users are exclusively misclassified by all\nmodels available. Even when individual models improve at the population level\nover time, we find these improvements rarely reduce the prevalence of systemic\nfailure. Instead, the benefits of these improvements predominantly accrue to\nindividuals who are already correctly classified by other models. In light of\nthese trends, we consider medical imaging for dermatology where the costs of\nsystemic failure are especially high. While traditional analyses reveal racial\nperformance disparities for both models and humans, ecosystem-level analysis\nreveals new forms of racial disparity in model predictions that do not present\nin human predictions. These examples demonstrate ecosystem-level analysis has\nunique strengths for characterizing the societal impact of machine learning.\n", "date": "2023-07-12", "categories": []}
{"id": "2310.14607", "title": "Confronting LLMs with Traditional ML: Rethinking the Fairness of Large\n  Language Models in Tabular Classifications", "abstract": "  Recent literature has suggested the potential of using large language models\n(LLMs) to make classifications for tabular tasks. However, LLMs have been shown\nto exhibit harmful social biases that reflect the stereotypes and inequalities\npresent in society. To this end, as well as the widespread use of tabular data\nin many high-stake applications, it is important to explore the following\nquestions: what sources of information do LLMs draw upon when making\nclassifications for tabular tasks; whether and to what extent are LLM\nclassifications for tabular data influenced by social biases and stereotypes;\nand what are the consequential implications for fairness?\n  Through a series of experiments, we delve into these questions and show that\nLLMs tend to inherit social biases from their training data which significantly\nimpact their fairness in tabular classification tasks. Furthermore, our\ninvestigations show that in the context of bias mitigation, though in-context\nlearning and finetuning have a moderate effect, the fairness metric gap between\ndifferent subgroups is still larger than that in traditional machine learning\nmodels, such as Random Forest and shallow Neural Networks. This observation\nemphasizes that the social biases are inherent within the LLMs themselves and\ninherited from their pretraining corpus, not only from the downstream task\ndatasets. Besides, we demonstrate that label-flipping of in-context examples\ncan significantly reduce biases, further highlighting the presence of inherent\nbias within LLMs.\n", "date": "2023-10-23", "categories": []}
{"id": "2310.15398", "title": "\"One-Size-Fits-All\"? Examining Expectations around What Constitute\n  \"Fair\" or \"Good\" NLG System Behaviors", "abstract": "  Fairness-related assumptions about what constitute appropriate NLG system\nbehaviors range from invariance, where systems are expected to behave\nidentically for social groups, to adaptation, where behaviors should instead\nvary across them. To illuminate tensions around invariance and adaptation, we\nconduct five case studies, in which we perturb different types of\nidentity-related language features (names, roles, locations, dialect, and\nstyle) in NLG system inputs. Through these cases studies, we examine people's\nexpectations of system behaviors, and surface potential caveats of these\ncontrasting yet commonly held assumptions. We find that motivations for\nadaptation include social norms, cultural differences, feature-specific\ninformation, and accommodation; in contrast, motivations for invariance include\nperspectives that favor prescriptivism, view adaptation as unnecessary or too\ndifficult for NLG systems to do appropriately, and are wary of false\nassumptions. Our findings highlight open challenges around what constitute\n\"fair\" or \"good\" NLG system behaviors.\n", "date": "2023-10-23", "categories": []}
{"id": "2404.02650", "title": "Towards detecting unanticipated bias in Large Language Models", "abstract": "  Over the last year, Large Language Models (LLMs) like ChatGPT have become\nwidely available and have exhibited fairness issues similar to those in\nprevious machine learning systems. Current research is primarily focused on\nanalyzing and quantifying these biases in training data and their impact on the\ndecisions of these models, alongside developing mitigation strategies. This\nresearch largely targets well-known biases related to gender, race, ethnicity,\nand language. However, it is clear that LLMs are also affected by other, less\nobvious implicit biases. The complex and often opaque nature of these models\nmakes detecting such biases challenging, yet this is crucial due to their\npotential negative impact in various applications. In this paper, we explore\nnew avenues for detecting these unanticipated biases in LLMs, focusing\nspecifically on Uncertainty Quantification and Explainable AI methods. These\napproaches aim to assess the certainty of model decisions and to make the\ninternal decision-making processes of LLMs more transparent, thereby\nidentifying and understanding biases that are not immediately apparent. Through\nthis research, we aim to contribute to the development of fairer and more\ntransparent AI systems.\n", "date": "2024-04-03", "categories": []}
{"id": "2305.09330", "title": "Consumer-side Fairness in Recommender Systems: A Systematic Survey of\n  Methods and Evaluation", "abstract": "  In the current landscape of ever-increasing levels of digitalization, we are\nfacing major challenges pertaining to scalability. Recommender systems have\nbecome irreplaceable both for helping users navigate the increasing amounts of\ndata and, conversely, aiding providers in marketing products to interested\nusers. The growing awareness of discrimination in machine learning methods has\nrecently motivated both academia and industry to research how fairness can be\nensured in recommender systems. For recommender systems, such issues are well\nexemplified by occupation recommendation, where biases in historical data may\nlead to recommender systems relating one gender to lower wages or to the\npropagation of stereotypes. In particular, consumer-side fairness, which\nfocuses on mitigating discrimination experienced by users of recommender\nsystems, has seen a vast number of diverse approaches for addressing different\ntypes of discrimination. The nature of said discrimination depends on the\nsetting and the applied fairness interpretation, of which there are many\nvariations. This survey serves as a systematic overview and discussion of the\ncurrent research on consumer-side fairness in recommender systems. To that end,\na novel taxonomy based on high-level fairness interpretation is proposed and\nused to categorize the research and their proposed fairness evaluation metrics.\nFinally, we highlight some suggestions for the future direction of the field.\n", "date": "2023-05-16", "categories": []}
{"id": "2312.02957", "title": "Classification for everyone : Building geography agnostic models for\n  fairer recognition", "abstract": "  In this paper, we analyze different methods to mitigate inherent geographical\nbiases present in state of the art image classification models. We first\nquantitatively present this bias in two datasets - The Dollar Street Dataset\nand ImageNet, using images with location information. We then present different\nmethods which can be employed to reduce this bias. Finally, we analyze the\neffectiveness of the different techniques on making these models more robust to\ngeographical locations of the images.\n", "date": "2023-12-05", "categories": []}
{"id": "2404.01521", "title": "Fair MP-BOOST: Fair and Interpretable Minipatch Boosting", "abstract": "  Ensemble methods, particularly boosting, have established themselves as\nhighly effective and widely embraced machine learning techniques for tabular\ndata. In this paper, we aim to leverage the robust predictive power of\ntraditional boosting methods while enhancing fairness and interpretability. To\nachieve this, we develop Fair MP-Boost, a stochastic boosting scheme that\nbalances fairness and accuracy by adaptively learning features and observations\nduring training. Specifically, Fair MP-Boost sequentially samples small subsets\nof observations and features, termed minipatches (MP), according to adaptively\nlearned feature and observation sampling probabilities. We devise these\nprobabilities by combining loss functions, or by combining feature importance\nscores to address accuracy and fairness simultaneously. Hence, Fair MP-Boost\nprioritizes important and fair features along with challenging instances, to\nselect the most relevant minipatches for learning. The learned probability\ndistributions also yield intrinsic interpretations of feature importance and\nimportant observations in Fair MP-Boost. Through empirical evaluation of\nsimulated and benchmark datasets, we showcase the interpretability, accuracy,\nand fairness of Fair MP-Boost.\n", "date": "2024-04-01", "categories": []}
{"id": "2404.01608", "title": "FAIRM: Learning invariant representations for algorithmic fairness and\n  domain generalization with minimax optimality", "abstract": "  Machine learning methods often assume that the test data have the same\ndistribution as the training data. However, this assumption may not hold due to\nmultiple levels of heterogeneity in applications, raising issues in algorithmic\nfairness and domain generalization. In this work, we address the problem of\nfair and generalizable machine learning by invariant principles. We propose a\ntraining environment-based oracle, FAIRM, which has desirable fairness and\ndomain generalization properties under a diversity-type condition. We then\nprovide an empirical FAIRM with finite-sample theoretical guarantees under weak\ndistributional assumptions. We then develop efficient algorithms to realize\nFAIRM in linear models and demonstrate the nonasymptotic performance with\nminimax optimality. We evaluate our method in numerical experiments with\nsynthetic data and MNIST data and show that it outperforms its counterparts.\n", "date": "2024-04-02", "categories": []}
{"id": "2404.01857", "title": "Detecting Gender Bias in Course Evaluations", "abstract": "  An outtake from the findnings of a master thesis studying gender bias in\ncourse evaluations through the lense of machine learning and nlp. We use\ndifferent methods to examine and explore the data and find differences in what\nstudents write about courses depending on gender of the examiner. Data from\nEnglish and Swedish courses are evaluated and compared, in order to capture\nmore nuance in the gender bias that might be found. Here we present the results\nfrom the work so far, but this is an ongoing project and there is more work to\ndo.\n", "date": "2024-04-02", "categories": []}
{"id": "2311.07884", "title": "Fair Abstractive Summarization of Diverse Perspectives", "abstract": "  People from different social and demographic groups express diverse\nperspectives and conflicting opinions on a broad set of topics such as product\nreviews, healthcare, law, and politics. A fair summary should provide a\ncomprehensive coverage of diverse perspectives without underrepresenting\ncertain groups. However, current work in summarization metrics and Large\nLanguage Models (LLMs) evaluation has not explored fair abstractive\nsummarization. In this paper, we systematically investigate fair abstractive\nsummarization for user-generated data. We first formally define fairness in\nabstractive summarization as not underrepresenting perspectives of any groups\nof people, and we propose four reference-free automatic metrics by measuring\nthe differences between target and source perspectives. We evaluate nine LLMs,\nincluding three GPT models, four LLaMA models, PaLM 2, and Claude, on six\ndatasets collected from social media, online reviews, and recorded transcripts.\nExperiments show that both the model-generated and the human-written reference\nsummaries suffer from low fairness. We conduct a comprehensive analysis of the\ncommon factors influencing fairness and propose three simple but effective\nmethods to alleviate unfair summarization. Our dataset and code are available\nat https://github.com/psunlpgroup/FairSumm.\n", "date": "2023-11-14", "categories": []}
{"id": "2403.19165", "title": "Evaluating Fair Feature Selection in Machine Learning for Healthcare", "abstract": "  With the universal adoption of machine learning in healthcare, the potential\nfor the automation of societal biases to further exacerbate health disparities\nposes a significant risk. We explore algorithmic fairness from the perspective\nof feature selection. Traditional feature selection methods identify features\nfor better decision making by removing resource-intensive, correlated, or\nnon-relevant features but overlook how these factors may differ across\nsubgroups. To counter these issues, we evaluate a fair feature selection method\nthat considers equal importance to all demographic groups. We jointly\nconsidered a fairness metric and an error metric within the feature selection\nprocess to ensure a balance between minimizing both bias and global\nclassification error. We tested our approach on three publicly available\nhealthcare datasets. On all three datasets, we observed improvements in\nfairness metrics coupled with a minimal degradation of balanced accuracy. Our\napproach addresses both distributive and procedural fairness within the fair\nmachine learning context.\n", "date": "2024-03-28", "categories": []}
{"id": "2404.00463", "title": "Addressing Both Statistical and Causal Gender Fairness in NLP Models", "abstract": "  Statistical fairness stipulates equivalent outcomes for every protected\ngroup, whereas causal fairness prescribes that a model makes the same\nprediction for an individual regardless of their protected characteristics.\nCounterfactual data augmentation (CDA) is effective for reducing bias in NLP\nmodels, yet models trained with CDA are often evaluated only on metrics that\nare closely tied to the causal fairness notion; similarly, sampling-based\nmethods designed to promote statistical fairness are rarely evaluated for\ncausal fairness. In this work, we evaluate both statistical and causal\ndebiasing methods for gender bias in NLP models, and find that while such\nmethods are effective at reducing bias as measured by the targeted metric, they\ndo not necessarily improve results on other bias metrics. We demonstrate that\ncombinations of statistical and causal debiasing techniques are able to reduce\nbias measured through both types of metrics.\n", "date": "2024-03-30", "categories": []}
{"id": "2403.19738", "title": "MIST: Mitigating Intersectional Bias with Disentangled Cross-Attention\n  Editing in Text-to-Image Diffusion Models", "abstract": "  Diffusion-based text-to-image models have rapidly gained popularity for their\nability to generate detailed and realistic images from textual descriptions.\nHowever, these models often reflect the biases present in their training data,\nespecially impacting marginalized groups. While prior efforts to debias\nlanguage models have focused on addressing specific biases, such as racial or\ngender biases, efforts to tackle intersectional bias have been limited.\nIntersectional bias refers to the unique form of bias experienced by\nindividuals at the intersection of multiple social identities. Addressing\nintersectional bias is crucial because it amplifies the negative effects of\ndiscrimination based on race, gender, and other identities. In this paper, we\nintroduce a method that addresses intersectional bias in diffusion-based\ntext-to-image models by modifying cross-attention maps in a disentangled\nmanner. Our approach utilizes a pre-trained Stable Diffusion model, eliminates\nthe need for an additional set of reference images, and preserves the original\nquality for unaltered concepts. Comprehensive experiments demonstrate that our\nmethod surpasses existing approaches in mitigating both single and\nintersectional biases across various attributes. We make our source code and\ndebiased models for various attributes available to encourage fairness in\ngenerative models and to support further research.\n", "date": "2024-03-28", "categories": []}
{"id": "2305.09535", "title": "What's the Problem, Linda? The Conjunction Fallacy as a Fairness Problem", "abstract": "  The field of Artificial Intelligence (AI) is focusing on creating automated\ndecision-making (ADM) systems that operate as close as possible to human-like\nintelligence. This effort has pushed AI researchers into exploring cognitive\nfields like psychology. The work of Daniel Kahneman and the late Amos Tversky\non biased human decision-making, including the study of the conjunction\nfallacy, has experienced a second revival because of this. Under the\nconjunction fallacy a human decision-maker will go against basic probability\nlaws and rank as more likely a conjunction over one of its parts. It has been\nproven overtime through a set of experiments with the Linda Problem being the\nmost famous one. Although this interdisciplinary effort is welcomed, we fear\nthat AI researchers ignore the driving force behind the conjunction fallacy as\ncaptured by the Linda Problem: the fact that Linda must be stereotypically\ndescribed as a woman. In this paper we revisit the Linda Problem and formulate\nit as a fairness problem. In doing so we introduce perception as a parameter of\ninterest through the structural causal perception framework. Using an\nillustrative decision-making example, we showcase the proposed conceptual\nframework and its potential impact for developing fair ADM systems.\n", "date": "2023-05-16", "categories": []}
{"id": "2403.19057", "title": "Equity in Healthcare: Analyzing Disparities in Machine Learning\n  Predictions of Diabetic Patient Readmissions", "abstract": "  This study investigates how machine learning (ML) models can predict hospital\nreadmissions for diabetic patients fairly and accurately across different\ndemographics (age, gender, race). We compared models like Deep Learning,\nGeneralized Linear Models, Gradient Boosting Machines (GBM), and Naive Bayes.\nGBM stood out with an F1-score of 84.3% and accuracy of 82.2%, accurately\npredicting readmissions across demographics. A fairness analysis was conducted\nacross all the models. GBM minimized disparities in predictions, achieving\nbalanced results across genders and races. It showed low False Discovery Rates\n(FDR) (6-7%) and False Positive Rates (FPR) (5%) for both genders.\nAdditionally, FDRs remained low for racial groups, such as African Americans\n(8%) and Asians (7%). Similarly, FPRs were consistent across age groups (4%)\nfor both patients under 40 and those above 40, indicating its precision and\nability to reduce bias. These findings emphasize the importance of choosing ML\nmodels carefully to ensure both accuracy and fairness for all patients. By\nshowcasing effectiveness of various models with fairness metrics, this study\npromotes personalized medicine and the need for fair ML algorithms in\nhealthcare. This can ultimately reduce disparities and improve outcomes for\ndiabetic patients of all backgrounds.\n", "date": "2024-03-27", "categories": []}
{"id": "2301.11104", "title": "Discovering and Mitigating Visual Biases through Keyword Explanation", "abstract": "  Addressing biases in computer vision models is crucial for real-world AI\ndeployments. However, mitigating visual biases is challenging due to their\nunexplainable nature, often identified indirectly through visualization or\nsample statistics, which necessitates additional human supervision for\ninterpretation. To tackle this issue, we propose the Bias-to-Text (B2T)\nframework, which interprets visual biases as keywords. Specifically, we extract\ncommon keywords from the captions of mispredicted images to identify potential\nbiases in the model. We then validate these keywords by measuring their\nsimilarity to the mispredicted images using a vision-language scoring model.\nThe keyword explanation form of visual bias offers several advantages, such as\na clear group naming for bias discovery and a natural extension for debiasing\nusing these group names. Our experiments demonstrate that B2T can identify\nknown biases, such as gender bias in CelebA, background bias in Waterbirds, and\ndistribution shifts in ImageNet-R/C. Additionally, B2T uncovers novel biases in\nlarger datasets, such as Dollar Street and ImageNet. For example, we discovered\na contextual bias between \"bee\" and \"flower\" in ImageNet. We also highlight\nvarious applications of B2T keywords, including debiased training, CLIP\nprompting, and model comparison.\n", "date": "2023-01-26", "categories": []}
{"id": "2309.06627", "title": "A Sequentially Fair Mechanism for Multiple Sensitive Attributes", "abstract": "  In the standard use case of Algorithmic Fairness, the goal is to eliminate\nthe relationship between a sensitive variable and a corresponding score.\nThroughout recent years, the scientific community has developed a host of\ndefinitions and tools to solve this task, which work well in many practical\napplications. However, the applicability and effectivity of these tools and\ndefinitions becomes less straightfoward in the case of multiple sensitive\nattributes. To tackle this issue, we propose a sequential framework, which\nallows to progressively achieve fairness across a set of sensitive features. We\naccomplish this by leveraging multi-marginal Wasserstein barycenters, which\nextends the standard notion of Strong Demographic Parity to the case with\nmultiple sensitive characteristics. This method also provides a closed-form\nsolution for the optimal, sequentially fair predictor, permitting a clear\ninterpretation of inter-sensitive feature correlations. Our approach seamlessly\nextends to approximate fairness, enveloping a framework accommodating the\ntrade-off between risk and unfairness. This extension permits a targeted\nprioritization of fairness improvements for a specific attribute within a set\nof sensitive attributes, allowing for a case specific adaptation. A data-driven\nestimation procedure for the derived solution is developed, and comprehensive\nnumerical experiments are conducted on both synthetic and real datasets. Our\nempirical findings decisively underscore the practical efficacy of our\npost-processing approach in fostering fair decision-making.\n", "date": "2023-09-12", "categories": []}
{"id": "2403.18216", "title": "Minimax Optimal Fair Classification with Bounded Demographic Disparity", "abstract": "  Mitigating the disparate impact of statistical machine learning methods is\ncrucial for ensuring fairness. While extensive research aims to reduce\ndisparity, the effect of using a \\emph{finite dataset} -- as opposed to the\nentire population -- remains unclear. This paper explores the statistical\nfoundations of fair binary classification with two protected groups, focusing\non controlling demographic disparity, defined as the difference in acceptance\nrates between the groups. Although fairness may come at the cost of accuracy\neven with infinite data, we show that using a finite sample incurs additional\ncosts due to the need to estimate group-specific acceptance thresholds. We\nstudy the minimax optimal classification error while constraining demographic\ndisparity to a user-specified threshold. To quantify the impact of fairness\nconstraints, we introduce a novel measure called \\emph{fairness-aware excess\nrisk} and derive a minimax lower bound on this measure that all classifiers\nmust satisfy. Furthermore, we propose FairBayes-DDP+, a group-wise thresholding\nmethod with an offset that we show attains the minimax lower bound. Our lower\nbound proofs involve several innovations. Experiments support that\nFairBayes-DDP+ controls disparity at the user-specified level, while being\nfaster and having a more favorable fairness-accuracy tradeoff than several\nbaselines.\n", "date": "2024-03-27", "categories": []}
{"id": "2403.17333", "title": "The Pursuit of Fairness in Artificial Intelligence Models: A Survey", "abstract": "  Artificial Intelligence (AI) models are now being utilized in all facets of\nour lives such as healthcare, education and employment. Since they are used in\nnumerous sensitive environments and make decisions that can be life altering,\npotential biased outcomes are a pressing matter. Developers should ensure that\nsuch models don't manifest any unexpected discriminatory practices like\npartiality for certain genders, ethnicities or disabled people. With the\nubiquitous dissemination of AI systems, researchers and practitioners are\nbecoming more aware of unfair models and are bound to mitigate bias in them.\nSignificant research has been conducted in addressing such issues to ensure\nmodels don't intentionally or unintentionally perpetuate bias. This survey\noffers a synopsis of the different ways researchers have promoted fairness in\nAI systems. We explore the different definitions of fairness existing in the\ncurrent literature. We create a comprehensive taxonomy by categorizing\ndifferent types of bias and investigate cases of biased AI in different\napplication domains. A thorough study is conducted of the approaches and\ntechniques employed by researchers to mitigate bias in AI models. Moreover, we\nalso delve into the impact of biased models on user experience and the ethical\nconsiderations to contemplate when developing and deploying such models. We\nhope this survey helps researchers and practitioners understand the intricate\ndetails of fairness and bias in AI systems. By sharing this thorough survey, we\naim to promote additional discourse in the domain of equitable and responsible\nAI.\n", "date": "2024-03-26", "categories": []}
{"id": "2306.04212", "title": "Migrate Demographic Group For Fair GNNs", "abstract": "  Graph Neural networks (GNNs) have been applied in many scenarios due to the\nsuperior performance of graph learning. However, fairness is always ignored\nwhen designing GNNs. As a consequence, biased information in training data can\neasily affect vanilla GNNs, causing biased results toward particular\ndemographic groups (divided by sensitive attributes, such as race and age).\nThere have been efforts to address the fairness issue. However, existing fair\ntechniques generally divide the demographic groups by raw sensitive attributes\nand assume that are fixed. The biased information correlated with raw sensitive\nattributes will run through the training process regardless of the implemented\nfair techniques. It is urgent to resolve this problem for training fair GNNs.\nTo tackle this problem, we propose a brand new framework, FairMigration, which\ncan dynamically migrate the demographic groups instead of keeping that fixed\nwith raw sensitive attributes. FairMigration is composed of two training\nstages. In the first stage, the GNNs are initially optimized by personalized\nself-supervised learning, and the demographic groups are adjusted dynamically.\nIn the second stage, the new demographic groups are frozen and supervised\nlearning is carried out under the constraints of new demographic groups and\nadversarial training. Extensive experiments reveal that FairMigration balances\nmodel performance and fairness well.\n", "date": "2023-06-07", "categories": []}
{"id": "2401.03695", "title": "A Large-Scale Empirical Study on Improving the Fairness of Image\n  Classification Models", "abstract": "  Fairness has been a critical issue that affects the adoption of deep learning\nmodels in real practice. To improve model fairness, many existing methods have\nbeen proposed and evaluated to be effective in their own contexts. However,\nthere is still no systematic evaluation among them for a comprehensive\ncomparison under the same context, which makes it hard to understand the\nperformance distinction among them, hindering the research progress and\npractical adoption of them. To fill this gap, this paper endeavours to conduct\nthe first large-scale empirical study to comprehensively compare the\nperformance of existing state-of-the-art fairness improving techniques.\nSpecifically, we target the widely-used application scenario of image\nclassification, and utilized three different datasets and five commonly-used\nperformance metrics to assess in total 13 methods from diverse categories. Our\nfindings reveal substantial variations in the performance of each method across\ndifferent datasets and sensitive attributes, indicating over-fitting on\nspecific datasets by many existing methods. Furthermore, different fairness\nevaluation metrics, due to their distinct focuses, yield significantly\ndifferent assessment results. Overall, we observe that pre-processing methods\nand in-processing methods outperform post-processing methods, with\npre-processing methods exhibiting the best performance. Our empirical study\noffers comprehensive recommendations for enhancing fairness in deep learning\nmodels. We approach the problem from multiple dimensions, aiming to provide a\nuniform evaluation platform and inspire researchers to explore more effective\nfairness solutions via a set of implications.\n", "date": "2024-01-08", "categories": []}
{"id": "2403.14282", "title": "How to be fair? A study of label and selection bias", "abstract": "  It is widely accepted that biased data leads to biased and thus potentially\nunfair models. Therefore, several measures for bias in data and model\npredictions have been proposed, as well as bias mitigation techniques whose aim\nis to learn models that are fair by design. Despite the myriad of mitigation\ntechniques developed in the past decade, however, it is still poorly understood\nunder what circumstances which methods work. Recently, Wick et al. showed, with\nexperiments on synthetic data, that there exist situations in which bias\nmitigation techniques lead to more accurate models when measured on unbiased\ndata. Nevertheless, in the absence of a thorough mathematical analysis, it\nremains unclear which techniques are effective under what circumstances. We\npropose to address this problem by establishing relationships between the type\nof bias and the effectiveness of a mitigation technique, where we categorize\nthe mitigation techniques by the bias measure they optimize. In this paper we\nillustrate this principle for label and selection bias on the one hand, and\ndemographic parity and ``We're All Equal'' on the other hand. Our theoretical\nanalysis allows to explain the results of Wick et al. and we also show that\nthere are situations where minimizing fairness measures does not result in the\nfairest possible distribution.\n", "date": "2024-03-21", "categories": []}
{"id": "2403.16201", "title": "From Discrete to Continuous: Deep Fair Clustering With Transferable\n  Representations", "abstract": "  We consider the problem of deep fair clustering, which partitions data into\nclusters via the representations extracted by deep neural networks while hiding\nsensitive data attributes. To achieve fairness, existing methods present a\nvariety of fairness-related objective functions based on the group fairness\ncriterion. However, these works typically assume that the sensitive attributes\nare discrete and do not work for continuous sensitive variables, such as the\nproportion of the female population in an area. Besides, the potential of the\nrepresentations learned from clustering tasks to improve performance on other\ntasks is ignored by existing works. In light of these limitations, we propose a\nflexible deep fair clustering method that can handle discrete and continuous\nsensitive attributes simultaneously. Specifically, we design an information\nbottleneck style objective function to learn fair and clustering-friendly\nrepresentations. Furthermore, we explore for the first time the transferability\nof the extracted representations to other downstream tasks. Unlike existing\nworks, we impose fairness at the representation level, which could guarantee\nfairness for the transferred task regardless of clustering results. To verify\nthe effectiveness of the proposed method, we perform extensive experiments on\ndatasets with discrete and continuous sensitive attributes, demonstrating the\nadvantage of our method in comparison with state-of-the-art methods.\n", "date": "2024-03-24", "categories": []}
{"id": "2403.14636", "title": "AI Fairness in Practice", "abstract": "  Reaching consensus on a commonly accepted definition of AI Fairness has long\nbeen a central challenge in AI ethics and governance. There is a broad spectrum\nof views across society on what the concept of fairness means and how it should\nbest be put to practice. In this workbook, we tackle this challenge by\nexploring how a context-based and society-centred approach to understanding AI\nFairness can help project teams better identify, mitigate, and manage the many\nways that unfair bias and discrimination can crop up across the AI project\nworkflow.\n  We begin by exploring how, despite the plurality of understandings about the\nmeaning of fairness, priorities of equality and non-discrimination have come to\nconstitute the broadly accepted core of its application as a practical\nprinciple. We focus on how these priorities manifest in the form of equal\nprotection from direct and indirect discrimination and from discriminatory\nharassment. These elements form ethical and legal criteria based upon which\ninstances of unfair bias and discrimination can be identified and mitigated\nacross the AI project workflow.\n  We then take a deeper dive into how the different contexts of the AI project\nlifecycle give rise to different fairness concerns. This allows us to identify\nseveral types of AI Fairness (Data Fairness, Application Fairness, Model Design\nand Development Fairness, Metric-Based Fairness, System Implementation\nFairness, and Ecosystem Fairness) that form the basis of a multi-lens approach\nto bias identification, mitigation, and management. Building on this, we\ndiscuss how to put the principle of AI Fairness into practice across the AI\nproject workflow through Bias Self-Assessment and Bias Risk Management as well\nas through the documentation of metric-based fairness criteria in a Fairness\nPosition Statement.\n", "date": "2024-02-19", "categories": []}
{"id": "2403.14727", "title": "Protected group bias and stereotypes in Large Language Models", "abstract": "  As modern Large Language Models (LLMs) shatter many state-of-the-art\nbenchmarks in a variety of domains, this paper investigates their behavior in\nthe domains of ethics and fairness, focusing on protected group bias. We\nconduct a two-part study: first, we solicit sentence continuations describing\nthe occupations of individuals from different protected groups, including\ngender, sexuality, religion, and race. Second, we have the model generate\nstories about individuals who hold different types of occupations. We collect\n>10k sentence completions made by a publicly available LLM, which we subject to\nhuman annotation. We find bias across minoritized groups, but in particular in\nthe domains of gender and sexuality, as well as Western bias, in model\ngenerations. The model not only reflects societal biases, but appears to\namplify them. The model is additionally overly cautious in replies to queries\nrelating to minoritized groups, providing responses that strongly emphasize\ndiversity and equity to an extent that other group characteristics are\novershadowed. This suggests that artificially constraining potentially harmful\noutputs may itself lead to harm, and should be applied in a careful and\ncontrolled manner.\n", "date": "2024-03-21", "categories": []}
{"id": "2302.01385", "title": "Hyper-parameter Tuning for Fair Classification without Sensitive\n  Attribute Access", "abstract": "  Fair machine learning methods seek to train models that balance model\nperformance across demographic subgroups defined over sensitive attributes like\nrace and gender. Although sensitive attributes are typically assumed to be\nknown during training, they may not be available in practice due to privacy and\nother logistical concerns. Recent work has sought to train fair models without\nsensitive attributes on training data. However, these methods need extensive\nhyper-parameter tuning to achieve good results, and hence assume that sensitive\nattributes are known on validation data. However, this assumption too might not\nbe practical. Here, we propose Antigone, a framework to train fair classifiers\nwithout access to sensitive attributes on either training or validation data.\nInstead, we generate pseudo sensitive attributes on the validation data by\ntraining a biased classifier and using the classifier's incorrectly (correctly)\nlabeled examples as proxies for minority (majority) groups. Since fairness\nmetrics like demographic parity, equal opportunity and subgroup accuracy can be\nestimated to within a proportionality constant even with noisy sensitive\nattribute information, we show theoretically and empirically that these proxy\nlabels can be used to maximize fairness under average accuracy constraints. Key\nto our results is a principled approach to select the hyper-parameters of the\nbiased classifier in a completely unsupervised fashion (meaning without access\nto ground truth sensitive attributes) that minimizes the gap between fairness\nestimated using noisy versus ground-truth sensitive labels.\n", "date": "2023-02-02", "categories": []}
{"id": "2403.13864", "title": "Optimal Transport for Fairness: Archival Data Repair using Small\n  Research Data Sets", "abstract": "  With the advent of the AI Act and other regulations, there is now an urgent\nneed for algorithms that repair unfairness in training data. In this paper, we\ndefine fairness in terms of conditional independence between protected\nattributes ($S$) and features ($X$), given unprotected attributes ($U$). We\naddress the important setting in which torrents of archival data need to be\nrepaired, using only a small proportion of these data, which are $S|U$-labelled\n(the research data). We use the latter to design optimal transport (OT)-based\nrepair plans on interpolated supports. This allows {\\em off-sample}, labelled,\narchival data to be repaired, subject to stationarity assumptions. It also\nsignificantly reduces the size of the supports of the OT plans, with\ncorrespondingly large savings in the cost of their design and of their {\\em\nsequential\\/} application to the off-sample data. We provide detailed\nexperimental results with simulated and benchmark real data (the Adult data\nset). Our performance figures demonstrate effective repair -- in the sense of\nquenching conditional dependence -- of large quantities of off-sample, labelled\n(archival) data.\n", "date": "2024-03-20", "categories": []}
{"id": "2301.12951", "title": "Unraveling Privacy Risks of Individual Fairness in Graph Neural Networks", "abstract": "  Graph neural networks (GNNs) have gained significant attraction due to their\nexpansive real-world applications. To build trustworthy GNNs, two aspects -\nfairness and privacy - have emerged as critical considerations. Previous\nstudies have separately examined the fairness and privacy aspects of GNNs,\nrevealing their trade-off with GNN performance. Yet, the interplay between\nthese two aspects remains unexplored. In this paper, we pioneer the exploration\nof the interaction between the privacy risks of edge leakage and the individual\nfairness of a GNN. Our theoretical analysis unravels that edge privacy risks\nunfortunately escalate when the nodes' individual fairness improves. Such an\nissue hinders the accomplishment of privacy and fairness of GNNs at the same\ntime. To balance fairness and privacy, we carefully introduce fairness-aware\nloss reweighting based on influence function and privacy-aware graph structure\nperturbation modules within a fine-tuning mechanism. Experimental results\nunderscore the effectiveness of our approach in achieving GNN fairness with\nlimited performance compromise and controlled privacy risks. This work\ncontributes to the comprehensively developing trustworthy GNNs by\nsimultaneously addressing both fairness and privacy aspects.\n", "date": "2023-01-30", "categories": []}
{"id": "2310.18430", "title": "MCRAGE: Synthetic Healthcare Data for Fairness", "abstract": "  In the field of healthcare, electronic health records (EHR) serve as crucial\ntraining data for developing machine learning models for diagnosis, treatment,\nand the management of healthcare resources. However, medical datasets are often\nimbalanced in terms of sensitive attributes such as race/ethnicity, gender, and\nage. Machine learning models trained on class-imbalanced EHR datasets perform\nsignificantly worse in deployment for individuals of the minority classes\ncompared to those from majority classes, which may lead to inequitable\nhealthcare outcomes for minority groups. To address this challenge, we propose\nMinority Class Rebalancing through Augmentation by Generative modeling\n(MCRAGE), a novel approach to augment imbalanced datasets using samples\ngenerated by a deep generative model. The MCRAGE process involves training a\nConditional Denoising Diffusion Probabilistic Model (CDDPM) capable of\ngenerating high-quality synthetic EHR samples from underrepresented classes. We\nuse this synthetic data to augment the existing imbalanced dataset, resulting\nin a more balanced distribution across all classes, which can be used to train\nless biased downstream models. We measure the performance of MCRAGE versus\nalternative approaches using Accuracy, F1 score and AUROC of these downstream\nmodels. We provide theoretical justification for our method in terms of recent\nconvergence results for DDPMs.\n", "date": "2023-10-27", "categories": []}
{"id": "2209.11812", "title": "Explanations, Fairness, and Appropriate Reliance in Human-AI\n  Decision-Making", "abstract": "  In this work, we study the effects of feature-based explanations on\ndistributive fairness of AI-assisted decisions, specifically focusing on the\ntask of predicting occupations from short textual bios. We also investigate how\nany effects are mediated by humans' fairness perceptions and their reliance on\nAI recommendations. Our findings show that explanations influence fairness\nperceptions, which, in turn, relate to humans' tendency to adhere to AI\nrecommendations. However, we see that such explanations do not enable humans to\ndiscern correct and incorrect AI recommendations. Instead, we show that they\nmay affect reliance irrespective of the correctness of AI recommendations.\nDepending on which features an explanation highlights, this can foster or\nhinder distributive fairness: when explanations highlight features that are\ntask-irrelevant and evidently associated with the sensitive attribute, this\nprompts overrides that counter AI recommendations that align with gender\nstereotypes. Meanwhile, if explanations appear task-relevant, this induces\nreliance behavior that reinforces stereotype-aligned errors. These results\nimply that feature-based explanations are not a reliable mechanism to improve\ndistributive fairness.\n", "date": "2022-09-23", "categories": []}
{"id": "2403.12069", "title": "Fairness Evaluation for Uplift Modeling in the Absence of Ground Truth", "abstract": "  The acceleration in the adoption of AI-based automated decision-making\nsystems poses a challenge for evaluating the fairness of algorithmic decisions,\nespecially in the absence of ground truth. When designing interventions, uplift\nmodeling is used extensively to identify candidates that are likely to benefit\nfrom treatment. However, these models remain particularly susceptible to\nfairness evaluation due to the lack of ground truth on the outcome measure\nsince a candidate cannot be in both treatment and control simultaneously. In\nthis article, we propose a framework that overcomes the missing ground truth\nproblem by generating surrogates to serve as a proxy for counterfactual labels\nof uplift modeling campaigns. We then leverage the surrogate ground truth to\nconduct a more comprehensive binary fairness evaluation. We show how to apply\nthe approach in a comprehensive study from a real-world marketing campaign for\npromotional offers and demonstrate its enhancement for fairness evaluation.\n", "date": "2024-02-12", "categories": []}
{"id": "2310.05598", "title": "On Prediction-Modelers and Decision-Makers: Why Fairness Requires More\n  Than a Fair Prediction Model", "abstract": "  An implicit ambiguity in the field of prediction-based decision-making\nregards the relation between the concepts of prediction and decision. Much of\nthe literature in the field tends to blur the boundaries between the two\nconcepts and often simply speaks of 'fair prediction.' In this paper, we point\nout that a differentiation of these concepts is helpful when implementing\nalgorithmic fairness. Even if fairness properties are related to the features\nof the used prediction model, what is more properly called 'fair' or 'unfair'\nis a decision system, not a prediction model. This is because fairness is about\nthe consequences on human lives, created by a decision, not by a prediction. We\nclarify the distinction between the concepts of prediction and decision and\nshow the different ways in which these two elements influence the final\nfairness properties of a prediction-based decision system. In addition to\nexploring this relationship conceptually and practically, we propose a\nframework that enables a better understanding and reasoning of the conceptual\nlogic of creating fairness in prediction-based decision-making. In our\nframework, we specify different roles, namely the 'prediction-modeler' and the\n'decision-maker,' and the information required from each of them for being able\nto implement fairness of the system. Our framework allows for deriving distinct\nresponsibilities for both roles and discussing some insights related to ethical\nand legal requirements. Our contribution is twofold. First, we shift the focus\nfrom abstract algorithmic fairness to context-dependent decision-making,\nrecognizing diverse actors with unique objectives and independent actions.\nSecond, we provide a conceptual framework that can help structure\nprediction-based decision problems with respect to fairness issues, identify\nresponsibilities, and implement fairness governance mechanisms in real-world\nscenarios.\n", "date": "2023-10-09", "categories": []}
{"id": "2403.10652", "title": "Improving Fairness in Credit Lending Models using Subgroup Threshold\n  Optimization", "abstract": "  In an effort to improve the accuracy of credit lending decisions, many\nfinancial intuitions are now using predictions from machine learning models.\nWhile such predictions enjoy many advantages, recent research has shown that\nthe predictions have the potential to be biased and unfair towards certain\nsubgroups of the population. To combat this, several techniques have been\nintroduced to help remove the bias and improve the overall fairness of the\npredictions. We introduce a new fairness technique, called \\textit{Subgroup\nThreshold Optimizer} (\\textit{STO}), that does not require any alternations to\nthe input training data nor does it require any changes to the underlying\nmachine learning algorithm, and thus can be used with any existing machine\nlearning pipeline. STO works by optimizing the classification thresholds for\nindividual subgroups in order to minimize the overall discrimination score\nbetween them. Our experiments on a real-world credit lending dataset show that\nSTO can reduce gender discrimination by over 90\\%.\n", "date": "2024-03-15", "categories": []}
{"id": "2403.10776", "title": "From Melting Pots to Misrepresentations: Exploring Harms in Generative\n  AI", "abstract": "  With the widespread adoption of advanced generative models such as Gemini and\nGPT, there has been a notable increase in the incorporation of such models into\nsociotechnical systems, categorized under AI-as-a-Service (AIaaS). Despite\ntheir versatility across diverse sectors, concerns persist regarding\ndiscriminatory tendencies within these models, particularly favoring selected\n`majority' demographics across various sociodemographic dimensions. Despite\nwidespread calls for diversification of media representations, marginalized\nracial and ethnic groups continue to face persistent distortion, stereotyping,\nand neglect within the AIaaS context. In this work, we provide a critical\nsummary of the state of research in the context of social harms to lead the\nconversation to focus on their implications. We also present open-ended\nresearch questions, guided by our discussion, to help define future research\npathways.\n", "date": "2024-03-16", "categories": []}
{"id": "2403.11048", "title": "JustQ: Automated Deployment of Fair and Accurate Quantum Neural Networks", "abstract": "  Despite the success of Quantum Neural Networks (QNNs) in decision-making\nsystems, their fairness remains unexplored, as the focus primarily lies on\naccuracy. This work conducts a design space exploration, unveiling QNN\nunfairness, and highlighting the significant influence of QNN deployment and\nquantum noise on accuracy and fairness. To effectively navigate the vast QNN\ndeployment design space, we propose JustQ, a framework for deploying fair and\naccurate QNNs on NISQ computers. It includes a complete NISQ error model,\nreinforcement learning-based deployment, and a flexible optimization objective\nincorporating both fairness and accuracy. Experimental results show JustQ\noutperforms previous methods, achieving superior accuracy and fairness. This\nwork pioneers fair QNN design on NISQ computers, paving the way for future\ninvestigations.\n", "date": "2024-03-17", "categories": []}
{"id": "2306.07261", "title": "Unprocessing Seven Years of Algorithmic Fairness", "abstract": "  Seven years ago, researchers proposed a postprocessing method to equalize the\nerror rates of a model across different demographic groups. The work launched\nhundreds of papers purporting to improve over the postprocessing baseline. We\nempirically evaluate these claims through thousands of model evaluations on\nseveral tabular datasets. We find that the fairness-accuracy Pareto frontier\nachieved by postprocessing contains all other methods we were feasibly able to\nevaluate. In doing so, we address two common methodological errors that have\nconfounded previous observations. One relates to the comparison of methods with\ndifferent unconstrained base models. The other concerns methods achieving\ndifferent levels of constraint relaxation. At the heart of our study is a\nsimple idea we call unprocessing that roughly corresponds to the inverse of\npostprocessing. Unprocessing allows for a direct comparison of methods using\ndifferent underlying models and levels of relaxation.\n", "date": "2023-06-12", "categories": []}
{"id": "2310.05725", "title": "Post-hoc Bias Scoring Is Optimal For Fair Classification", "abstract": "  We consider a binary classification problem under group fairness constraints,\nwhich can be one of Demographic Parity (DP), Equalized Opportunity (EOp), or\nEqualized Odds (EO). We propose an explicit characterization of Bayes optimal\nclassifier under the fairness constraints, which turns out to be a simple\nmodification rule of the unconstrained classifier. Namely, we introduce a novel\ninstance-level measure of bias, which we call bias score, and the modification\nrule is a simple linear rule on top of the finite amount of bias scores.Based\non this characterization, we develop a post-hoc approach that allows us to\nadapt to fairness constraints while maintaining high accuracy. In the case of\nDP and EOp constraints, the modification rule is thresholding a single bias\nscore, while in the case of EO constraints we are required to fit a linear\nmodification rule with 2 parameters. The method can also be applied for\ncomposite group-fairness criteria, such as ones involving several sensitive\nattributes.\n", "date": "2023-10-09", "categories": []}
{"id": "2311.07604", "title": "Finetuning Text-to-Image Diffusion Models for Fairness", "abstract": "  The rapid adoption of text-to-image diffusion models in society underscores\nan urgent need to address their biases. Without interventions, these biases\ncould propagate a skewed worldview and restrict opportunities for minority\ngroups. In this work, we frame fairness as a distributional alignment problem.\nOur solution consists of two main technical contributions: (1) a distributional\nalignment loss that steers specific characteristics of the generated images\ntowards a user-defined target distribution, and (2) adjusted direct finetuning\nof diffusion model's sampling process (adjusted DFT), which leverages an\nadjusted gradient to directly optimize losses defined on the generated images.\nEmpirically, our method markedly reduces gender, racial, and their\nintersectional biases for occupational prompts. Gender bias is significantly\nreduced even when finetuning just five soft tokens. Crucially, our method\nsupports diverse perspectives of fairness beyond absolute equality, which is\ndemonstrated by controlling age to a $75\\%$ young and $25\\%$ old distribution\nwhile simultaneously debiasing gender and race. Finally, our method is\nscalable: it can debias multiple concepts at once by simply including these\nprompts in the finetuning data. We share code and various fair diffusion model\nadaptors at https://sail-sg.github.io/finetune-fair-diffusion/.\n", "date": "2023-11-11", "categories": []}
{"id": "2403.09148", "title": "Evaluating LLMs for Gender Disparities in Notable Persons", "abstract": "  This study examines the use of Large Language Models (LLMs) for retrieving\nfactual information, addressing concerns over their propensity to produce\nfactually incorrect \"hallucinated\" responses or to altogether decline to even\nanswer prompt at all. Specifically, it investigates the presence of\ngender-based biases in LLMs' responses to factual inquiries. This paper takes a\nmulti-pronged approach to evaluating GPT models by evaluating fairness across\nmultiple dimensions of recall, hallucinations and declinations. Our findings\nreveal discernible gender disparities in the responses generated by GPT-3.5.\nWhile advancements in GPT-4 have led to improvements in performance, they have\nnot fully eradicated these gender disparities, notably in instances where\nresponses are declined. The study further explores the origins of these\ndisparities by examining the influence of gender associations in prompts and\nthe homogeneity in the responses.\n", "date": "2024-03-14", "categories": []}
{"id": "2402.18133", "title": "Classes Are Not Equal: An Empirical Study on Image Recognition Fairness", "abstract": "  In this paper, we present an empirical study on image recognition fairness,\ni.e., extreme class accuracy disparity on balanced data like ImageNet. We\nexperimentally demonstrate that classes are not equal and the fairness issue is\nprevalent for image classification models across various datasets, network\narchitectures, and model capacities. Moreover, several intriguing properties of\nfairness are identified. First, the unfairness lies in problematic\nrepresentation rather than classifier bias. Second, with the proposed concept\nof Model Prediction Bias, we investigate the origins of problematic\nrepresentation during optimization. Our findings reveal that models tend to\nexhibit greater prediction biases for classes that are more challenging to\nrecognize. It means that more other classes will be confused with harder\nclasses. Then the False Positives (FPs) will dominate the learning in\noptimization, thus leading to their poor accuracy. Further, we conclude that\ndata augmentation and representation learning algorithms improve overall\nperformance by promoting fairness to some degree in image classification. The\nCode is available at\nhttps://github.com/dvlab-research/Parametric-Contrastive-Learning.\n", "date": "2024-02-28", "categories": []}
{"id": "2403.07724", "title": "Balancing Fairness and Accuracy in Data-Restricted Binary Classification", "abstract": "  Applications that deal with sensitive information may have restrictions\nplaced on the data available to a machine learning (ML) classifier. For\nexample, in some applications, a classifier may not have direct access to\nsensitive attributes, affecting its ability to produce accurate and fair\ndecisions. This paper proposes a framework that models the trade-off between\naccuracy and fairness under four practical scenarios that dictate the type of\ndata available for analysis. Prior works examine this trade-off by analyzing\nthe outputs of a scoring function that has been trained to implicitly learn the\nunderlying distribution of the feature vector, class label, and sensitive\nattribute of a dataset. In contrast, our framework directly analyzes the\nbehavior of the optimal Bayesian classifier on this underlying distribution by\nconstructing a discrete approximation it from the dataset itself. This approach\nenables us to formulate multiple convex optimization problems, which allow us\nto answer the question: How is the accuracy of a Bayesian classifier affected\nin different data restricting scenarios when constrained to be fair? Analysis\nis performed on a set of fairness definitions that include group and individual\nfairness. Experiments on three datasets demonstrate the utility of the proposed\nframework as a tool for quantifying the trade-offs among different fairness\nnotions and their distributional dependencies.\n", "date": "2024-03-12", "categories": []}
{"id": "2403.07780", "title": "FairRR: Pre-Processing for Group Fairness through Randomized Response", "abstract": "  The increasing usage of machine learning models in consequential\ndecision-making processes has spurred research into the fairness of these\nsystems. While significant work has been done to study group fairness in the\nin-processing and post-processing setting, there has been little that\ntheoretically connects these results to the pre-processing domain. This paper\nproposes that achieving group fairness in downstream models can be formulated\nas finding the optimal design matrix in which to modify a response variable in\na Randomized Response framework. We show that measures of group fairness can be\ndirectly controlled for with optimal model utility, proposing a pre-processing\nalgorithm called FairRR that yields excellent downstream model utility and\nfairness.\n", "date": "2024-03-12", "categories": []}
{"id": "2403.07857", "title": "Fairness Feedback Loops: Training on Synthetic Data Amplifies Bias", "abstract": "  Model-induced distribution shifts (MIDS) occur as previous model outputs\npollute new model training sets over generations of models. This is known as\nmodel collapse in the case of generative models, and performative prediction or\nunfairness feedback loops for supervised models. When a model induces a\ndistribution shift, it also encodes its mistakes, biases, and unfairnesses into\nthe ground truth of its data ecosystem. We introduce a framework that allows us\nto track multiple MIDS over many generations, finding that they can lead to\nloss in performance, fairness, and minoritized group representation, even in\ninitially unbiased datasets. Despite these negative consequences, we identify\nhow models might be used for positive, intentional, interventions in their data\necosystems, providing redress for historical discrimination through a framework\ncalled algorithmic reparation (AR). We simulate AR interventions by curating\nrepresentative training batches for stochastic gradient descent to demonstrate\nhow AR can improve upon the unfairnesses of models and data ecosystems subject\nto other MIDS. Our work takes an important step towards identifying,\nmitigating, and taking accountability for the unfair feedback loops enabled by\nthe idea that ML systems are inherently neutral and objective.\n", "date": "2024-03-12", "categories": []}
{"id": "2206.11673", "title": "Is your model predicting the past?", "abstract": "  When does a machine learning model predict the future of individuals and when\ndoes it recite patterns that predate the individuals? In this work, we propose\na distinction between these two pathways of prediction, supported by\ntheoretical, empirical, and normative arguments. At the center of our proposal\nis a family of simple and efficient statistical tests, called backward\nbaselines, that demonstrate if, and to what extent, a model recounts the past.\nOur statistical theory provides guidance for interpreting backward baselines,\nestablishing equivalences between different baselines and familiar statistical\nconcepts. Concretely, we derive a meaningful backward baseline for auditing a\nprediction system as a black box, given only background variables and the\nsystem's predictions. Empirically, we evaluate the framework on different\nprediction tasks derived from longitudinal panel surveys, demonstrating the\nease and effectiveness of incorporating backward baselines into the practice of\nmachine learning.\n", "date": "2022-06-23", "categories": []}
{"id": "2208.05845", "title": "Analyzing Fairness in Deepfake Detection With Massively Annotated\n  Databases", "abstract": "  In recent years, image and video manipulations with Deepfake have become a\nsevere concern for security and society. Many detection models and datasets\nhave been proposed to detect Deepfake data reliably. However, there is an\nincreased concern that these models and training databases might be biased and,\nthus, cause Deepfake detectors to fail. In this work, we investigate factors\ncausing biased detection in public Deepfake datasets by (a) creating\nlarge-scale demographic and non-demographic attribute annotations with 47\ndifferent attributes for five popular Deepfake datasets and (b) comprehensively\nanalysing attributes resulting in AI-bias of three state-of-the-art Deepfake\ndetection backbone models on these datasets. The analysis shows how various\nattributes influence a large variety of distinctive attributes (from over 65M\nlabels) on the detection performance which includes demographic (age, gender,\nethnicity) and non-demographic (hair, skin, accessories, etc.) attributes. The\nresults examined datasets show limited diversity and, more importantly, show\nthat the utilised Deepfake detection backbone models are strongly affected by\ninvestigated attributes making them not fair across attributes. The Deepfake\ndetection backbone methods trained on such imbalanced/biased datasets result in\nincorrect detection results leading to generalisability, fairness, and security\nissues. Our findings and annotated datasets will guide future research to\nevaluate and mitigate bias in Deepfake detection techniques. The annotated\ndatasets and the corresponding code are publicly available.\n", "date": "2022-08-11", "categories": []}
{"id": "2301.09003", "title": "Blacks is to Anger as Whites is to Joy? Understanding Latent Affective\n  Bias in Large Pre-trained Neural Language Models", "abstract": "  Groundbreaking inventions and highly significant performance improvements in\ndeep learning based Natural Language Processing are witnessed through the\ndevelopment of transformer based large Pre-trained Language Models (PLMs). The\nwide availability of unlabeled data within human generated data deluge along\nwith self-supervised learning strategy helps to accelerate the success of large\nPLMs in language generation, language understanding, etc. But at the same time,\nlatent historical bias/unfairness in human minds towards a particular gender,\nrace, etc., encoded unintentionally/intentionally into the corpora harms and\nquestions the utility and efficacy of large PLMs in many real-world\napplications, particularly for the protected groups. In this paper, we present\nan extensive investigation towards understanding the existence of \"Affective\nBias\" in large PLMs to unveil any biased association of emotions such as anger,\nfear, joy, etc., towards a particular gender, race or religion with respect to\nthe downstream task of textual emotion detection. We conduct our exploration of\naffective bias from the very initial stage of corpus level affective bias\nanalysis by searching for imbalanced distribution of affective words within a\ndomain, in large scale corpora that are used to pre-train and fine-tune PLMs.\nLater, to quantify affective bias in model predictions, we perform an extensive\nset of class-based and intensity-based evaluations using various bias\nevaluation corpora. Our results show the existence of statistically significant\naffective bias in the PLM based emotion detection systems, indicating biased\nassociation of certain emotions towards a particular gender, race, and\nreligion.\n", "date": "2023-01-21", "categories": []}
{"id": "2301.11562", "title": "Arbitrariness and Social Prediction: The Confounding Role of Variance in\n  Fair Classification", "abstract": "  Variance in predictions across different trained models is a significant,\nunder-explored source of error in fair binary classification. In practice, the\nvariance on some data examples is so large that decisions can be effectively\narbitrary. To investigate this problem, we take an experimental approach and\nmake four overarching contributions: We: 1) Define a metric called\nself-consistency, derived from variance, which we use as a proxy for measuring\nand reducing arbitrariness; 2) Develop an ensembling algorithm that abstains\nfrom classification when a prediction would be arbitrary; 3) Conduct the\nlargest to-date empirical study of the role of variance (vis-a-vis\nself-consistency and arbitrariness) in fair binary classification; and, 4)\nRelease a toolkit that makes the US Home Mortgage Disclosure Act (HMDA)\ndatasets easily usable for future research. Altogether, our experiments reveal\nshocking insights about the reliability of conclusions on benchmark datasets.\nMost fair binary classification benchmarks are close-to-fair when taking into\naccount the amount of arbitrariness present in predictions -- before we even\ntry to apply any fairness interventions. This finding calls into question the\npractical utility of common algorithmic fairness methods, and in turn suggests\nthat we should reconsider how we choose to measure fairness in binary\nclassification.\n", "date": "2023-01-27", "categories": []}
{"id": "2306.04107", "title": "BeMap: Balanced Message Passing for Fair Graph Neural Network", "abstract": "  Fairness in graph neural networks has been actively studied recently.\nHowever, existing works often do not explicitly consider the role of message\npassing in introducing or amplifying the bias. In this paper, we first\ninvestigate the problem of bias amplification in message passing. We\nempirically and theoretically demonstrate that message passing could amplify\nthe bias when the 1-hop neighbors from different demographic groups are\nunbalanced. Guided by such analyses, we propose BeMap, a fair message passing\nmethod, that leverages a balance-aware sampling strategy to balance the number\nof the 1-hop neighbors of each node among different demographic groups.\nExtensive experiments on node classification demonstrate the efficacy of BeMap\nin mitigating bias while maintaining classification accuracy. The code is\navailable at https://github.com/xiaolin-cs/BeMap.\n", "date": "2023-06-07", "categories": []}
{"id": "2306.09264", "title": "Harvard Glaucoma Fairness: A Retinal Nerve Disease Dataset for Fairness\n  Learning and Fair Identity Normalization", "abstract": "  Fairness (also known as equity interchangeably) in machine learning is\nimportant for societal well-being, but limited public datasets hinder its\nprogress. Currently, no dedicated public medical datasets with imaging data for\nfairness learning are available, though minority groups suffer from more health\nissues. To address this gap, we introduce Harvard Glaucoma Fairness\n(Harvard-GF), a retinal nerve disease dataset with both 2D and 3D imaging data\nand balanced racial groups for glaucoma detection. Glaucoma is the leading\ncause of irreversible blindness globally with Blacks having doubled glaucoma\nprevalence than other races. We also propose a fair identity normalization\n(FIN) approach to equalize the feature importance between different identity\ngroups. Our FIN approach is compared with various the-state-of-the-art fairness\nlearning methods with superior performance in the racial, gender, and ethnicity\nfairness tasks with 2D and 3D imaging data, which demonstrate the utilities of\nour dataset Harvard-GF for fairness learning. To facilitate fairness\ncomparisons between different models, we propose an equity-scaled performance\nmeasure, which can be flexibly used to compare all kinds of performance metrics\nin the context of fairness. The dataset and code are publicly accessible via\n\\url{https://ophai.hms.harvard.edu/datasets/harvard-glaucoma-fairness-3300-samples/}.\n", "date": "2023-06-15", "categories": []}
{"id": "2311.05866", "title": "Fair Supervised Learning with A Simple Random Sampler of Sensitive\n  Attributes", "abstract": "  As the data-driven decision process becomes dominating for industrial\napplications, fairness-aware machine learning arouses great attention in\nvarious areas. This work proposes fairness penalties learned by neural networks\nwith a simple random sampler of sensitive attributes for non-discriminatory\nsupervised learning. In contrast to many existing works that critically rely on\nthe discreteness of sensitive attributes and response variables, the proposed\npenalty is able to handle versatile formats of the sensitive attributes, so it\nis more extensively applicable in practice than many existing algorithms. This\npenalty enables us to build a computationally efficient group-level\nin-processing fairness-aware training framework. Empirical evidence shows that\nour framework enjoys better utility and fairness measures on popular benchmark\ndata sets than competing methods. We also theoretically characterize estimation\nerrors and loss of utility of the proposed neural-penalized risk minimization\nproblem.\n", "date": "2023-11-10", "categories": []}
{"id": "2403.05975", "title": "Measuring Bias in a Ranked List using Term-based Representations", "abstract": "  In most recent studies, gender bias in document ranking is evaluated with the\nNFaiRR metric, which measures bias in a ranked list based on an aggregation\nover the unbiasedness scores of each ranked document. This perspective in\nmeasuring the bias of a ranked list has a key limitation: individual documents\nof a ranked list might be biased while the ranked list as a whole balances the\ngroups' representations. To address this issue, we propose a novel metric\ncalled TExFAIR (term exposure-based fairness), which is based on two new\nextensions to a generic fairness evaluation framework, attention-weighted\nranking fairness (AWRF). TExFAIR assesses fairness based on the term-based\nrepresentation of groups in a ranked list: (i) an explicit definition of\nassociating documents to groups based on probabilistic term-level associations,\nand (ii) a rank-biased discounting factor (RBDF) for counting\nnon-representative documents towards the measurement of the fairness of a\nranked list. We assess TExFAIR on the task of measuring gender bias in passage\nranking, and study the relationship between TExFAIR and NFaiRR. Our experiments\nshow that there is no strong correlation between TExFAIR and NFaiRR, which\nindicates that TExFAIR measures a different dimension of fairness than NFaiRR.\nWith TExFAIR, we extend the AWRF framework to allow for the evaluation of\nfairness in settings with term-based representations of groups in documents in\na ranked list.\n", "date": "2024-03-09", "categories": []}
{"id": "2403.06812", "title": "Monotone Individual Fairness", "abstract": "  We revisit the problem of online learning with individual fairness, where an\nonline learner strives to maximize predictive accuracy while ensuring that\nsimilar individuals are treated similarly. We first extend the frameworks of\nGillen et al. (2018); Bechavod et al. (2020), which rely on feedback from human\nauditors regarding fairness violations, as we consider auditing schemes that\nare capable of aggregating feedback from any number of auditors, using a rich\nclass we term monotone aggregation functions. We then prove a characterization\nfor such auditing schemes, practically reducing the analysis of auditing for\nindividual fairness by multiple auditors to that of auditing by\n(instance-specific) single auditors. Using our generalized framework, we\npresent an oracle-efficient algorithm achieving an upper bound frontier of\n$(\\mathcal{O}(T^{1/2+2b}),\\mathcal{O}(T^{3/4-b}))$ respectively for regret,\nnumber of fairness violations, for $0\\leq b \\leq 1/4$. We then study an online\nclassification setting where label feedback is available for\npositively-predicted individuals only, and present an oracle-efficient\nalgorithm achieving an upper bound frontier of\n$(\\mathcal{O}(T^{2/3+2b}),\\mathcal{O}(T^{5/6-b}))$ for regret, number of\nfairness violations, for $0\\leq b \\leq 1/6$. In both settings, our algorithms\nimprove on the best known bounds for oracle-efficient algorithms. Furthermore,\nour algorithms offer significant improvements in computational efficiency,\ngreatly reducing the number of required calls to an (offline) optimization\noracle per round, to $\\tilde{\\mathcal{O}}(\\alpha^{-2})$ in the full information\nsetting, and $\\tilde{\\mathcal{O}}(\\alpha^{-2} + k^2T^{1/3})$ in the partial\ninformation setting, where $\\alpha$ is the sensitivity for reporting fairness\nviolations, and $k$ is the number of individuals in a round.\n", "date": "2024-03-11", "categories": []}
{"id": "2310.20673", "title": "Balancing Act: Constraining Disparate Impact in Sparse Models", "abstract": "  Model pruning is a popular approach to enable the deployment of large deep\nlearning models on edge devices with restricted computational or storage\ncapacities. Although sparse models achieve performance comparable to that of\ntheir dense counterparts at the level of the entire dataset, they exhibit high\naccuracy drops for some data sub-groups. Existing methods to mitigate this\ndisparate impact induced by pruning (i) rely on surrogate metrics that address\nthe problem indirectly and have limited interpretability; or (ii) scale poorly\nwith the number of protected sub-groups in terms of computational cost. We\npropose a constrained optimization approach that directly addresses the\ndisparate impact of pruning: our formulation bounds the accuracy change between\nthe dense and sparse models, for each sub-group. This choice of constraints\nprovides an interpretable success criterion to determine if a pruned model\nachieves acceptable disparity levels. Experimental results demonstrate that our\ntechnique scales reliably to problems involving large models and hundreds of\nprotected sub-groups.\n", "date": "2023-10-31", "categories": []}
{"id": "2401.10632", "title": "Interventional Fairness on Partially Known Causal Graphs: A Constrained\n  Optimization Approach", "abstract": "  Fair machine learning aims to prevent discrimination against individuals or\nsub-populations based on sensitive attributes such as gender and race. In\nrecent years, causal inference methods have been increasingly used in fair\nmachine learning to measure unfairness by causal effects. However, current\nmethods assume that the true causal graph is given, which is often not true in\nreal-world applications. To address this limitation, this paper proposes a\nframework for achieving causal fairness based on the notion of interventions\nwhen the true causal graph is partially known. The proposed approach involves\nmodeling fair prediction using a Partially Directed Acyclic Graph (PDAG),\nspecifically, a class of causal DAGs that can be learned from observational\ndata combined with domain knowledge. The PDAG is used to measure causal\nfairness, and a constrained optimization problem is formulated to balance\nbetween fairness and accuracy. Results on both simulated and real-world\ndatasets demonstrate the effectiveness of this method.\n", "date": "2024-01-19", "categories": []}
{"id": "2403.05114", "title": "APPLE: Adversarial Privacy-aware Perturbations on Latent Embedding for\n  Unfairness Mitigation", "abstract": "  Ensuring fairness in deep-learning-based segmentors is crucial for health\nequity. Much effort has been dedicated to mitigating unfairness in the training\ndatasets or procedures. However, with the increasing prevalence of foundation\nmodels in medical image analysis, it is hard to train fair models from scratch\nwhile preserving utility. In this paper, we propose a novel method, Adversarial\nPrivacy-aware Perturbations on Latent Embedding (APPLE), that can improve the\nfairness of deployed segmentors by introducing a small latent feature perturber\nwithout updating the weights of the original model. By adding perturbation to\nthe latent vector, APPLE decorates the latent vector of segmentors such that no\nfairness-related features can be passed to the decoder of the segmentors while\npreserving the architecture and parameters of the segmentor. Experiments on two\nsegmentation datasets and five segmentors (three U-Net-like and two SAM-like)\nillustrate the effectiveness of our proposed method compared to several\nunfairness mitigation methods.\n", "date": "2024-03-08", "categories": []}
{"id": "2403.05235", "title": "Fairness-Aware Interpretable Modeling (FAIM) for Trustworthy Machine\n  Learning in Healthcare", "abstract": "  The escalating integration of machine learning in high-stakes fields such as\nhealthcare raises substantial concerns about model fairness. We propose an\ninterpretable framework - Fairness-Aware Interpretable Modeling (FAIM), to\nimprove model fairness without compromising performance, featuring an\ninteractive interface to identify a \"fairer\" model from a set of\nhigh-performing models and promoting the integration of data-driven evidence\nand clinical expertise to enhance contextualized fairness. We demonstrated\nFAIM's value in reducing sex and race biases by predicting hospital admission\nwith two real-world databases, MIMIC-IV-ED and SGH-ED. We show that for both\ndatasets, FAIM models not only exhibited satisfactory discriminatory\nperformance but also significantly mitigated biases as measured by\nwell-established fairness metrics, outperforming commonly used bias-mitigation\nmethods. Our approach demonstrates the feasibility of improving fairness\nwithout sacrificing performance and provides an a modeling mode that invites\ndomain experts to engage, fostering a multidisciplinary effort toward tailored\nAI fairness.\n", "date": "2024-03-08", "categories": []}
{"id": "2305.18761", "title": "Identifying Spurious Biases Early in Training through the Lens of\n  Simplicity Bias", "abstract": "  Neural networks trained with (stochastic) gradient descent have an inductive\nbias towards learning simpler solutions. This makes them highly prone to\nlearning spurious correlations in the training data, that may not hold at test\ntime. In this work, we provide the first theoretical analysis of the effect of\nsimplicity bias on learning spurious correlations. Notably, we show that\nexamples with spurious features are provably separable based on the model's\noutput early in training. We further illustrate that if spurious features have\na small enough noise-to-signal ratio, the network's output on the majority of\nexamples is almost exclusively determined by the spurious features, leading to\npoor worst-group test accuracy. Finally, we propose SPARE, which identifies\nspurious correlations early in training and utilizes importance sampling to\nalleviate their effect. Empirically, we demonstrate that SPARE outperforms\nstate-of-the-art methods by up to 21.1% in worst-group accuracy, while being up\nto 12x faster. We also show that SPARE is a highly effective but lightweight\nmethod to discover spurious correlations.\n", "date": "2023-05-30", "categories": []}
{"id": "2403.04445", "title": "Classist Tools: Social Class Correlates with Performance in NLP", "abstract": "  Since the foundational work of William Labov on the social stratification of\nlanguage (Labov, 1964), linguistics has made concentrated efforts to explore\nthe links between sociodemographic characteristics and language production and\nperception. But while there is strong evidence for socio-demographic\ncharacteristics in language, they are infrequently used in Natural Language\nProcessing (NLP). Age and gender are somewhat well represented, but Labov's\noriginal target, socioeconomic status, is noticeably absent. And yet it\nmatters. We show empirically that NLP disadvantages less-privileged\nsocioeconomic groups. We annotate a corpus of 95K utterances from movies with\nsocial class, ethnicity and geographical language variety and measure the\nperformance of NLP systems on three tasks: language modelling, automatic speech\nrecognition, and grammar error correction. We find significant performance\ndisparities that can be attributed to socioeconomic status as well as ethnicity\nand geographical differences. With NLP technologies becoming ever more\nubiquitous and quotidian, they must accommodate all language varieties to avoid\ndisadvantaging already marginalised groups. We argue for the inclusion of\nsocioeconomic class in future language technologies.\n", "date": "2024-03-07", "categories": []}
{"id": "2311.17695", "title": "Fair Text-to-Image Diffusion via Fair Mapping", "abstract": "  In this paper, we address the limitations of existing text-to-image diffusion\nmodels in generating demographically fair results when given human-related\ndescriptions. These models often struggle to disentangle the target language\ncontext from sociocultural biases, resulting in biased image generation. To\novercome this challenge, we propose Fair Mapping, a flexible, model-agnostic,\nand lightweight approach that modifies a pre-trained text-to-image diffusion\nmodel by controlling the prompt to achieve fair image generation. One key\nadvantage of our approach is its high efficiency. It only requires updating an\nadditional linear network with few parameters at a low computational cost. By\ndeveloping a linear network that maps conditioning embeddings into a debiased\nspace, we enable the generation of relatively balanced demographic results\nbased on the specified text condition. With comprehensive experiments on face\nimage generation, we show that our method significantly improves image\ngeneration fairness with almost the same image quality compared to conventional\ndiffusion models when prompted with descriptions related to humans. By\neffectively addressing the issue of implicit language bias, our method produces\nmore fair and diverse image outputs.\n", "date": "2023-11-29", "categories": []}
{"id": "2211.10209", "title": "On the Alignment of Group Fairness with Attribute Privacy", "abstract": "  Group fairness and privacy are fundamental aspects in designing trustworthy\nmachine learning models. Previous research has highlighted conflicts between\ngroup fairness and different privacy notions. We are the first to demonstrate\nthe alignment of group fairness with the specific privacy notion of attribute\nprivacy in a blackbox setting. Attribute privacy, quantified by the resistance\nto attribute inference attacks (AIAs), requires indistinguishability in the\ntarget model's output predictions. Group fairness guarantees this thereby\nmitigating AIAs and achieving attribute privacy. To demonstrate this, we first\nintroduce AdaptAIA, an enhancement of existing AIAs, tailored for real-world\ndatasets with class imbalances in sensitive attributes. Through theoretical and\nextensive empirical analyses, we demonstrate the efficacy of two standard group\nfairness algorithms (i.e., adversarial debiasing and exponentiated gradient\ndescent) against AdaptAIA. Additionally, since using group fairness results in\nattribute privacy, it acts as a defense against AIAs, which is currently\nlacking. Overall, we show that group fairness aligns with attribute privacy at\nno additional cost other than the already existing trade-off with model\nutility.\n", "date": "2022-11-18", "categories": []}
{"id": "2307.11333", "title": "Demystifying Local and Global Fairness Trade-offs in Federated Learning\n  Using Partial Information Decomposition", "abstract": "  This work presents an information-theoretic perspective to group fairness\ntrade-offs in federated learning (FL) with respect to sensitive attributes,\nsuch as gender, race, etc. Existing works often focus on either $\\textit{global\nfairness}$ (overall disparity of the model across all clients) or\n$\\textit{local fairness}$ (disparity of the model at each client), without\nalways considering their trade-offs. There is a lack of understanding regarding\nthe interplay between global and local fairness in FL, particularly under data\nheterogeneity, and if and when one implies the other. To address this gap, we\nleverage a body of work in information theory called partial information\ndecomposition (PID), which first identifies three sources of unfairness in FL,\nnamely, $\\textit{Unique Disparity}$, $\\textit{Redundant Disparity}$, and\n$\\textit{Masked Disparity}$. We demonstrate how these three disparities\ncontribute to global and local fairness using canonical examples. This\ndecomposition helps us derive fundamental limits on the trade-off between\nglobal and local fairness, highlighting where they agree or disagree. We\nintroduce the $\\textit{Accuracy and Global-Local Fairness Optimality Problem\n(AGLFOP)}$, a convex optimization that defines the theoretical limits of\naccuracy and fairness trade-offs, identifying the best possible performance any\nFL strategy can attain given a dataset and client distribution. We also present\nexperimental results on synthetic datasets and the ADULT dataset to support our\ntheoretical findings.\n", "date": "2023-07-21", "categories": []}
{"id": "2403.02726", "title": "Bias in Generative AI", "abstract": "  This study analyzed images generated by three popular generative artificial\nintelligence (AI) tools - Midjourney, Stable Diffusion, and DALLE 2 -\nrepresenting various occupations to investigate potential bias in AI\ngenerators. Our analysis revealed two overarching areas of concern in these AI\ngenerators, including (1) systematic gender and racial biases, and (2) subtle\nbiases in facial expressions and appearances. Firstly, we found that all three\nAI generators exhibited bias against women and African Americans. Moreover, we\nfound that the evident gender and racial biases uncovered in our analysis were\neven more pronounced than the status quo when compared to labor force\nstatistics or Google images, intensifying the harmful biases we are actively\nstriving to rectify in our society. Secondly, our study uncovered more nuanced\nprejudices in the portrayal of emotions and appearances. For example, women\nwere depicted as younger with more smiles and happiness, while men were\ndepicted as older with more neutral expressions and anger, posing a risk that\ngenerative AI models may unintentionally depict women as more submissive and\nless competent than men. Such nuanced biases, by their less overt nature, might\nbe more problematic as they can permeate perceptions unconsciously and may be\nmore difficult to rectify. Although the extent of bias varied depending on the\nmodel, the direction of bias remained consistent in both commercial and\nopen-source AI generators. As these tools become commonplace, our study\nhighlights the urgency to identify and mitigate various biases in generative\nAI, reinforcing the commitment to ensuring that AI technologies benefit all of\nhumanity for a more inclusive future.\n", "date": "2024-03-05", "categories": []}
{"id": "2310.13822", "title": "Adversarial Attacks on Fairness of Graph Neural Networks", "abstract": "  Fairness-aware graph neural networks (GNNs) have gained a surge of attention\nas they can reduce the bias of predictions on any demographic group (e.g.,\nfemale) in graph-based applications. Although these methods greatly improve the\nalgorithmic fairness of GNNs, the fairness can be easily corrupted by carefully\ndesigned adversarial attacks. In this paper, we investigate the problem of\nadversarial attacks on fairness of GNNs and propose G-FairAttack, a general\nframework for attacking various types of fairness-aware GNNs in terms of\nfairness with an unnoticeable effect on prediction utility. In addition, we\npropose a fast computation technique to reduce the time complexity of\nG-FairAttack. The experimental study demonstrates that G-FairAttack\nsuccessfully corrupts the fairness of different types of GNNs while keeping the\nattack unnoticeable. Our study on fairness attacks sheds light on potential\nvulnerabilities in fairness-aware GNNs and guides further research on the\nrobustness of GNNs in terms of fairness.\n", "date": "2023-10-20", "categories": []}
{"id": "2401.16439", "title": "Distribution-Specific Auditing For Subgroup Fairness", "abstract": "  We study the problem of auditing classifiers with the notion of statistical\nsubgroup fairness. Kearns et al. (2018) has shown that the problem of auditing\ncombinatorial subgroups fairness is as hard as agnostic learning. Essentially\nall work on remedying statistical measures of discrimination against subgroups\nassumes access to an oracle for this problem, despite the fact that no\nefficient algorithms are known for it. If we assume the data distribution is\nGaussian, or even merely log-concave, then a recent line of work has discovered\nefficient agnostic learning algorithms for halfspaces. Unfortunately, the\nreduction of Kearns et al. was formulated in terms of weak, \"distribution-free\"\nlearning, and thus did not establish a connection for families such as\nlog-concave distributions.\n  In this work, we give positive and negative results on auditing for Gaussian\ndistributions: On the positive side, we present an alternative approach to\nleverage these advances in agnostic learning and thereby obtain the first\npolynomial-time approximation scheme (PTAS) for auditing nontrivial\ncombinatorial subgroup fairness: we show how to audit statistical notions of\nfairness over homogeneous halfspace subgroups when the features are Gaussian.\nOn the negative side, we find that under cryptographic assumptions, no\npolynomial-time algorithm can guarantee any nontrivial auditing, even under\nGaussian feature distributions, for general halfspace subgroups.\n", "date": "2024-01-27", "categories": []}
{"id": "2403.00198", "title": "AXOLOTL: Fairness through Assisted Self-Debiasing of Large Language\n  Model Outputs", "abstract": "  Pre-trained Large Language Models (LLMs) have significantly advanced natural\nlanguage processing capabilities but are susceptible to biases present in their\ntraining data, leading to unfair outcomes in various applications. While\nnumerous strategies have been proposed to mitigate bias, they often require\nextensive computational resources and may compromise model performance. In this\nwork, we introduce AXOLOTL, a novel post-processing framework, which operates\nagnostically across tasks and models, leveraging public APIs to interact with\nLLMs without direct access to internal parameters. Through a three-step process\nresembling zero-shot learning, AXOLOTL identifies biases, proposes resolutions,\nand guides the model to self-debias its outputs. This approach minimizes\ncomputational costs and preserves model performance, making AXOLOTL a promising\ntool for debiasing LLM outputs with broad applicability and ease of use.\n", "date": "2024-03-01", "categories": []}
{"id": "2403.00277", "title": "Gender Bias in Large Language Models across Multiple Languages", "abstract": "  With the growing deployment of large language models (LLMs) across various\napplications, assessing the influence of gender biases embedded in LLMs becomes\ncrucial. The topic of gender bias within the realm of natural language\nprocessing (NLP) has gained considerable focus, particularly in the context of\nEnglish. Nonetheless, the investigation of gender bias in languages other than\nEnglish is still relatively under-explored and insufficiently analyzed. In this\nwork, We examine gender bias in LLMs-generated outputs for different languages.\nWe use three measurements: 1) gender bias in selecting descriptive words given\nthe gender-related context. 2) gender bias in selecting gender-related pronouns\n(she/he) given the descriptive words. 3) gender bias in the topics of\nLLM-generated dialogues. We investigate the outputs of the GPT series of LLMs\nin various languages using our three measurement methods. Our findings revealed\nsignificant gender biases across all the languages we examined.\n", "date": "2024-03-01", "categories": []}
{"id": "2403.00625", "title": "Bias Mitigation in Fine-tuning Pre-trained Models for Enhanced Fairness\n  and Efficiency", "abstract": "  Fine-tuning pre-trained models is a widely employed technique in numerous\nreal-world applications. However, fine-tuning these models on new tasks can\nlead to unfair outcomes. This is due to the absence of generalization\nguarantees for fairness properties, regardless of whether the original\npre-trained model was developed with fairness considerations. To tackle this\nissue, we introduce an efficient and robust fine-tuning framework specifically\ndesigned to mitigate biases in new tasks. Our empirical analysis shows that the\nparameters in the pre-trained model that affect predictions for different\ndemographic groups are different, so based on this observation, we employ a\ntransfer learning strategy that neutralizes the importance of these influential\nweights, determined using Fisher information across demographic groups.\nAdditionally, we integrate this weight importance neutralization strategy with\na matrix factorization technique, which provides a low-rank approximation of\nthe weight matrix using fewer parameters, reducing the computational demands.\nExperiments on multiple pre-trained models and new tasks demonstrate the\neffectiveness of our method.\n", "date": "2024-03-01", "categories": []}
{"id": "2312.03151", "title": "Multitask Learning Can Improve Worst-Group Outcomes", "abstract": "  In order to create machine learning systems that serve a variety of users\nwell, it is vital to not only achieve high average performance but also ensure\nequitable outcomes across diverse groups. However, most machine learning\nmethods are designed to improve a model's average performance on a chosen end\ntask without consideration for their impact on worst group error. Multitask\nlearning (MTL) is one such widely used technique. In this paper, we seek not\nonly to understand the impact of MTL on worst-group accuracy but also to\nexplore its potential as a tool to address the challenge of group-wise\nfairness. We primarily consider the standard setting of fine-tuning a\npre-trained model, where, following recent work \\citep{gururangan2020don,\ndery2023aang}, we multitask the end task with the pre-training objective\nconstructed from the end task data itself. In settings with few or no group\nannotations, we find that multitasking often, but not consistently, achieves\nbetter worst-group accuracy than Just-Train-Twice (JTT;\n\\citet{pmlr-v139-liu21f}) -- a representative distributionally robust\noptimization (DRO) method. Leveraging insights from synthetic data experiments,\nwe propose to modify standard MTL by regularizing the joint multitask\nrepresentation space. We run a large number of fine-tuning experiments across\ncomputer vision and natural language processing datasets and find that our\nregularized MTL approach \\emph{consistently} outperforms JTT on both average\nand worst-group outcomes. Our official code can be found here:\n\\href{https://github.com/atharvajk98/MTL-group-robustness.git}{\\url{https://github.com/atharvajk98/MTL-group-robustness}}.\n", "date": "2023-12-05", "categories": []}
{"id": "2402.18803", "title": "To Pool or Not To Pool: Analyzing the Regularizing Effects of Group-Fair\n  Training on Shared Models", "abstract": "  In fair machine learning, one source of performance disparities between\ngroups is over-fitting to groups with relatively few training samples. We\nderive group-specific bounds on the generalization error of welfare-centric\nfair machine learning that benefit from the larger sample size of the majority\ngroup. We do this by considering group-specific Rademacher averages over a\nrestricted hypothesis class, which contains the family of models likely to\nperform well with respect to a fair learning objective (e.g., a power-mean).\nOur simulations demonstrate these bounds improve over a naive method, as\nexpected by theory, with particularly significant improvement for smaller group\nsizes.\n", "date": "2024-02-29", "categories": []}
{"id": "2402.18502", "title": "Few-Shot Fairness: Unveiling LLM's Potential for Fairness-Aware\n  Classification", "abstract": "  Employing Large Language Models (LLM) in various downstream applications such\nas classification is crucial, especially for smaller companies lacking the\nexpertise and resources required for fine-tuning a model. Fairness in LLMs\nhelps ensure inclusivity, equal representation based on factors such as race,\ngender and promotes responsible AI deployment. As the use of LLMs has become\nincreasingly prevalent, it is essential to assess whether LLMs can generate\nfair outcomes when subjected to considerations of fairness. In this study, we\nintroduce a framework outlining fairness regulations aligned with various\nfairness definitions, with each definition being modulated by varying degrees\nof abstraction. We explore the configuration for in-context learning and the\nprocedure for selecting in-context demonstrations using RAG, while\nincorporating fairness rules into the process. Experiments conducted with\ndifferent LLMs indicate that GPT-4 delivers superior results in terms of both\naccuracy and fairness compared to other models. This work is one of the early\nattempts to achieve fairness in prediction tasks by utilizing LLMs through\nin-context learning.\n", "date": "2024-02-28", "categories": []}
{"id": "2303.00968", "title": "Dynamic fairness-aware recommendation through multi-agent social choice", "abstract": "  Algorithmic fairness in the context of personalized recommendation presents\nsignificantly different challenges to those commonly encountered in\nclassification tasks. Researchers studying classification have generally\nconsidered fairness to be a matter of achieving equality of outcomes between a\nprotected and unprotected group, and built algorithmic interventions on this\nbasis. We argue that fairness in real-world application settings in general,\nand especially in the context of personalized recommendation, is much more\ncomplex and multi-faceted, requiring a more general approach. We propose a\nmodel to formalize multistakeholder fairness in recommender systems as a two\nstage social choice problem. In particular, we express recommendation fairness\nas a novel combination of an allocation and an aggregation problem, which\nintegrate both fairness concerns and personalized recommendation provisions,\nand derive new recommendation techniques based on this formulation. Simulations\ndemonstrate the ability of the framework to integrate multiple fairness\nconcerns in a dynamic way.\n", "date": "2023-03-02", "categories": []}
{"id": "2303.16564", "title": "Implicit Visual Bias Mitigation by Posterior Estimate Sharpening of a\n  Bayesian Neural Network", "abstract": "  The fairness of a deep neural network is strongly affected by dataset bias\nand spurious correlations, both of which are usually present in modern\nfeature-rich and complex visual datasets. Due to the difficulty and variability\nof the task, no single de-biasing method has been universally successful. In\nparticular, implicit methods not requiring explicit knowledge of bias variables\nare especially relevant for real-world applications. We propose a novel\nimplicit mitigation method using a Bayesian neural network, allowing us to\nleverage the relationship between epistemic uncertainties and the presence of\nbias or spurious correlations in a sample. Our proposed posterior estimate\nsharpening procedure encourages the network to focus on core features that do\nnot contribute to high uncertainties. Experimental results on three benchmark\ndatasets demonstrate that Bayesian networks with sharpened posterior estimates\nperform comparably to prior existing methods and show potential worthy of\nfurther exploration.\n", "date": "2023-03-29", "categories": []}
{"id": "2309.08621", "title": "Exploring Social Choice Mechanisms for Recommendation Fairness in SCRUF", "abstract": "  Fairness problems in recommender systems often have a complexity in practice\nthat is not adequately captured in simplified research formulations. A social\nchoice formulation of the fairness problem, operating within a multi-agent\narchitecture of fairness concerns, offers a flexible and multi-aspect\nalternative to fairness-aware recommendation approaches. Leveraging social\nchoice allows for increased generality and the possibility of tapping into\nwell-studied social choice algorithms for resolving the tension between\nmultiple, competing fairness concerns. This paper explores a range of options\nfor choice mechanisms in multi-aspect fairness applications using both real and\nsynthetic data and shows that different classes of choice and allocation\nmechanisms yield different but consistent fairness / accuracy tradeoffs. We\nalso show that a multi-agent formulation offers flexibility in adapting to user\npopulation dynamics.\n", "date": "2023-09-10", "categories": []}
{"id": "2311.05227", "title": "Kantian Deontology Meets AI Alignment: Towards Morally Grounded Fairness\n  Metrics", "abstract": "  Deontological ethics, specifically understood through Immanuel Kant, provides\na moral framework that emphasizes the importance of duties and principles,\nrather than the consequences of action. Understanding that despite the\nprominence of deontology, it is currently an overlooked approach in fairness\nmetrics, this paper explores the compatibility of a Kantian deontological\nframework in fairness metrics, part of the AI alignment field. We revisit\nKant's critique of utilitarianism, which is the primary approach in AI fairness\nmetrics and argue that fairness principles should align with the Kantian\ndeontological framework. By integrating Kantian ethics into AI alignment, we\nnot only bring in a widely-accepted prominent moral theory but also strive for\na more morally grounded AI landscape that better balances outcomes and\nprocedures in pursuit of fairness and justice.\n", "date": "2023-11-09", "categories": []}
{"id": "2402.16458", "title": "ID-XCB: Data-independent Debiasing for Fair and Accurate\n  Transformer-based Cyberbullying Detection", "abstract": "  Swear words are a common proxy to collect datasets with cyberbullying\nincidents. Our focus is on measuring and mitigating biases derived from\nspurious associations between swear words and incidents occurring as a result\nof such data collection strategies. After demonstrating and quantifying these\nbiases, we introduce ID-XCB, the first data-independent debiasing technique\nthat combines adversarial training, bias constraints and debias fine-tuning\napproach aimed at alleviating model attention to bias-inducing words without\nimpacting overall model performance. We explore ID-XCB on two popular\nsession-based cyberbullying datasets along with comprehensive ablation and\ngeneralisation studies. We show that ID-XCB learns robust cyberbullying\ndetection capabilities while mitigating biases, outperforming state-of-the-art\ndebiasing methods in both performance and bias mitigation. Our quantitative and\nqualitative analyses demonstrate its generalisability to unseen data.\n", "date": "2024-02-26", "categories": []}
{"id": "2402.16979", "title": "Algorithmic Arbitrariness in Content Moderation", "abstract": "  Machine learning (ML) is widely used to moderate online content. Despite its\nscalability relative to human moderation, the use of ML introduces unique\nchallenges to content moderation. One such challenge is predictive\nmultiplicity: multiple competing models for content classification may perform\nequally well on average, yet assign conflicting predictions to the same\ncontent. This multiplicity can result from seemingly innocuous choices during\nmodel development, such as random seed selection for parameter initialization.\nWe experimentally demonstrate how content moderation tools can arbitrarily\nclassify samples as toxic, leading to arbitrary restrictions on speech. We\ndiscuss these findings in terms of human rights set out by the International\nCovenant on Civil and Political Rights (ICCPR), namely freedom of expression,\nnon-discrimination, and procedural justice. We analyze (i) the extent of\npredictive multiplicity among state-of-the-art LLMs used for detecting toxic\ncontent; (ii) the disparate impact of this arbitrariness across social groups;\nand (iii) how model multiplicity compares to unambiguous human classifications.\nOur findings indicate that the up-scaled algorithmic moderation risks\nlegitimizing an algorithmic leviathan, where an algorithm disproportionately\nmanages human rights. To mitigate such risks, our study underscores the need to\nidentify and increase the transparency of arbitrariness in content moderation\napplications. Since algorithmic content moderation is being fueled by pressing\nsocial concerns, such as disinformation and hate speech, our discussion on\nharms raises concerns relevant to policy debates. Our findings also contribute\nto content moderation and intermediary liability laws being discussed and\npassed in many countries, such as the Digital Services Act in the European\nUnion, the Online Safety Act in the United Kingdom, and the Fake News Bill in\nBrazil.\n", "date": "2024-02-26", "categories": []}
{"id": "2402.17101", "title": "T-HITL Effectively Addresses Problematic Associations in Image\n  Generation and Maintains Overall Visual Quality", "abstract": "  Generative AI image models may inadvertently generate problematic\nrepresentations of people. Past research has noted that millions of users\nengage daily across the world with these models and that the models, including\nthrough problematic representations of people, have the potential to compound\nand accelerate real-world discrimination and other harms (Bianchi et al, 2023).\nIn this paper, we focus on addressing the generation of problematic\nassociations between demographic groups and semantic concepts that may reflect\nand reinforce negative narratives embedded in social data. Building on\nsociological literature (Blumer, 1958) and mapping representations to model\nbehaviors, we have developed a taxonomy to study problematic associations in\nimage generation models. We explore the effectiveness of fine tuning at the\nmodel level as a method to address these associations, identifying a potential\nreduction in visual quality as a limitation of traditional fine tuning. We also\npropose a new methodology with twice-human-in-the-loop (T-HITL) that promises\nimprovements in both reducing problematic associations and also maintaining\nvisual quality. We demonstrate the effectiveness of T-HITL by providing\nevidence of three problematic associations addressed by T-HITL at the model\nlevel. Our contributions to scholarship are two-fold. By defining problematic\nassociations in the context of machine learning models and generative AI, we\nintroduce a conceptual and technical taxonomy for addressing some of these\nassociations. Finally, we provide a method, T-HITL, that addresses these\nassociations and simultaneously maintains visual quality of image model\ngenerations. This mitigation need not be a tradeoff, but rather an enhancement.\n", "date": "2024-02-27", "categories": []}
{"id": "2402.15561", "title": "Fair Multivariate Adaptive Regression Splines for Ensuring Equity and\n  Transparency", "abstract": "  Predictive analytics is widely used in various domains, including education,\nto inform decision-making and improve outcomes. However, many predictive models\nare proprietary and inaccessible for evaluation or modification by researchers\nand practitioners, limiting their accountability and ethical design. Moreover,\npredictive models are often opaque and incomprehensible to the officials who\nuse them, reducing their trust and utility. Furthermore, predictive models may\nintroduce or exacerbate bias and inequity, as they have done in many sectors of\nsociety. Therefore, there is a need for transparent, interpretable, and fair\npredictive models that can be easily adopted and adapted by different\nstakeholders. In this paper, we propose a fair predictive model based on\nmultivariate adaptive regression splines(MARS) that incorporates fairness\nmeasures in the learning process. MARS is a non-parametric regression model\nthat performs feature selection, handles non-linear relationships, generates\ninterpretable decision rules, and derives optimal splitting criteria on the\nvariables. Specifically, we integrate fairness into the knot optimization\nalgorithm and provide theoretical and empirical evidence of how it results in a\nfair knot placement. We apply our fairMARS model to real-world data and\ndemonstrate its effectiveness in terms of accuracy and equity. Our paper\ncontributes to the advancement of responsible and ethical predictive analytics\nfor social good.\n", "date": "2024-02-23", "categories": []}
{"id": "2402.15477", "title": "Debiasing Machine Learning Models by Using Weakly Supervised Learning", "abstract": "  We tackle the problem of bias mitigation of algorithmic decisions in a\nsetting where both the output of the algorithm and the sensitive variable are\ncontinuous. Most of prior work deals with discrete sensitive variables, meaning\nthat the biases are measured for subgroups of persons defined by a label,\nleaving out important algorithmic bias cases, where the sensitive variable is\ncontinuous. Typical examples are unfair decisions made with respect to the age\nor the financial status. In our work, we then propose a bias mitigation\nstrategy for continuous sensitive variables, based on the notion of endogeneity\nwhich comes from the field of econometrics. In addition to solve this new\nproblem, our bias mitigation strategy is a weakly supervised learning method\nwhich requires that a small portion of the data can be measured in a fair\nmanner. It is model agnostic, in the sense that it does not make any hypothesis\non the prediction model. It also makes use of a reasonably large amount of\ninput observations and their corresponding predictions. Only a small fraction\nof the true output predictions should be known. This therefore limits the need\nfor expert interventions. Results obtained on synthetic data show the\neffectiveness of our approach for examples as close as possible to real-life\napplications in econometrics.\n", "date": "2024-02-23", "categories": []}
{"id": "2402.14305", "title": "Towards Efficient Pareto-optimal Utility-Fairness between Groups in\n  Repeated Rankings", "abstract": "  In this paper, we tackle the problem of computing a sequence of rankings with\nthe guarantee of the Pareto-optimal balance between (1) maximizing the utility\nof the consumers and (2) minimizing unfairness between producers of the items.\nSuch a multi-objective optimization problem is typically solved using a\ncombination of a scalarization method and linear programming on bi-stochastic\nmatrices, representing the distribution of possible rankings of items. However,\nthe above-mentioned approach relies on Birkhoff-von Neumann (BvN)\ndecomposition, of which the computational complexity is $\\mathcal{O}(n^5)$ with\n$n$ being the number of items, making it impractical for large-scale systems.\nTo address this drawback, we introduce a novel approach to the above problem by\nusing the Expohedron - a permutahedron whose points represent all achievable\nexposures of items. On the Expohedron, we profile the Pareto curve which\ncaptures the trade-off between group fairness and user utility by identifying a\nfinite number of Pareto optimal solutions. We further propose an efficient\nmethod by relaxing our optimization problem on the Expohedron's circumscribed\n$n$-sphere, which significantly improve the running time. Moreover, the\napproximate Pareto curve is asymptotically close to the real Pareto optimal\ncurve as the number of substantial solutions increases. Our methods are\napplicable with different ranking merits that are non-decreasing functions of\nitem relevance. The effectiveness of our methods are validated through\nexperiments on both synthetic and real-world datasets.\n", "date": "2024-02-22", "categories": []}
{"id": "2402.14815", "title": "Demographic Bias of Expert-Level Vision-Language Foundation Models in\n  Medical Imaging", "abstract": "  Advances in artificial intelligence (AI) have achieved expert-level\nperformance in medical imaging applications. Notably, self-supervised\nvision-language foundation models can detect a broad spectrum of pathologies\nwithout relying on explicit training annotations. However, it is crucial to\nensure that these AI models do not mirror or amplify human biases, thereby\ndisadvantaging historically marginalized groups such as females or Black\npatients. The manifestation of such biases could systematically delay essential\nmedical care for certain patient subgroups. In this study, we investigate the\nalgorithmic fairness of state-of-the-art vision-language foundation models in\nchest X-ray diagnosis across five globally-sourced datasets. Our findings\nreveal that compared to board-certified radiologists, these foundation models\nconsistently underdiagnose marginalized groups, with even higher rates seen in\nintersectional subgroups, such as Black female patients. Such demographic\nbiases present over a wide range of pathologies and demographic attributes.\nFurther analysis of the model embedding uncovers its significant encoding of\ndemographic information. Deploying AI systems with these biases in medical\nimaging can intensify pre-existing care disparities, posing potential\nchallenges to equitable healthcare access and raising ethical questions about\ntheir clinical application.\n", "date": "2024-02-22", "categories": []}
{"id": "2308.10149", "title": "A Survey on Fairness in Large Language Models", "abstract": "  Large Language Models (LLMs) have shown powerful performance and development\nprospects and are widely deployed in the real world. However, LLMs can capture\nsocial biases from unprocessed training data and propagate the biases to\ndownstream tasks. Unfair LLM systems have undesirable social impacts and\npotential harms. In this paper, we provide a comprehensive review of related\nresearch on fairness in LLMs. Considering the influence of parameter magnitude\nand training paradigm on research strategy, we divide existing fairness\nresearch into oriented to medium-sized LLMs under pre-training and fine-tuning\nparadigms and oriented to large-sized LLMs under prompting paradigms. First,\nfor medium-sized LLMs, we introduce evaluation metrics and debiasing methods\nfrom the perspectives of intrinsic bias and extrinsic bias, respectively. Then,\nfor large-sized LLMs, we introduce recent fairness research, including fairness\nevaluation, reasons for bias, and debiasing methods. Finally, we discuss and\nprovide insight on the challenges and future directions for the development of\nfairness in LLMs.\n", "date": "2023-08-20", "categories": []}
{"id": "2312.04404", "title": "On the Impact of Multi-dimensional Local Differential Privacy on\n  Fairness", "abstract": "  Automated decision systems are increasingly used to make consequential\ndecisions in people's lives. Due to the sensitivity of the manipulated data as\nwell as the resulting decisions, several ethical concerns need to be addressed\nfor the appropriate use of such technologies, in particular, fairness and\nprivacy. Unlike previous work, which focused on centralized differential\nprivacy (DP) or local DP (LDP) for a single sensitive attribute, in this paper,\nwe examine the impact of LDP in the presence of several sensitive attributes\n(i.e., multi-dimensional data) on fairness. Detailed empirical analysis on\nsynthetic and benchmark datasets revealed very relevant observations. In\nparticular, (1) multi-dimensional LDP is an efficient approach to reduce\ndisparity, (2) the multi-dimensional approach of LDP (independent vs. combined)\nmatters only at low privacy guarantees, and (3) the outcome Y distribution has\nan important effect on which group is more sensitive to the obfuscation. Last,\nwe summarize our findings in the form of recommendations to guide practitioners\nin adopting effective privacy-preserving practices while maintaining fairness\nand utility in ML applications.\n", "date": "2023-12-07", "categories": []}
{"id": "2402.13379", "title": "Referee-Meta-Learning for Fast Adaptation of Locational Fairness", "abstract": "  When dealing with data from distinct locations, machine learning algorithms\ntend to demonstrate an implicit preference of some locations over the others,\nwhich constitutes biases that sabotage the spatial fairness of the algorithm.\nThis unfairness can easily introduce biases in subsequent decision-making given\nbroad adoptions of learning-based solutions in practice. However, locational\nbiases in AI are largely understudied. To mitigate biases over locations, we\npropose a locational meta-referee (Meta-Ref) to oversee the few-shot\nmeta-training and meta-testing of a deep neural network. Meta-Ref dynamically\nadjusts the learning rates for training samples of given locations to advocate\na fair performance across locations, through an explicit consideration of\nlocational biases and the characteristics of input data. We present a\nthree-phase training framework to learn both a meta-learning-based predictor\nand an integrated Meta-Ref that governs the fairness of the model. Once trained\nwith a distribution of spatial tasks, Meta-Ref is applied to samples from new\nspatial tasks (i.e., regions outside the training area) to promote fairness\nduring the fine-tune step. We carried out experiments with two case studies on\ncrop monitoring and transportation safety, which show Meta-Ref can improve\nlocational fairness while keeping the overall prediction quality at a similar\nlevel.\n", "date": "2024-02-20", "categories": []}
{"id": "2402.13771", "title": "Mask-up: Investigating Biases in Face Re-identification for Masked Faces", "abstract": "  AI based Face Recognition Systems (FRSs) are now widely distributed and\ndeployed as MLaaS solutions all over the world, moreso since the COVID-19\npandemic for tasks ranging from validating individuals' faces while buying SIM\ncards to surveillance of citizens. Extensive biases have been reported against\nmarginalized groups in these systems and have led to highly discriminatory\noutcomes. The post-pandemic world has normalized wearing face masks but FRSs\nhave not kept up with the changing times. As a result, these systems are\nsusceptible to mask based face occlusion. In this study, we audit four\ncommercial and nine open-source FRSs for the task of face re-identification\nbetween different varieties of masked and unmasked images across five benchmark\ndatasets (total 14,722 images). These simulate a realistic\nvalidation/surveillance task as deployed in all major countries around the\nworld. Three of the commercial and five of the open-source FRSs are highly\ninaccurate; they further perpetuate biases against non-White individuals, with\nthe lowest accuracy being 0%. A survey for the same task with 85 human\nparticipants also results in a low accuracy of 40%. Thus a human-in-the-loop\nmoderation in the pipeline does not alleviate the concerns, as has been\nfrequently hypothesized in literature. Our large-scale study shows that\ndevelopers, lawmakers and users of such services need to rethink the design\nprinciples behind FRSs, especially for the task of face re-identification,\ntaking cognizance of observed biases.\n", "date": "2024-02-21", "categories": []}
{"id": "2303.08485", "title": "Can Fairness be Automated? Guidelines and Opportunities for\n  Fairness-aware AutoML", "abstract": "  The field of automated machine learning (AutoML) introduces techniques that\nautomate parts of the development of machine learning (ML) systems,\naccelerating the process and reducing barriers for novices. However, decisions\nderived from ML models can reproduce, amplify, or even introduce unfairness in\nour societies, causing harm to (groups of) individuals. In response,\nresearchers have started to propose AutoML systems that jointly optimize\nfairness and predictive performance to mitigate fairness-related harm. However,\nfairness is a complex and inherently interdisciplinary subject, and solely\nposing it as an optimization problem can have adverse side effects. With this\nwork, we aim to raise awareness among developers of AutoML systems about such\nlimitations of fairness-aware AutoML, while also calling attention to the\npotential of AutoML as a tool for fairness research. We present a comprehensive\noverview of different ways in which fairness-related harm can arise and the\nensuing implications for the design of fairness-aware AutoML. We conclude that\nwhile fairness cannot be automated, fairness-aware AutoML can play an important\nrole in the toolbox of ML practitioners. We highlight several open technical\nchallenges for future work in this direction. Additionally, we advocate for the\ncreation of more user-centered assistive systems designed to tackle challenges\nencountered in fairness work\n", "date": "2023-03-15", "categories": []}
{"id": "2306.14518", "title": "Toward Fairness Through Fair Multi-Exit Framework for Dermatological\n  Disease Diagnosis", "abstract": "  Fairness has become increasingly pivotal in medical image recognition.\nHowever, without mitigating bias, deploying unfair medical AI systems could\nharm the interests of underprivileged populations. In this paper, we observe\nthat while features extracted from the deeper layers of neural networks\ngenerally offer higher accuracy, fairness conditions deteriorate as we extract\nfeatures from deeper layers. This phenomenon motivates us to extend the concept\nof multi-exit frameworks. Unlike existing works mainly focusing on accuracy,\nour multi-exit framework is fairness-oriented; the internal classifiers are\ntrained to be more accurate and fairer, with high extensibility to apply to\nmost existing fairness-aware frameworks. During inference, any instance with\nhigh confidence from an internal classifier is allowed to exit early.\nExperimental results show that the proposed framework can improve the fairness\ncondition over the state-of-the-art in two dermatological disease datasets.\n", "date": "2023-06-26", "categories": []}
{"id": "2402.13061", "title": "Toward Fairness via Maximum Mean Discrepancy Regularization on Logits\n  Space", "abstract": "  Fairness has become increasingly pivotal in machine learning for high-risk\napplications such as machine learning in healthcare and facial recognition.\nHowever, we see the deficiency in the previous logits space constraint methods.\nTherefore, we propose a novel framework, Logits-MMD, that achieves the fairness\ncondition by imposing constraints on output logits with Maximum Mean\nDiscrepancy. Moreover, quantitative analysis and experimental results show that\nour framework has a better property that outperforms previous methods and\nachieves state-of-the-art on two facial recognition datasets and one animal\ndataset. Finally, we show experimental results and demonstrate that our debias\napproach achieves the fairness condition effectively.\n", "date": "2024-02-20", "categories": []}
{"id": "2306.17828", "title": "On the Cause of Unfairness: A Training Sample Perspective", "abstract": "  Identifying the causes of a model's unfairness is an important yet relatively\nunexplored task. We look into this problem through the lens of training data -\nthe major source of unfairness. We ask the following questions: How would the\nunfairness of a model change if its training samples (1) were collected from a\ndifferent (e.g. demographic) group, (2) were labeled differently, or (3) whose\nfeatures were modified? In other words, we quantify the influence of training\nsamples on unfairness by counterfactually changing samples based on predefined\nconcepts, i.e. data attributes such as features, labels, and sensitive\nattributes. Our framework not only can help practitioners understand the\nobserved unfairness and mitigate it by repairing their training data, but also\nleads to many other applications, e.g. detecting mislabeling, fixing imbalanced\nrepresentations, and detecting fairness-targeted poisoning attacks.\n", "date": "2023-06-30", "categories": []}
{"id": "2401.16457", "title": "Effective Controllable Bias Mitigation for Classification and Retrieval\n  using Gate Adapters", "abstract": "  Bias mitigation of Language Models has been the topic of many studies with a\nrecent focus on learning separate modules like adapters for on-demand\ndebiasing. Besides optimizing for a modularized debiased model, it is often\ncritical in practice to control the degree of bias reduction at inference time,\ne.g., in order to tune for a desired performance-fairness trade-off in search\nresults or to control the strength of debiasing in classification tasks. In\nthis paper, we introduce Controllable Gate Adapter (ConGater), a novel modular\ngating mechanism with adjustable sensitivity parameters, which allows for a\ngradual transition from the biased state of the model to the fully debiased\nversion at inference time. We demonstrate ConGater performance by (1)\nconducting adversarial debiasing experiments with three different models on\nthree classification tasks with four protected attributes, and (2) reducing the\nbias of search results through fairness list-wise regularization to enable\nadjusting a trade-off between performance and fairness metrics. Our experiments\non the classification tasks show that compared to baselines of the same\ncaliber, ConGater can maintain higher task performance while containing less\ninformation regarding the attributes. Our results on the retrieval task show\nthat the fully debiased ConGater can achieve the same fairness performance\nwhile maintaining more than twice as high task performance than recent strong\nbaselines. Overall, besides strong performance ConGater enables the continuous\ntransitioning between biased and debiased states of models, enhancing\npersonalization of use and interpretability through controllability.\n", "date": "2024-01-29", "categories": []}
{"id": "2402.11190", "title": "Disclosure and Mitigation of Gender Bias in LLMs", "abstract": "  Large Language Models (LLMs) can generate biased responses. Yet previous\ndirect probing techniques contain either gender mentions or predefined gender\nstereotypes, which are challenging to comprehensively collect. Hence, we\npropose an indirect probing framework based on conditional generation. This\napproach aims to induce LLMs to disclose their gender bias even without\nexplicit gender or stereotype mentions. We explore three distinct strategies to\ndisclose explicit and implicit gender bias in LLMs. Our experiments demonstrate\nthat all tested LLMs exhibit explicit and/or implicit gender bias, even when\ngender stereotypes are not present in the inputs. In addition, an increased\nmodel size or model alignment amplifies bias in most cases. Furthermore, we\ninvestigate three methods to mitigate bias in LLMs via Hyperparameter Tuning,\nInstruction Guiding, and Debias Tuning. Remarkably, these methods prove\neffective even in the absence of explicit genders or stereotypes.\n", "date": "2024-02-17", "categories": []}
{"id": "2402.11710", "title": "A Note on Bias to Complete", "abstract": "  Minimizing social bias strengthens societal bonds, promoting shared\nunderstanding and better decision-making. We revisit the definition of bias by\ndiscovering new bias types (e.g., societal status) in dynamic environments and\ndescribe them relative to context, such as culture, region, time, and personal\nbackground. Our framework includes eight hypotheses about bias and a minimizing\nbias strategy for each assumption as well as five methods as proposed solutions\nin LLM. The realization of the framework is yet to be completed.\n", "date": "2024-02-18", "categories": []}
{"id": "2402.10756", "title": "Towards Cohesion-Fairness Harmony: Contrastive Regularization in\n  Individual Fair Graph Clustering", "abstract": "  Conventional fair graph clustering methods face two primary challenges: i)\nThey prioritize balanced clusters at the expense of cluster cohesion by\nimposing rigid constraints, ii) Existing methods of both individual and\ngroup-level fairness in graph partitioning mostly rely on eigen decompositions\nand thus, generally lack interpretability. To address these issues, we propose\niFairNMTF, an individual Fairness Nonnegative Matrix Tri-Factorization model\nwith contrastive fairness regularization that achieves balanced and cohesive\nclusters. By introducing fairness regularization, our model allows for\ncustomizable accuracy-fairness trade-offs, thereby enhancing user autonomy\nwithout compromising the interpretability provided by nonnegative matrix\ntri-factorization. Experimental evaluations on real and synthetic datasets\ndemonstrate the superior flexibility of iFairNMTF in achieving fairness and\nclustering performance.\n", "date": "2024-02-16", "categories": []}
{"id": "2402.10762", "title": "On Explaining Unfairness: An Overview", "abstract": "  Algorithmic fairness and explainability are foundational elements for\nachieving responsible AI. In this paper, we focus on their interplay, a\nresearch area that is recently receiving increasing attention. To this end, we\nfirst present two comprehensive taxonomies, each representing one of the two\ncomplementary fields of study: fairness and explanations. Then, we categorize\nexplanations for fairness into three types: (a) Explanations to enhance\nfairness metrics, (b) Explanations to help us understand the causes of\n(un)fairness, and (c) Explanations to assist us in designing methods for\nmitigating unfairness. Finally, based on our fairness and explanation\ntaxonomies, we present undiscovered literature paths revealing gaps that can\nserve as valuable insights for future research.\n", "date": "2024-02-16", "categories": []}
{"id": "2402.02816", "title": "Intersectional Two-sided Fairness in Recommendation", "abstract": "  Fairness of recommender systems (RS) has attracted increasing attention\nrecently. Based on the involved stakeholders, the fairness of RS can be divided\ninto user fairness, item fairness, and two-sided fairness which considers both\nuser and item fairness simultaneously. However, we argue that the\nintersectional two-sided unfairness may still exist even if the RS is two-sided\nfair, which is observed and shown by empirical studies on real-world data in\nthis paper, and has not been well-studied previously. To mitigate this problem,\nwe propose a novel approach called Intersectional Two-sided Fairness\nRecommendation (ITFR). Our method utilizes a sharpness-aware loss to perceive\ndisadvantaged groups, and then uses collaborative loss balance to develop\nconsistent distinguishing abilities for different intersectional groups.\nAdditionally, predicted score normalization is leveraged to align positive\npredicted scores to fairly treat positives in different intersectional groups.\nExtensive experiments and analyses on three public datasets show that our\nproposed approach effectively alleviates the intersectional two-sided\nunfairness and consistently outperforms previous state-of-the-art methods.\n", "date": "2024-02-05", "categories": []}
{"id": "2402.00910", "title": "Addressing Bias Through Ensemble Learning and Regularized Fine-Tuning", "abstract": "  Addressing biases in AI models is crucial for ensuring fair and accurate\npredictions. However, obtaining large, unbiased datasets for training can be\nchallenging. This paper proposes a comprehensive approach using multiple\nmethods to remove bias in AI models, with only a small dataset and a\npotentially biased pretrained model. We train multiple models with the\ncounter-bias of the pre-trained model through data splitting, local training,\nand regularized fine-tuning, gaining potentially counter-biased models. Then,\nwe employ ensemble learning for all models to reach unbiased predictions. To\nfurther accelerate the inference time of our ensemble model, we conclude our\nsolution with knowledge distillation that results in a single unbiased neural\nnetwork. We demonstrate the effectiveness of our approach through experiments\non the CIFAR10 and HAM10000 datasets, showcasing promising results. This work\ncontributes to the ongoing effort to create more unbiased and reliable AI\nmodels, even with limited data availability.\n", "date": "2024-02-01", "categories": []}
{"id": "2402.09326", "title": "Stability and Multigroup Fairness in Ranking with Uncertain Predictions", "abstract": "  Rankings are ubiquitous across many applications, from search engines to\nhiring committees. In practice, many rankings are derived from the output of\npredictors. However, when predictors trained for classification tasks have\nintrinsic uncertainty, it is not obvious how this uncertainty should be\nrepresented in the derived rankings. Our work considers ranking functions: maps\nfrom individual predictions for a classification task to distributions over\nrankings. We focus on two aspects of ranking functions: stability to\nperturbations in predictions and fairness towards both individuals and\nsubgroups. Not only is stability an important requirement for its own sake, but\n-- as we show -- it composes harmoniously with individual fairness in the sense\nof Dwork et al. (2012). While deterministic ranking functions cannot be stable\naside from trivial scenarios, we show that the recently proposed uncertainty\naware (UA) ranking functions of Singh et al. (2021) are stable. Our main result\nis that UA rankings also achieve multigroup fairness through successful\ncomposition with multiaccurate or multicalibrated predictors. Our work\ndemonstrates that UA rankings naturally interpolate between group and\nindividual level fairness guarantees, while simultaneously satisfying stability\nguarantees important whenever machine-learned predictions are used.\n", "date": "2024-02-14", "categories": []}
{"id": "2402.09328", "title": "Connecting Algorithmic Fairness to Quality Dimensions in Machine\n  Learning in Official Statistics and Survey Production", "abstract": "  National Statistical Organizations (NSOs) increasingly draw on Machine\nLearning (ML) to improve the timeliness and cost-effectiveness of their\nproducts. When introducing ML solutions, NSOs must ensure that high standards\nwith respect to robustness, reproducibility, and accuracy are upheld as\ncodified, e.g., in the Quality Framework for Statistical Algorithms (QF4SA;\nYung et al. 2022). At the same time, a growing body of research focuses on\nfairness as a pre-condition of a safe deployment of ML to prevent disparate\nsocial impacts in practice. However, fairness has not yet been explicitly\ndiscussed as a quality aspect in the context of the application of ML at NSOs.\nWe employ Yung et al. (2022)'s QF4SA quality framework and present a mapping of\nits quality dimensions to algorithmic fairness. We thereby extend the QF4SA\nframework in several ways: we argue for fairness as its own quality dimension,\nwe investigate the interaction of fairness with other dimensions, and we\nexplicitly address data, both on its own and its interaction with applied\nmethodology. In parallel with empirical illustrations, we show how our mapping\ncan contribute to methodology in the domains of official statistics,\nalgorithmic fairness, and trustworthy machine learning.\n", "date": "2024-02-14", "categories": []}
{"id": "2311.15108", "title": "Leveraging Diffusion Perturbations for Measuring Fairness in Computer\n  Vision", "abstract": "  Computer vision models have been known to encode harmful biases, leading to\nthe potentially unfair treatment of historically marginalized groups, such as\npeople of color. However, there remains a lack of datasets balanced along\ndemographic traits that can be used to evaluate the downstream fairness of\nthese models. In this work, we demonstrate that diffusion models can be\nleveraged to create such a dataset. We first use a diffusion model to generate\na large set of images depicting various occupations. Subsequently, each image\nis edited using inpainting to generate multiple variants, where each variant\nrefers to a different perceived race. Using this dataset, we benchmark several\nvision-language models on a multi-class occupation classification task. We find\nthat images generated with non-Caucasian labels have a significantly higher\noccupation misclassification rate than images generated with Caucasian labels,\nand that several misclassifications are suggestive of racial biases. We measure\na model's downstream fairness by computing the standard deviation in the\nprobability of predicting the true occupation label across the different\nperceived identity groups. Using this fairness metric, we find significant\ndisparities between the evaluated vision-and-language models. We hope that our\nwork demonstrates the potential value of diffusion methods for fairness\nevaluations.\n", "date": "2023-11-25", "categories": []}
{"id": "2402.07519", "title": "MAFIA: Multi-Adapter Fused Inclusive LanguAge Models", "abstract": "  Pretrained Language Models (PLMs) are widely used in NLP for various tasks.\nRecent studies have identified various biases that such models exhibit and have\nproposed methods to correct these biases. However, most of the works address a\nlimited set of bias dimensions independently such as gender, race, or religion.\nMoreover, the methods typically involve finetuning the full model to maintain\nthe performance on the downstream task. In this work, we aim to modularly\ndebias a pretrained language model across multiple dimensions. Previous works\nextensively explored debiasing PLMs using limited US-centric counterfactual\ndata augmentation (CDA). We use structured knowledge and a large generative\nmodel to build a diverse CDA across multiple bias dimensions in a\nsemi-automated way. We highlight how existing debiasing methods do not consider\ninteractions between multiple societal biases and propose a debiasing model\nthat exploits the synergy amongst various societal biases and enables\nmulti-bias debiasing simultaneously. An extensive evaluation on multiple tasks\nand languages demonstrates the efficacy of our approach.\n", "date": "2024-02-12", "categories": []}
{"id": "2402.05786", "title": "Prompting Fairness: Artificial Intelligence as Game Players", "abstract": "  Utilitarian games such as dictator games to measure fairness have been\nstudied in the social sciences for decades. These games have given us insight\ninto not only how humans view fairness but also in what conditions the\nfrequency of fairness, altruism and greed increase or decrease. While these\ngames have traditionally been focused on humans, the rise of AI gives us the\nability to study how these models play these games. AI is becoming a constant\nin human interaction and examining how these models portray fairness in game\nplay can give us some insight into how AI makes decisions. Over 101 rounds of\nthe dictator game, I conclude that AI has a strong sense of fairness that is\ndependant of it it deems the person it is playing with as trustworthy, framing\nhas a strong effect on how much AI gives a recipient when designated the\ntrustee, and there may be evidence that AI experiences inequality aversion just\nas humans.\n", "date": "2024-02-08", "categories": []}
{"id": "2207.11345", "title": "Toward Fairness in Speech Recognition: Discovery and mitigation of\n  performance disparities", "abstract": "  As for other forms of AI, speech recognition has recently been examined with\nrespect to performance disparities across different user cohorts. One approach\nto achieve fairness in speech recognition is to (1) identify speaker cohorts\nthat suffer from subpar performance and (2) apply fairness mitigation measures\ntargeting the cohorts discovered. In this paper, we report on initial findings\nwith both discovery and mitigation of performance disparities using data from a\nproduct-scale AI assistant speech recognition system. We compare cohort\ndiscovery based on geographic and demographic information to a more scalable\nmethod that groups speakers without human labels, using speaker embedding\ntechnology. For fairness mitigation, we find that oversampling of\nunderrepresented cohorts, as well as modeling speaker cohort membership by\nadditional input variables, reduces the gap between top- and bottom-performing\ncohorts, without deteriorating overall recognition accuracy.\n", "date": "2022-07-22", "categories": []}
{"id": "2402.05252", "title": "Learning Fair Ranking Policies via Differentiable Optimization of\n  Ordered Weighted Averages", "abstract": "  Learning to Rank (LTR) is one of the most widely used machine learning\napplications. It is a key component in platforms with profound societal\nimpacts, including job search, healthcare information retrieval, and social\nmedia content feeds. Conventional LTR models have been shown to produce biases\nresults, stimulating a discourse on how to address the disparities introduced\nby ranking systems that solely prioritize user relevance. However, while\nseveral models of fair learning to rank have been proposed, they suffer from\ndeficiencies either in accuracy or efficiency, thus limiting their\napplicability to real-world ranking platforms. This paper shows how\nefficiently-solvable fair ranking models, based on the optimization of Ordered\nWeighted Average (OWA) functions, can be integrated into the training loop of\nan LTR model to achieve favorable balances between fairness, user utility, and\nruntime efficiency. In particular, this paper is the first to show how to\nbackpropagate through constrained optimizations of OWA objectives, enabling\ntheir use in integrated prediction and decision models.\n", "date": "2024-02-07", "categories": []}
{"id": "2402.05575", "title": "Simultaneously Achieving Group Exposure Fairness and Within-Group\n  Meritocracy in Stochastic Bandits", "abstract": "  Existing approaches to fairness in stochastic multi-armed bandits (MAB)\nprimarily focus on exposure guarantee to individual arms. When arms are\nnaturally grouped by certain attribute(s), we propose Bi-Level Fairness, which\nconsiders two levels of fairness. At the first level, Bi-Level Fairness\nguarantees a certain minimum exposure to each group. To address the unbalanced\nallocation of pulls to individual arms within a group, we consider meritocratic\nfairness at the second level, which ensures that each arm is pulled according\nto its merit within the group. Our work shows that we can adapt a UCB-based\nalgorithm to achieve a Bi-Level Fairness by providing (i) anytime Group\nExposure Fairness guarantees and (ii) ensuring individual-level Meritocratic\nFairness within each group. We first show that one can decompose regret bounds\ninto two components: (a) regret due to anytime group exposure fairness and (b)\nregret due to meritocratic fairness within each group. Our proposed algorithm\nBF-UCB balances these two regrets optimally to achieve the upper bound of\n$O(\\sqrt{T})$ on regret; $T$ being the stopping time. With the help of\nsimulated experiments, we further show that BF-UCB achieves sub-linear regret;\nprovides better group and individual exposure guarantees compared to existing\nalgorithms; and does not result in a significant drop in reward with respect to\nUCB algorithm, which does not impose any fairness constraint.\n", "date": "2024-02-08", "categories": []}
{"id": "2402.04383", "title": "FairWire: Fair Graph Generation", "abstract": "  Machine learning over graphs has recently attracted growing attention due to\nits ability to analyze and learn complex relations within critical\ninterconnected systems. However, the disparate impact that is amplified by the\nuse of biased graph structures in these algorithms has raised significant\nconcerns for the deployment of them in real-world decision systems. In\naddition, while synthetic graph generation has become pivotal for privacy and\nscalability considerations, the impact of generative learning algorithms on the\nstructural bias has not yet been investigated. Motivated by this, this work\nfocuses on the analysis and mitigation of structural bias for both real and\nsynthetic graphs. Specifically, we first theoretically analyze the sources of\nstructural bias that result in disparity for the predictions of dyadic\nrelations. To alleviate the identified bias factors, we design a novel fairness\nregularizer that offers a versatile use. Faced with the bias amplification in\ngraph generation models that is brought to light in this work, we further\npropose a fair graph generation framework, FairWire, by leveraging our fair\nregularizer design in a generative model. Experimental results on real-world\nnetworks validate that the proposed tools herein deliver effective structural\nbias mitigation for both real and synthetic graphs.\n", "date": "2024-02-06", "categories": []}
{"id": "2402.04489", "title": "De-amplifying Bias from Differential Privacy in Language Model\n  Fine-tuning", "abstract": "  Fairness and privacy are two important values machine learning (ML)\npractitioners often seek to operationalize in models. Fairness aims to reduce\nmodel bias for social/demographic sub-groups. Privacy via differential privacy\n(DP) mechanisms, on the other hand, limits the impact of any individual's\ntraining data on the resulting model. The trade-offs between privacy and\nfairness goals of trustworthy ML pose a challenge to those wishing to address\nboth. We show that DP amplifies gender, racial, and religious bias when\nfine-tuning large language models (LLMs), producing models more biased than\nones fine-tuned without DP. We find the cause of the amplification to be a\ndisparity in convergence of gradients across sub-groups. Through the case of\nbinary gender bias, we demonstrate that Counterfactual Data Augmentation (CDA),\na known method for addressing bias, also mitigates bias amplification by DP. As\na consequence, DP and CDA together can be used to fine-tune models while\nmaintaining both fairness and privacy.\n", "date": "2024-02-07", "categories": []}
{"id": "2310.19391", "title": "Causal Fair Metric: Bridging Causality, Individual Fairness, and\n  Adversarial Robustness", "abstract": "  Despite the essential need for comprehensive considerations in responsible\nAI, factors like robustness, fairness, and causality are often studied in\nisolation. Adversarial perturbation, used to identify vulnerabilities in\nmodels, and individual fairness, aiming for equitable treatment of similar\nindividuals, despite initial differences, both depend on metrics to generate\ncomparable input data instances. Previous attempts to define such joint metrics\noften lack general assumptions about data or structural causal models and were\nunable to reflect counterfactual proximity. To address this, our paper\nintroduces a causal fair metric formulated based on causal structures\nencompassing sensitive attributes and protected causal perturbation. To enhance\nthe practicality of our metric, we propose metric learning as a method for\nmetric estimation and deployment in real-world problems in the absence of\nstructural causal models. We also demonstrate the application of our novel\nmetric in classifiers. Empirical evaluation of real-world and synthetic\ndatasets illustrates the effectiveness of our proposed metric in achieving an\naccurate classifier with fairness, resilience to adversarial perturbations, and\na nuanced understanding of causal relationships.\n", "date": "2023-10-30", "categories": []}
{"id": "2401.15270", "title": "SimFair: Physics-Guided Fairness-Aware Learning with Simulation Models", "abstract": "  Fairness-awareness has emerged as an essential building block for the\nresponsible use of artificial intelligence in real applications. In many cases,\ninequity in performance is due to the change in distribution over different\nregions. While techniques have been developed to improve the transferability of\nfairness, a solution to the problem is not always feasible with no samples from\nthe new regions, which is a bottleneck for pure data-driven attempts.\nFortunately, physics-based mechanistic models have been studied for many\nproblems with major social impacts. We propose SimFair, a physics-guided\nfairness-aware learning framework, which bridges the data limitation by\nintegrating physical-rule-based simulation and inverse modeling into the\ntraining design. Using temperature prediction as an example, we demonstrate the\neffectiveness of the proposed SimFair in fairness preservation.\n", "date": "2024-01-27", "categories": []}
{"id": "2402.03011", "title": "On the Impact of Output Perturbation on Fairness in Binary Linear\n  Classification", "abstract": "  We theoretically study how differential privacy interacts with both\nindividual and group fairness in binary linear classification. More precisely,\nwe focus on the output perturbation mechanism, a classic approach in\nprivacy-preserving machine learning. We derive high-probability bounds on the\nlevel of individual and group fairness that the perturbed models can achieve\ncompared to the original model. Hence, for individual fairness, we prove that\nthe impact of output perturbation on the level of fairness is bounded but grows\nwith the dimension of the model. For group fairness, we show that this impact\nis determined by the distribution of so-called angular margins, that is signed\nmargins of the non-private model re-scaled by the norm of each example.\n", "date": "2024-02-05", "categories": []}
{"id": "2402.03531", "title": "Fairness and Privacy Guarantees in Federated Contextual Bandits", "abstract": "  This paper considers the contextual multi-armed bandit (CMAB) problem with\nfairness and privacy guarantees in a federated environment. We consider\nmerit-based exposure as the desired fair outcome, which provides exposure to\neach action in proportion to the reward associated. We model the algorithm's\neffectiveness using fairness regret, which captures the difference between fair\noptimal policy and the policy output by the algorithm. Applying fair CMAB\nalgorithm to each agent individually leads to fairness regret linear in the\nnumber of agents. We propose that collaborative -- federated learning can be\nmore effective and provide the algorithm Fed-FairX-LinUCB that also ensures\ndifferential privacy. The primary challenge in extending the existing privacy\nframework is designing the communication protocol for communicating required\ninformation across agents. A naive protocol can either lead to weaker privacy\nguarantees or higher regret. We design a novel communication protocol that\nallows for (i) Sub-linear theoretical bounds on fairness regret for\nFed-FairX-LinUCB and comparable bounds for the private counterpart,\nPriv-FairX-LinUCB (relative to single-agent learning), (ii) Effective use of\nprivacy budget in Priv-FairX-LinUCB. We demonstrate the efficacy of our\nproposed algorithm with extensive simulations-based experiments. We show that\nboth Fed-FairX-LinUCB and Priv-FairX-LinUCB achieve near-optimal fairness\nregret.\n", "date": "2024-02-05", "categories": []}
{"id": "2402.01811", "title": "A Distributionally Robust Optimisation Approach to Fair Credit Scoring", "abstract": "  Credit scoring has been catalogued by the European Commission and the\nExecutive Office of the US President as a high-risk classification task, a key\nconcern being the potential harms of making loan approval decisions based on\nmodels that would be biased against certain groups. To address this concern,\nrecent credit scoring research has considered a range of fairness-enhancing\ntechniques put forward by the machine learning community to reduce bias and\nunfair treatment in classification systems. While the definition of fairness or\nthe approach they follow to impose it may vary, most of these techniques,\nhowever, disregard the robustness of the results. This can create situations\nwhere unfair treatment is effectively corrected in the training set, but when\nproducing out-of-sample classifications, unfair treatment is incurred again.\nInstead, in this paper, we will investigate how to apply Distributionally\nRobust Optimisation (DRO) methods to credit scoring, thereby empirically\nevaluating how they perform in terms of fairness, ability to classify\ncorrectly, and the robustness of the solution against changes in the marginal\nproportions. In so doing, we find DRO methods to provide a substantial\nimprovement in terms of fairness, with almost no loss in performance. These\nresults thus indicate that DRO can improve fairness in credit scoring, provided\nthat further advances are made in efficiently implementing these systems. In\naddition, our analysis suggests that many of the commonly used fairness metrics\nare unsuitable for a credit scoring setting, as they depend on the choice of\nclassification threshold.\n", "date": "2024-02-02", "categories": []}
{"id": "2402.01472", "title": "Synthetic Data for the Mitigation of Demographic Biases in Face\n  Recognition", "abstract": "  This study investigates the possibility of mitigating the demographic biases\nthat affect face recognition technologies through the use of synthetic data.\nDemographic biases have the potential to impact individuals from specific\ndemographic groups, and can be identified by observing disparate performance of\nface recognition systems across demographic groups. They primarily arise from\nthe unequal representations of demographic groups in the training data. In\nrecent times, synthetic data have emerged as a solution to some problems that\naffect face recognition systems. In particular, during the generation process\nit is possible to specify the desired demographic and facial attributes of\nimages, in order to control the demographic distribution of the synthesized\ndataset, and fairly represent the different demographic groups. We propose to\nfine-tune with synthetic data existing face recognition systems that present\nsome demographic biases. We use synthetic datasets generated with GANDiffFace,\na novel framework able to synthesize datasets for face recognition with\ncontrollable demographic distribution and realistic intra-class variations. We\nconsider multiple datasets representing different demographic groups for\ntraining and evaluation. Also, we fine-tune different face recognition systems,\nand evaluate their demographic fairness with different metrics. Our results\nsupport the proposed approach and the use of synthetic data to mitigate\ndemographic biases in face recognition.\n", "date": "2024-02-02", "categories": []}
{"id": "2206.14397", "title": "Fair Machine Learning in Healthcare: A Review", "abstract": "  The digitization of healthcare data coupled with advances in computational\ncapabilities has propelled the adoption of machine learning (ML) in healthcare.\nHowever, these methods can perpetuate or even exacerbate existing disparities,\nleading to fairness concerns such as the unequal distribution of resources and\ndiagnostic inaccuracies among different demographic groups. Addressing these\nfairness problem is paramount to prevent further entrenchment of social\ninjustices. In this survey, we analyze the intersection of fairness in machine\nlearning and healthcare disparities. We adopt a framework based on the\nprinciples of distributive justice to categorize fairness concerns into two\ndistinct classes: equal allocation and equal performance. We provide a critical\nreview of the associated fairness metrics from a machine learning standpoint\nand examine biases and mitigation strategies across the stages of the ML\nlifecycle, discussing the relationship between biases and their\ncountermeasures. The paper concludes with a discussion on the pressing\nchallenges that remain unaddressed in ensuring fairness in healthcare ML, and\nproposes several new research directions that hold promise for developing\nethical and equitable ML applications in healthcare.\n", "date": "2022-06-29", "categories": []}
{"id": "2401.16784", "title": "Graph Fairness Learning under Distribution Shifts", "abstract": "  Graph neural networks (GNNs) have achieved remarkable performance on\ngraph-structured data. However, GNNs may inherit prejudice from the training\ndata and make discriminatory predictions based on sensitive attributes, such as\ngender and race. Recently, there has been an increasing interest in ensuring\nfairness on GNNs, but all of them are under the assumption that the training\nand testing data are under the same distribution, i.e., training data and\ntesting data are from the same graph. Will graph fairness performance decrease\nunder distribution shifts? How does distribution shifts affect graph fairness\nlearning? All these open questions are largely unexplored from a theoretical\nperspective. To answer these questions, we first theoretically identify the\nfactors that determine bias on a graph. Subsequently, we explore the factors\ninfluencing fairness on testing graphs, with a noteworthy factor being the\nrepresentation distances of certain groups between the training and testing\ngraph. Motivated by our theoretical analysis, we propose our framework\nFatraGNN. Specifically, to guarantee fairness performance on unknown testing\ngraphs, we propose a graph generator to produce numerous graphs with\nsignificant bias and under different distributions. Then we minimize the\nrepresentation distances for each certain group between the training graph and\ngenerated graphs. This empowers our model to achieve high classification and\nfairness performance even on generated graphs with significant bias, thereby\neffectively handling unknown testing graphs. Experiments on real-world and\nsemi-synthetic datasets demonstrate the effectiveness of our model in terms of\nboth accuracy and fairness.\n", "date": "2024-01-30", "categories": []}
{"id": "2311.04892", "title": "Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMs", "abstract": "  Recent works have showcased the ability of LLMs to embody diverse personas in\ntheir responses, exemplified by prompts like 'You are Yoda. Explain the Theory\nof Relativity.' While this ability allows personalization of LLMs and enables\nhuman behavior simulation, its effect on LLMs' capabilities remains unclear. To\nfill this gap, we present the first extensive study of the unintended\nside-effects of persona assignment on the ability of LLMs to perform basic\nreasoning tasks. Our study covers 24 reasoning datasets, 4 LLMs, and 19 diverse\npersonas (e.g. an Asian person) spanning 5 socio-demographic groups. Our\nexperiments unveil that LLMs harbor deep rooted bias against various\nsocio-demographics underneath a veneer of fairness. While they overtly reject\nstereotypes when explicitly asked ('Are Black people less skilled at\nmathematics?'), they manifest stereotypical and erroneous presumptions when\nasked to answer questions while adopting a persona. These can be observed as\nabstentions in responses, e.g., 'As a Black person, I can't answer this\nquestion as it requires math knowledge', and generally result in a substantial\nperformance drop. Our experiments with ChatGPT-3.5 show that this bias is\nubiquitous - 80% of our personas demonstrate bias; it is significant - some\ndatasets show performance drops of 70%+; and can be especially harmful for\ncertain groups - some personas suffer statistically significant drops on 80%+\nof the datasets. Overall, all 4 LLMs exhibit this bias to varying extents, with\nGPT-4-Turbo showing the least but still a problematic amount of bias (evident\nin 42% of the personas). Further analysis shows that these persona-induced\nerrors can be hard-to-discern and hard-to-avoid. Our findings serve as a\ncautionary tale that the practice of assigning personas to LLMs - a trend on\nthe rise - can surface their deep-rooted biases and have unforeseeable and\ndetrimental side-effects.\n", "date": "2023-11-08", "categories": []}
{"id": "2401.16088", "title": "Fairness in Algorithmic Recourse Through the Lens of Substantive\n  Equality of Opportunity", "abstract": "  Algorithmic recourse -- providing recommendations to those affected\nnegatively by the outcome of an algorithmic system on how they can take action\nand change that outcome -- has gained attention as a means of giving persons\nagency in their interactions with artificial intelligence (AI) systems. Recent\nwork has shown that even if an AI decision-making classifier is ``fair''\n(according to some reasonable criteria), recourse itself may be unfair due to\ndifferences in the initial circumstances of individuals, compounding\ndisparities for marginalized populations and requiring them to exert more\neffort than others. There is a need to define more methods and metrics for\nevaluating fairness in recourse that span a range of normative views of the\nworld, and specifically those that take into account time. Time is a critical\nelement in recourse because the longer it takes an individual to act, the more\nthe setting may change due to model or data drift.\n  This paper seeks to close this research gap by proposing two notions of\nfairness in recourse that are in normative alignment with substantive equality\nof opportunity, and that consider time. The first considers the (often\nrepeated) effort individuals exert per successful recourse event, and the\nsecond considers time per successful recourse event. Building upon an\nagent-based framework for simulating recourse, this paper demonstrates how much\neffort is needed to overcome disparities in initial circumstances. We then\nproposes an intervention to improve the fairness of recourse by rewarding\neffort, and compare it to existing strategies.\n", "date": "2024-01-29", "categories": []}
{"id": "2401.16197", "title": "Geospatial Disparities: A Case Study on Real Estate Prices in Paris", "abstract": "  Driven by an increasing prevalence of trackers, ever more IoT sensors, and\nthe declining cost of computing power, geospatial information has come to play\na pivotal role in contemporary predictive models. While enhancing prognostic\nperformance, geospatial data also has the potential to perpetuate many\nhistorical socio-economic patterns, raising concerns about a resurgence of\nbiases and exclusionary practices, with their disproportionate impacts on\nsociety. Addressing this, our paper emphasizes the crucial need to identify and\nrectify such biases and calibration errors in predictive models, particularly\nas algorithms become more intricate and less interpretable. The increasing\ngranularity of geospatial information further introduces ethical concerns, as\nchoosing different geographical scales may exacerbate disparities akin to\nredlining and exclusionary zoning. To address these issues, we propose a\ntoolkit for identifying and mitigating biases arising from geospatial data.\nExtending classical fairness definitions, we incorporate an ordinal regression\ncase with spatial attributes, deviating from the binary classification focus.\nThis extension allows us to gauge disparities stemming from data aggregation\nlevels and advocates for a less interfering correction approach. Illustrating\nour methodology using a Parisian real estate dataset, we showcase practical\napplications and scrutinize the implications of choosing geographical\naggregation levels for fairness and calibration measures.\n", "date": "2024-01-29", "categories": []}
{"id": "2401.16350", "title": "FedFair^3: Unlocking Threefold Fairness in Federated Learning", "abstract": "  Federated Learning (FL) is an emerging paradigm in machine learning without\nexposing clients' raw data. In practical scenarios with numerous clients,\nencouraging fair and efficient client participation in federated learning is of\nutmost importance, which is also challenging given the heterogeneity in data\ndistribution and device properties. Existing works have proposed different\nclient-selection methods that consider fairness; however, they fail to select\nclients with high utilities while simultaneously achieving fair accuracy\nlevels. In this paper, we propose a fair client-selection approach that unlocks\nthreefold fairness in federated learning. In addition to having a fair\nclient-selection strategy, we enforce an equitable number of rounds for client\nparticipation and ensure a fair accuracy distribution over the clients. The\nexperimental results demonstrate that FedFair^3, in comparison to the\nstate-of-the-art baselines, achieves 18.15% less accuracy variance on the IID\ndata and 54.78% on the non-IID data, without decreasing the global accuracy.\nFurthermore, it shows 24.36% less wall-clock training time on average.\n", "date": "2024-01-29", "categories": []}
{"id": "2401.14702", "title": "FairSample: Training Fair and Accurate Graph Convolutional Neural\n  Networks Efficiently", "abstract": "  Fairness in Graph Convolutional Neural Networks (GCNs) becomes a more and\nmore important concern as GCNs are adopted in many crucial applications.\nSocietal biases against sensitive groups may exist in many real world graphs.\nGCNs trained on those graphs may be vulnerable to being affected by such\nbiases. In this paper, we adopt the well-known fairness notion of demographic\nparity and tackle the challenge of training fair and accurate GCNs efficiently.\nWe present an in-depth analysis on how graph structure bias, node attribute\nbias, and model parameters may affect the demographic parity of GCNs. Our\ninsights lead to FairSample, a framework that jointly mitigates the three types\nof biases. We employ two intuitive strategies to rectify graph structures.\nFirst, we inject edges across nodes that are in different sensitive groups but\nsimilar in node features. Second, to enhance model fairness and retain model\nquality, we develop a learnable neighbor sampling policy using reinforcement\nlearning. To address the bias in node features and model parameters, FairSample\nis complemented by a regularization objective to optimize fairness.\n", "date": "2024-01-26", "categories": []}
{"id": "2307.16778", "title": "KoBBQ: Korean Bias Benchmark for Question Answering", "abstract": "  The Bias Benchmark for Question Answering (BBQ) is designed to evaluate\nsocial biases of language models (LMs), but it is not simple to adapt this\nbenchmark to cultural contexts other than the US because social biases depend\nheavily on the cultural context. In this paper, we present KoBBQ, a Korean bias\nbenchmark dataset, and we propose a general framework that addresses\nconsiderations for cultural adaptation of a dataset. Our framework includes\npartitioning the BBQ dataset into three classes--Simply-Transferred (can be\nused directly after cultural translation), Target-Modified (requires\nlocalization in target groups), and Sample-Removed (does not fit Korean\nculture)-- and adding four new categories of bias specific to Korean culture.\nWe conduct a large-scale survey to collect and validate the social biases and\nthe targets of the biases that reflect the stereotypes in Korean culture. The\nresulting KoBBQ dataset comprises 268 templates and 76,048 samples across 12\ncategories of social bias. We use KoBBQ to measure the accuracy and bias scores\nof several state-of-the-art multilingual LMs. The results clearly show\ndifferences in the bias of LMs as measured by KoBBQ and a machine-translated\nversion of BBQ, demonstrating the need for and utility of a well-constructed,\nculturally-aware social bias benchmark.\n", "date": "2023-07-31", "categories": []}
{"id": "2401.13935", "title": "A New Paradigm for Counterfactual Reasoning in Fairness and Recourse", "abstract": "  Counterfactuals and counterfactual reasoning underpin numerous techniques for\nauditing and understanding artificial intelligence (AI) systems. The\ntraditional paradigm for counterfactual reasoning in this literature is the\ninterventional counterfactual, where hypothetical interventions are imagined\nand simulated. For this reason, the starting point for causal reasoning about\nlegal protections and demographic data in AI is an imagined intervention on a\nlegally-protected characteristic, such as ethnicity, race, gender, disability,\nage, etc. We ask, for example, what would have happened had your race been\ndifferent? An inherent limitation of this paradigm is that some demographic\ninterventions -- like interventions on race -- may not translate into the\nformalisms of interventional counterfactuals. In this work, we explore a new\nparadigm based instead on the backtracking counterfactual, where rather than\nimagine hypothetical interventions on legally-protected characteristics, we\nimagine alternate initial conditions while holding these characteristics fixed.\nWe ask instead, what would explain a counterfactual outcome for you as you\nactually are or could be? This alternate framework allows us to address many of\nthe same social concerns, but to do so while asking fundamentally different\nquestions that do not rely on demographic interventions.\n", "date": "2024-01-25", "categories": []}
{"id": "2401.14343", "title": "Class-attribute Priors: Adapting Optimization to Heterogeneity and\n  Fairness Objective", "abstract": "  Modern classification problems exhibit heterogeneities across individual\nclasses: Each class may have unique attributes, such as sample size, label\nquality, or predictability (easy vs difficult), and variable importance at\ntest-time. Without care, these heterogeneities impede the learning process,\nmost notably, when optimizing fairness objectives. Confirming this, under a\ngaussian mixture setting, we show that the optimal SVM classifier for balanced\naccuracy needs to be adaptive to the class attributes. This motivates us to\npropose CAP: An effective and general method that generates a class-specific\nlearning strategy (e.g. hyperparameter) based on the attributes of that class.\nThis way, optimization process better adapts to heterogeneities. CAP leads to\nsubstantial improvements over the naive approach of assigning separate\nhyperparameters to each class. We instantiate CAP for loss function design and\npost-hoc logit adjustment, with emphasis on label-imbalanced problems. We show\nthat CAP is competitive with prior art and its flexibility unlocks clear\nbenefits for fairness objectives beyond balanced accuracy. Finally, we evaluate\nCAP on problems with label noise as well as weighted test objectives to\nshowcase how CAP can jointly adapt to different heterogeneities.\n", "date": "2024-01-25", "categories": []}
{"id": "2208.06680", "title": "Locating disparities in machine learning", "abstract": "  Machine learning can provide predictions with disparate outcomes, in which\nsubgroups of the population (e.g., defined by age, gender, or other sensitive\nattributes) are systematically disadvantaged. In order to comply with upcoming\nlegislation, practitioners need to locate such disparate outcomes. However,\nprevious literature typically detects disparities through statistical\nprocedures for when the sensitive attribute is specified a priori. This limits\napplicability in real-world settings where datasets are high dimensional and,\non top of that, sensitive attributes may be unknown. As a remedy, we propose a\ndata-driven framework called Automatic Location of Disparities (ALD) which aims\nat locating disparities in machine learning. ALD meets several demands from\nindustry: ALD (1) is applicable to arbitrary machine learning classifiers; (2)\noperates on different definitions of disparities (e.g., statistical parity or\nequalized odds); and (3) deals with both categorical and continuous predictors\neven if disparities arise from complex and multi-way interactions known as\nintersectionality (e. g., age above 60 and female). ALD produces interpretable\naudit reports as output. We demonstrate the effectiveness of ALD based on both\nsynthetic and real-world datasets. As a result, we empower practitioners to\neffectively locate and mitigate disparities in machine learning algorithms,\nconduct algorithmic audits, and protect individuals from discrimination.\n", "date": "2022-08-13", "categories": []}
{"id": "2302.11944", "title": "Counterfactual Situation Testing: Uncovering Discrimination under\n  Fairness given the Difference", "abstract": "  We present counterfactual situation testing (CST), a causal data mining\nframework for detecting discrimination in classifiers. CST aims to answer in an\nactionable and meaningful way the intuitive question \"what would have been the\nmodel outcome had the individual, or complainant, been of a different protected\nstatus?\" It extends the legally-grounded situation testing of Thanh et al.\n(2011) by operationalizing the notion of fairness given the difference using\ncounterfactual reasoning. For any complainant, we find and compare similar\nprotected and non-protected instances in the dataset used by the classifier to\nconstruct a control and test group, where a difference between the decision\noutcomes of the two groups implies potential individual discrimination. Unlike\nsituation testing, which builds both groups around the complainant, we build\nthe test group on the complainant's counterfactual generated using causal\nknowledge. The counterfactual is intended to reflect how the protected\nattribute when changed affects the seemingly neutral attributes used by the\nclassifier, which is taken for granted in many frameworks for discrimination.\nUnder CST, we compare similar individuals within each group but dissimilar\nindividuals across both groups due to the possible difference between the\ncomplainant and its counterfactual. Evaluating our framework on two\nclassification scenarios, we show that it uncovers a greater number of cases\nthan situation testing, even when the classifier satisfies the counterfactual\nfairness condition of Kusner et al. (2017).\n", "date": "2023-02-23", "categories": []}
{"id": "2309.04447", "title": "Impact of Blur and Resolution on Demographic Disparities in 1-to-Many\n  Facial Identification", "abstract": "  Most studies to date that have examined demographic variations in face\nrecognition accuracy have analyzed 1-to-1 matching accuracy, using images that\ncould be described as \"government ID quality\". This paper analyzes the accuracy\nof 1-to-many facial identification across demographic groups, and in the\npresence of blur and reduced resolution in the probe image as might occur in\n\"surveillance camera quality\" images. Cumulative match characteristic curves\n(CMC) are not appropriate for comparing propensity for rank-one recognition\nerrors across demographics, and so we use three metrics for our analysis: (1)\nthe well-known d' metric between mated and non-mated score distributions, and\nintroduced in this work, (2) absolute score difference between thresholds in\nthe high-similarity tail of the non-mated and the low-similarity tail of the\nmated distribution, and (3) distribution of (mated - non-mated rank-one scores)\nacross the set of probe images. We find that demographic variation in 1-to-many\naccuracy does not entirely follow what has been observed in 1-to-1 matching\naccuracy. Also, different from 1-to-1 accuracy, demographic comparison of\n1-to-many accuracy can be affected by different numbers of identities and\nimages across demographics. More importantly, we show that increased blur in\nthe probe image, or reduced resolution of the face in the probe image, can\nsignificantly increase the false positive identification rate. And we show that\nthe demographic variation in these high blur or low resolution conditions is\nmuch larger for male / female than for African-American / Caucasian. The point\nthat 1-to-many accuracy can potentially collapse in the context of processing\n\"surveillance camera quality\" probe images against a \"government ID quality\"\ngallery is an important one.\n", "date": "2023-09-08", "categories": []}
{"id": "2310.03258", "title": "Assessing Electricity Service Unfairness with Transfer Counterfactual\n  Learning", "abstract": "  Energy justice is a growing area of interest in interdisciplinary energy\nresearch. However, identifying systematic biases in the energy sector remains\nchallenging due to confounding variables, intricate heterogeneity in\ncounterfactual effects, and limited data availability. First, this paper\ndemonstrates how one can evaluate counterfactual unfairness in a power system\nby analyzing the average causal effect of a specific protected attribute.\nSubsequently, we use subgroup analysis to handle model heterogeneity and\nintroduce a novel method for estimating counterfactual unfairness based on\ntransfer learning, which helps to alleviate the data scarcity in each subgroup.\nIn our numerical analysis, we apply our method to a unique large-scale\ncustomer-level power outage data set and investigate the counterfactual effect\nof demographic factors, such as income and age of the population, on power\noutage durations. Our results indicate that low-income and elderly-populated\nareas consistently experience longer power outages under both daily and\npost-disaster operations, and such discrimination is exacerbated under severe\nconditions. These findings suggest a widespread, systematic issue of injustice\nin the power service systems and emphasize the necessity for focused\ninterventions in disadvantaged communities.\n", "date": "2023-10-05", "categories": []}
{"id": "2306.14624", "title": "Insights From Insurance for Fair Machine Learning", "abstract": "  We argue that insurance can act as an analogon for the social situatedness of\nmachine learning systems, hence allowing machine learning scholars to take\ninsights from the rich and interdisciplinary insurance literature. Tracing the\ninteraction of uncertainty, fairness and responsibility in insurance provides a\nfresh perspective on fairness in machine learning. We link insurance fairness\nconceptions to their machine learning relatives, and use this bridge to\nproblematize fairness as calibration. In this process, we bring to the\nforefront two themes that have been largely overlooked in the machine learning\nliterature: responsibility and aggregate-individual tensions.\n", "date": "2023-06-26", "categories": []}
{"id": "2401.12720", "title": "A Comprehensive View of the Biases of Toxicity and Sentiment Analysis\n  Methods Towards Utterances with African American English Expressions", "abstract": "  Language is a dynamic aspect of our culture that changes when expressed in\ndifferent technologies/communities. Online social networks have enabled the\ndiffusion and evolution of different dialects, including African American\nEnglish (AAE). However, this increased usage is not without barriers. One\nparticular barrier is how sentiment (Vader, TextBlob, and Flair) and toxicity\n(Google's Perspective and the open-source Detoxify) methods present biases\ntowards utterances with AAE expressions. Consider Google's Perspective to\nunderstand bias. Here, an utterance such as ``All n*ggers deserve to die\nrespectfully. The police murder us.'' it reaches a higher toxicity than\n``African-Americans deserve to die respectfully. The police murder us.''. This\nscore difference likely arises because the tool cannot understand the\nre-appropriation of the term ``n*gger''. One explanation for this bias is that\nAI models are trained on limited datasets, and using such a term in training\ndata is more likely to appear in a toxic utterance. While this may be\nplausible, the tool will make mistakes regardless. Here, we study bias on two\nWeb-based (YouTube and Twitter) datasets and two spoken English datasets. Our\nanalysis shows how most models present biases towards AAE in most settings. We\nisolate the impact of AAE expression usage via linguistic control features from\nthe Linguistic Inquiry and Word Count (LIWC) software, grammatical control\nfeatures extracted via Part-of-Speech (PoS) tagging from Natural Language\nProcessing (NLP) models, and the semantic of utterances by comparing sentence\nembeddings from recent language models. We present consistent results on how a\nheavy usage of AAE expressions may cause the speaker to be considered\nsubstantially more toxic, even when speaking about nearly the same subject. Our\nstudy complements similar analyses focusing on small datasets and/or one method\nonly.\n", "date": "2024-01-23", "categories": []}
{"id": "2401.11288", "title": "Long-Term Fair Decision Making through Deep Generative Models", "abstract": "  This paper studies long-term fair machine learning which aims to mitigate\ngroup disparity over the long term in sequential decision-making systems. To\ndefine long-term fairness, we leverage the temporal causal graph and use the\n1-Wasserstein distance between the interventional distributions of different\ndemographic groups at a sufficiently large time step as the quantitative\nmetric. Then, we propose a three-phase learning framework where the decision\nmodel is trained on high-fidelity data generated by a deep generative model. We\nformulate the optimization problem as a performative risk minimization and\nadopt the repeated gradient descent algorithm for learning. The empirical\nevaluation shows the efficacy of the proposed method using both synthetic and\nsemi-synthetic datasets.\n", "date": "2024-01-20", "categories": []}
{"id": "2401.11601", "title": "Robust Evaluation Measures for Evaluating Social Biases in Masked\n  Language Models", "abstract": "  Many evaluation measures are used to evaluate social biases in masked\nlanguage models (MLMs). However, we find that these previously proposed\nevaluation measures are lacking robustness in scenarios with limited datasets.\nThis is because these measures are obtained by comparing the\npseudo-log-likelihood (PLL) scores of the stereotypical and anti-stereotypical\nsamples using an indicator function. The disadvantage is the limited mining of\nthe PLL score sets without capturing its distributional information. In this\npaper, we represent a PLL score set as a Gaussian distribution and use Kullback\nLeibler (KL) divergence and Jensen Shannon (JS) divergence to construct\nevaluation measures for the distributions of stereotypical and\nanti-stereotypical PLL scores. Experimental results on the publicly available\ndatasets StereoSet (SS) and CrowS-Pairs (CP) show that our proposed measures\nare significantly more robust and interpretable than those proposed previously.\n", "date": "2024-01-21", "categories": []}
{"id": "2401.10535", "title": "The \"Colonial Impulse\" of Natural Language Processing: An Audit of\n  Bengali Sentiment Analysis Tools and Their Identity-based Biases", "abstract": "  While colonization has sociohistorically impacted people's identities across\nvarious dimensions, those colonial values and biases continue to be perpetuated\nby sociotechnical systems. One category of sociotechnical systems--sentiment\nanalysis tools--can also perpetuate colonial values and bias, yet less\nattention has been paid to how such tools may be complicit in perpetuating\ncoloniality, although they are often used to guide various practices (e.g.,\ncontent moderation). In this paper, we explore potential bias in sentiment\nanalysis tools in the context of Bengali communities that have experienced and\ncontinue to experience the impacts of colonialism. Drawing on identity\ncategories most impacted by colonialism amongst local Bengali communities, we\nfocused our analytic attention on gender, religion, and nationality. We\nconducted an algorithmic audit of all sentiment analysis tools for Bengali,\navailable on the Python package index (PyPI) and GitHub. Despite similar\nsemantic content and structure, our analyses showed that in addition to\ninconsistencies in output from different tools, Bengali sentiment analysis\ntools exhibit bias between different identity categories and respond\ndifferently to different ways of identity expression. Connecting our findings\nwith colonially shaped sociocultural structures of Bengali communities, we\ndiscuss the implications of downstream bias of sentiment analysis tools.\n", "date": "2024-01-19", "categories": []}
{"id": "2310.05055", "title": "FairTune: Optimizing Parameter Efficient Fine Tuning for Fairness in\n  Medical Image Analysis", "abstract": "  Training models with robust group fairness properties is crucial in ethically\nsensitive application areas such as medical diagnosis. Despite the growing body\nof work aiming to minimise demographic bias in AI, this problem remains\nchallenging. A key reason for this challenge is the fairness generalisation\ngap: High-capacity deep learning models can fit all training data nearly\nperfectly, and thus also exhibit perfect fairness during training. In this\ncase, bias emerges only during testing when generalisation performance differs\nacross subgroups. This motivates us to take a bi-level optimisation perspective\non fair learning: Optimising the learning strategy based on validation\nfairness. Specifically, we consider the highly effective workflow of adapting\npre-trained models to downstream medical imaging tasks using\nparameter-efficient fine-tuning (PEFT) techniques. There is a trade-off between\nupdating more parameters, enabling a better fit to the task of interest vs.\nfewer parameters, potentially reducing the generalisation gap. To manage this\ntradeoff, we propose FairTune, a framework to optimise the choice of PEFT\nparameters with respect to fairness. We demonstrate empirically that FairTune\nleads to improved fairness on a range of medical imaging datasets. The code is\navailable at https://github.com/Raman1121/FairTune\n", "date": "2023-10-08", "categories": []}
{"id": "2401.08691", "title": "Towards Responsible AI in Banking: Addressing Bias for Fair\n  Decision-Making", "abstract": "  In an era characterized by the pervasive integration of artificial\nintelligence into decision-making processes across diverse industries, the\ndemand for trust has never been more pronounced. This thesis embarks on a\ncomprehensive exploration of bias and fairness, with a particular emphasis on\ntheir ramifications within the banking sector, where AI-driven decisions bear\nsubstantial societal consequences. In this context, the seamless integration of\nfairness, explainability, and human oversight is of utmost importance,\nculminating in the establishment of what is commonly referred to as\n\"Responsible AI\". This emphasizes the critical nature of addressing biases\nwithin the development of a corporate culture that aligns seamlessly with both\nAI regulations and universal human rights standards, particularly in the realm\nof automated decision-making systems. Nowadays, embedding ethical principles\ninto the development, training, and deployment of AI models is crucial for\ncompliance with forthcoming European regulations and for promoting societal\ngood. This thesis is structured around three fundamental pillars: understanding\nbias, mitigating bias, and accounting for bias. These contributions are\nvalidated through their practical application in real-world scenarios, in\ncollaboration with Intesa Sanpaolo. This collaborative effort not only\ncontributes to our understanding of fairness but also provides practical tools\nfor the responsible implementation of AI-based decision-making systems. In line\nwith open-source principles, we have released Bias On Demand and FairView as\naccessible Python packages, further promoting progress in the field of AI\nfairness.\n", "date": "2024-01-13", "categories": []}
{"id": "2211.13709", "title": "Undesirable Biases in NLP: Addressing Challenges of Measurement", "abstract": "  As Large Language Models and Natural Language Processing (NLP) technology\nrapidly develop and spread into daily life, it becomes crucial to anticipate\nhow their use could harm people. One problem that has received a lot of\nattention in recent years is that this technology has displayed harmful biases,\nfrom generating derogatory stereotypes to producing disparate outcomes for\ndifferent social groups. Although a lot of effort has been invested in\nassessing and mitigating these biases, our methods of measuring the biases of\nNLP models have serious problems and it is often unclear what they actually\nmeasure. In this paper, we provide an interdisciplinary approach to discussing\nthe issue of NLP model bias by adopting the lens of psychometrics -- a field\nspecialized in the measurement of concepts like bias that are not directly\nobservable. In particular, we will explore two central notions from\npsychometrics, the construct validity and the reliability of measurement tools,\nand discuss how they can be applied in the context of measuring model bias. Our\ngoal is to provide NLP practitioners with methodological tools for designing\nbetter bias measures, and to inspire them more generally to explore tools from\npsychometrics when working on bias measurement tools.\n", "date": "2022-11-24", "categories": []}
{"id": "2306.10155", "title": "Fairness in Multi-Task Learning via Wasserstein Barycenters", "abstract": "  Algorithmic Fairness is an established field in machine learning that aims to\nreduce biases in data. Recent advances have proposed various methods to ensure\nfairness in a univariate environment, where the goal is to de-bias a single\ntask. However, extending fairness to a multi-task setting, where more than one\nobjective is optimised using a shared representation, remains underexplored. To\nbridge this gap, we develop a method that extends the definition of Strong\nDemographic Parity to multi-task learning using multi-marginal Wasserstein\nbarycenters. Our approach provides a closed form solution for the optimal fair\nmulti-task predictor including both regression and binary classification tasks.\nWe develop a data-driven estimation procedure for the solution and run\nnumerical experiments on both synthetic and real datasets. The empirical\nresults highlight the practical value of our post-processing methodology in\npromoting fair decision-making.\n", "date": "2023-06-16", "categories": []}
{"id": "2401.07174", "title": "On the (In)Compatibility between Group Fairness and Individual Fairness", "abstract": "  We study the compatibility between the optimal statistical parity solutions\nand individual fairness. While individual fairness seeks to treat similar\nindividuals similarly, optimal statistical parity aims to provide similar\ntreatment to individuals who share relative similarity within their respective\nsensitive groups. The two fairness perspectives, while both desirable from a\nfairness perspective, often come into conflict in applications. Our goal in\nthis work is to analyze the existence of this conflict and its potential\nsolution. In particular, we establish sufficient (sharp) conditions for the\ncompatibility between the optimal (post-processing) statistical parity $L^2$\nlearning and the ($K$-Lipschitz or $(\\epsilon,\\delta)$) individual fairness\nrequirements. Furthermore, when there exists a conflict between the two, we\nfirst relax the former to the Pareto frontier (or equivalently the optimal\ntrade-off) between $L^2$ error and statistical disparity, and then analyze the\ncompatibility between the frontier and the individual fairness requirements.\nOur analysis identifies regions along the Pareto frontier that satisfy\nindividual fairness requirements. (Lastly, we provide individual fairness\nguarantees for the composition of a trained model and the optimal\npost-processing step so that one can determine the compatibility of the\npost-processed model.) This provides practitioners with a valuable approach to\nattain Pareto optimality for statistical parity while adhering to the\nconstraints of individual fairness.\n", "date": "2024-01-13", "categories": []}
{"id": "2401.07697", "title": "Data vs. Model Machine Learning Fairness Testing: An Empirical Study", "abstract": "  Although several fairness definitions and bias mitigation techniques exist in\nthe literature, all existing solutions evaluate fairness of Machine Learning\n(ML) systems after the training stage. In this paper, we take the first steps\ntowards evaluating a more holistic approach by testing for fairness both before\nand after model training. We evaluate the effectiveness of the proposed\napproach and position it within the ML development lifecycle, using an\nempirical analysis of the relationship between model dependent and independent\nfairness metrics. The study uses 2 fairness metrics, 4 ML algorithms, 5\nreal-world datasets and 1600 fairness evaluation cycles. We find a linear\nrelationship between data and model fairness metrics when the distribution and\nthe size of the training data changes. Our results indicate that testing for\nfairness prior to training can be a ``cheap'' and effective means of catching a\nbiased data collection process early; detecting data drifts in production\nsystems and minimising execution of full training cycles thus reducing\ndevelopment time and costs.\n", "date": "2024-01-15", "categories": []}
{"id": "2401.08053", "title": "SCoFT: Self-Contrastive Fine-Tuning for Equitable Image Generation", "abstract": "  Accurate representation in media is known to improve the well-being of the\npeople who consume it. Generative image models trained on large web-crawled\ndatasets such as LAION are known to produce images with harmful stereotypes and\nmisrepresentations of cultures. We improve inclusive representation in\ngenerated images by (1) engaging with communities to collect a culturally\nrepresentative dataset that we call the Cross-Cultural Understanding Benchmark\n(CCUB) and (2) proposing a novel Self-Contrastive Fine-Tuning (SCoFT) method\nthat leverages the model's known biases to self-improve. SCoFT is designed to\nprevent overfitting on small datasets, encode only high-level information from\nthe data, and shift the generated distribution away from misrepresentations\nencoded in a pretrained model. Our user study conducted on 51 participants from\n5 different countries based on their self-selected national cultural\naffiliation shows that fine-tuning on CCUB consistently generates images with\nhigher cultural relevance and fewer stereotypes when compared to the Stable\nDiffusion baseline, which is further improved with our SCoFT technique.\n", "date": "2024-01-16", "categories": []}
{"id": "2309.04027", "title": "TIDE: Textual Identity Detection for Evaluating and Augmenting\n  Classification and Language Models", "abstract": "  Machine learning models can perpetuate unintended biases from unfair and\nimbalanced datasets. Evaluating and debiasing these datasets and models is\nespecially hard in text datasets where sensitive attributes such as race,\ngender, and sexual orientation may not be available. When these models are\ndeployed into society, they can lead to unfair outcomes for historically\nunderrepresented groups. In this paper, we present a dataset coupled with an\napproach to improve text fairness in classifiers and language models. We create\na new, more comprehensive identity lexicon, TIDAL, which includes 15,123\nidentity terms and associated sense context across three demographic\ncategories. We leverage TIDAL to develop an identity annotation and\naugmentation tool that can be used to improve the availability of identity\ncontext and the effectiveness of ML fairness techniques. We evaluate our\napproaches using human contributors, and additionally run experiments focused\non dataset and model debiasing. Results show our assistive annotation technique\nimproves the reliability and velocity of human-in-the-loop processes. Our\ndataset and methods uncover more disparities during evaluation, and also\nproduce more fair models during remediation. These approaches provide a\npractical path forward for scaling classifier and generative model fairness in\nreal-world settings.\n", "date": "2023-09-07", "categories": []}
{"id": "2401.06318", "title": "Striking a Balance in Fairness for Dynamic Systems Through Reinforcement\n  Learning", "abstract": "  While significant advancements have been made in the field of fair machine\nlearning, the majority of studies focus on scenarios where the decision model\noperates on a static population. In this paper, we study fairness in dynamic\nsystems where sequential decisions are made. Each decision may shift the\nunderlying distribution of features or user behavior. We model the dynamic\nsystem through a Markov Decision Process (MDP). By acknowledging that\ntraditional fairness notions and long-term fairness are distinct requirements\nthat may not necessarily align with one another, we propose an algorithmic\nframework to integrate various fairness considerations with reinforcement\nlearning using both pre-processing and in-processing approaches. Three case\nstudies show that our method can strike a balance between traditional fairness\nnotions, long-term fairness, and utility.\n", "date": "2024-01-12", "categories": []}
{"id": "2302.03350", "title": "To Be Forgotten or To Be Fair: Unveiling Fairness Implications of\n  Machine Unlearning Methods", "abstract": "  The right to be forgotten (RTBF) is motivated by the desire of people not to\nbe perpetually disadvantaged by their past deeds. For this, data deletion needs\nto be deep and permanent, and should be removed from machine learning models.\nResearchers have proposed machine unlearning algorithms which aim to erase\nspecific data from trained models more efficiently. However, these methods\nmodify how data is fed into the model and how training is done, which may\nsubsequently compromise AI ethics from the fairness perspective. To help\nsoftware engineers make responsible decisions when adopting these unlearning\nmethods, we present the first study on machine unlearning methods to reveal\ntheir fairness implications. We designed and conducted experiments on two\ntypical machine unlearning methods (SISA and AmnesiacML) along with a\nretraining method (ORTR) as baseline using three fairness datasets under three\ndifferent deletion strategies. Experimental results show that under non-uniform\ndata deletion, SISA leads to better fairness compared with ORTR and AmnesiacML,\nwhile initial training and uniform data deletion do not necessarily affect the\nfairness of all three methods. These findings have exposed an important\nresearch problem in software engineering, and can help practitioners better\nunderstand the potential trade-offs on fairness when considering solutions for\nRTBF.\n", "date": "2023-02-07", "categories": []}
{"id": "2308.02053", "title": "The Unequal Opportunities of Large Language Models: Revealing\n  Demographic Bias through Job Recommendations", "abstract": "  Large Language Models (LLMs) have seen widespread deployment in various\nreal-world applications. Understanding these biases is crucial to comprehend\nthe potential downstream consequences when using LLMs to make decisions,\nparticularly for historically disadvantaged groups. In this work, we propose a\nsimple method for analyzing and comparing demographic bias in LLMs, through the\nlens of job recommendations. We demonstrate the effectiveness of our method by\nmeasuring intersectional biases within ChatGPT and LLaMA, two cutting-edge\nLLMs. Our experiments primarily focus on uncovering gender identity and\nnationality bias; however, our method can be extended to examine biases\nassociated with any intersection of demographic identities. We identify\ndistinct biases in both models toward various demographic identities, such as\nboth models consistently suggesting low-paying jobs for Mexican workers or\npreferring to recommend secretarial roles to women. Our study highlights the\nimportance of measuring the bias of LLMs in downstream applications to\nunderstand the potential for harm and inequitable outcomes.\n", "date": "2023-08-03", "categories": []}
{"id": "2311.06826", "title": "Fairness Hacking: The Malicious Practice of Shrouding Unfairness in\n  Algorithms", "abstract": "  Fairness in machine learning (ML) is an ever-growing field of research due to\nthe manifold potential for harm from algorithmic discrimination. To prevent\nsuch harm, a large body of literature develops new approaches to quantify\nfairness. Here, we investigate how one can divert the quantification of\nfairness by describing a practice we call \"fairness hacking\" for the purpose of\nshrouding unfairness in algorithms. This impacts end-users who rely on learning\nalgorithms, as well as the broader community interested in fair AI practices.\nWe introduce two different categories of fairness hacking in reference to the\nestablished concept of p-hacking. The first category, intra-metric fairness\nhacking, describes the misuse of a particular metric by adding or removing\nsensitive attributes from the analysis. In this context, countermeasures that\nhave been developed to prevent or reduce p-hacking can be applied to similarly\nprevent or reduce fairness hacking. The second category of fairness hacking is\ninter-metric fairness hacking. Inter-metric fairness hacking is the search for\na specific fair metric with given attributes. We argue that countermeasures to\nprevent or reduce inter-metric fairness hacking are still in their infancy.\nFinally, we demonstrate both types of fairness hacking using real datasets. Our\npaper intends to serve as a guidance for discussions within the fair ML\ncommunity to prevent or reduce the misuse of fairness metrics, and thus reduce\noverall harm from ML applications.\n", "date": "2023-11-12", "categories": []}
{"id": "2211.11183", "title": "Causal Fairness Assessment of Treatment Allocation with Electronic\n  Health Records", "abstract": "  Healthcare continues to grapple with the persistent issue of treatment\ndisparities, sparking concerns regarding the equitable allocation of treatments\nin clinical practice. While various fairness metrics have emerged to assess\nfairness in decision-making processes, a growing focus has been on\ncausality-based fairness concepts due to their capacity to mitigate confounding\neffects and reason about bias. However, the application of causal fairness\nnotions in evaluating the fairness of clinical decision-making with electronic\nhealth record (EHR) data remains an understudied domain. This study aims to\naddress the methodological gap in assessing causal fairness of treatment\nallocation with electronic health records data. We propose a causal fairness\nalgorithm to assess fairness in clinical decision-making. Our algorithm\naccounts for the heterogeneity of patient populations and identifies potential\nunfairness in treatment allocation by conditioning on patients who have the\nsame likelihood to benefit from the treatment. We apply this framework to a\npatient cohort with coronary artery disease derived from an EHR database to\nevaluate the fairness of treatment decisions. In addition, we investigate the\nimpact of social determinants of health on the assessment of causal fairness of\ntreatment allocation.\n", "date": "2022-11-21", "categories": []}
{"id": "2310.07535", "title": "Fairness under Covariate Shift: Improving Fairness-Accuracy tradeoff\n  with few Unlabeled Test Samples", "abstract": "  Covariate shift in the test data is a common practical phenomena that can\nsignificantly downgrade both the accuracy and the fairness performance of the\nmodel. Ensuring fairness across different sensitive groups under covariate\nshift is of paramount importance due to societal implications like criminal\njustice. We operate in the unsupervised regime where only a small set of\nunlabeled test samples along with a labeled training set is available. Towards\nimproving fairness under this highly challenging yet realistic scenario, we\nmake three contributions. First is a novel composite weighted entropy based\nobjective for prediction accuracy which is optimized along with a\nrepresentation matching loss for fairness. We experimentally verify that\noptimizing with our loss formulation outperforms a number of state-of-the-art\nbaselines in the pareto sense with respect to the fairness-accuracy tradeoff on\nseveral standard datasets. Our second contribution is a new setting we term\nAsymmetric Covariate Shift that, to the best of our knowledge, has not been\nstudied before. Asymmetric covariate shift occurs when distribution of\ncovariates of one group shifts significantly compared to the other groups and\nthis happens when a dominant group is over-represented. While this setting is\nextremely challenging for current baselines, We show that our proposed method\nsignificantly outperforms them. Our third contribution is theoretical, where we\nshow that our weighted entropy term along with prediction loss on the training\nset approximates test loss under covariate shift. Empirically and through\nformal sample complexity bounds, we show that this approximation to the unseen\ntest loss does not depend on importance sampling variance which affects many\nother baselines.\n", "date": "2023-10-11", "categories": []}
{"id": "2401.03097", "title": "Adaptive Boosting with Fairness-aware Reweighting Technique for Fair\n  Classification", "abstract": "  Machine learning methods based on AdaBoost have been widely applied to\nvarious classification problems across many mission-critical applications\nincluding healthcare, law and finance. However, there is a growing concern\nabout the unfairness and discrimination of data-driven classification models,\nwhich is inevitable for classical algorithms including AdaBoost. In order to\nachieve fair classification, a novel fair AdaBoost (FAB) approach is proposed\nthat is an interpretable fairness-improving variant of AdaBoost. We mainly\ninvestigate binary classification problems and focus on the fairness of three\ndifferent indicators (i.e., accuracy, false positive rate and false negative\nrate). By utilizing a fairness-aware reweighting technique for base\nclassifiers, the proposed FAB approach can achieve fair classification while\nmaintaining the advantage of AdaBoost with negligible sacrifice of predictive\nperformance. In addition, a hyperparameter is introduced in FAB to show\npreferences for the fairness-accuracy trade-off. An upper bound for the target\nloss function that quantifies error rate and unfairness is theoretically\nderived for FAB, which provides a strict theoretical support for the\nfairness-improving methods designed for AdaBoost. The effectiveness of the\nproposed method is demonstrated on three real-world datasets (i.e., Adult,\nCOMPAS and HSLS) with respect to the three fairness indicators. The results are\naccordant with theoretic analyses, and show that (i) FAB significantly improves\nclassification fairness at a small cost of accuracy compared with AdaBoost; and\n(ii) FAB outperforms state-of-the-art fair classification methods including\nequalized odds method, exponentiated gradient method, and disparate\nmistreatment method in terms of the fairness-accuracy trade-off.\n", "date": "2024-01-06", "categories": []}
{"id": "2401.02552", "title": "Long-term Fairness For Real-time Decision Making: A Constrained Online\n  Optimization Approach", "abstract": "  Machine learning (ML) has demonstrated remarkable capabilities across many\nreal-world systems, from predictive modeling to intelligent automation.\nHowever, the widespread integration of machine learning also makes it necessary\nto ensure machine learning-driven decision-making systems do not violate\nethical principles and values of society in which they operate. As ML-driven\ndecisions proliferate, particularly in cases involving sensitive attributes\nsuch as gender, race, and age, to name a few, the need for equity and\nimpartiality has emerged as a fundamental concern. In situations demanding\nreal-time decision-making, fairness objectives become more nuanced and complex:\ninstantaneous fairness to ensure equity in every time slot, and long-term\nfairness to ensure fairness over a period of time. There is a growing awareness\nthat real-world systems that operate over long periods and require fairness\nover different timelines. However, existing approaches mainly address dynamic\ncosts with time-invariant fairness constraints, often disregarding the\nchallenges posed by time-varying fairness constraints. To bridge this gap, this\nwork introduces a framework for ensuring long-term fairness within dynamic\ndecision-making systems characterized by time-varying fairness constraints. We\nformulate the decision problem with fairness constraints over a period as a\nconstrained online optimization problem. A novel online algorithm, named\nLoTFair, is presented that solves the problem 'on the fly'. We prove that\nLoTFair can make overall fairness violations negligible while maintaining the\nperformance over the long run.\n", "date": "2024-01-04", "categories": []}
{"id": "2305.06055", "title": "A Classification of Feedback Loops and Their Relation to Biases in\n  Automated Decision-Making Systems", "abstract": "  Prediction-based decision-making systems are becoming increasingly prevalent\nin various domains. Previous studies have demonstrated that such systems are\nvulnerable to runaway feedback loops, e.g., when police are repeatedly sent\nback to the same neighborhoods regardless of the actual rate of criminal\nactivity, which exacerbate existing biases. In practice, the automated\ndecisions have dynamic feedback effects on the system itself that can\nperpetuate over time, making it difficult for short-sighted design choices to\ncontrol the system's evolution. While researchers started proposing longer-term\nsolutions to prevent adverse outcomes (such as bias towards certain groups),\nthese interventions largely depend on ad hoc modeling assumptions and a\nrigorous theoretical understanding of the feedback dynamics in ML-based\ndecision-making systems is currently missing. In this paper, we use the\nlanguage of dynamical systems theory, a branch of applied mathematics that\ndeals with the analysis of the interconnection of systems with dynamic\nbehaviors, to rigorously classify the different types of feedback loops in the\nML-based decision-making pipeline. By reviewing existing scholarly work, we\nshow that this classification covers many examples discussed in the algorithmic\nfairness community, thereby providing a unifying and principled framework to\nstudy feedback loops. By qualitative analysis, and through a simulation example\nof recommender systems, we show which specific types of ML biases are affected\nby each type of feedback loop. We find that the existence of feedback loops in\nthe ML-based decision-making pipeline can perpetuate, reinforce, or even reduce\nML biases.\n", "date": "2023-05-10", "categories": []}
{"id": "2401.02012", "title": "Fast & Fair: Efficient Second-Order Robust Optimization for Fairness in\n  Machine Learning", "abstract": "  This project explores adversarial training techniques to develop fairer Deep\nNeural Networks (DNNs) to mitigate the inherent bias they are known to exhibit.\nDNNs are susceptible to inheriting bias with respect to sensitive attributes\nsuch as race and gender, which can lead to life-altering outcomes (e.g.,\ndemographic bias in facial recognition software used to arrest a suspect). We\npropose a robust optimization problem, which we demonstrate can improve\nfairness in several datasets, both synthetic and real-world, using an affine\nlinear model. Leveraging second order information, we are able to find a\nsolution to our optimization problem more efficiently than a purely first order\nmethod.\n", "date": "2024-01-04", "categories": []}
{"id": "2401.02183", "title": "FairGridSearch: A Framework to Compare Fairness-Enhancing Models", "abstract": "  Machine learning models are increasingly used in critical decision-making\napplications. However, these models are susceptible to replicating or even\namplifying bias present in real-world data. While there are various bias\nmitigation methods and base estimators in the literature, selecting the optimal\nmodel for a specific application remains challenging.\n  This paper focuses on binary classification and proposes FairGridSearch, a\nnovel framework for comparing fairness-enhancing models. FairGridSearch enables\nexperimentation with different model parameter combinations and recommends the\nbest one. The study applies FairGridSearch to three popular datasets (Adult,\nCOMPAS, and German Credit) and analyzes the impacts of metric selection, base\nestimator choice, and classification threshold on model fairness.\n  The results highlight the significance of selecting appropriate accuracy and\nfairness metrics for model evaluation. Additionally, different base estimators\nand classification threshold values affect the effectiveness of bias mitigation\nmethods and fairness stability respectively, but the effects are not consistent\nacross all datasets. Based on these findings, future research on fairness in\nmachine learning should consider a broader range of factors when building fair\nmodels, going beyond bias mitigation methods alone.\n", "date": "2024-01-04", "categories": []}
{"id": "2305.09073", "title": "Consensus and Subjectivity of Skin Tone Annotation for ML Fairness", "abstract": "  Understanding different human attributes and how they affect model behavior\nmay become a standard need for all model creation and usage, from traditional\ncomputer vision tasks to the newest multimodal generative AI systems. In\ncomputer vision specifically, we have relied on datasets augmented with\nperceived attribute signals (e.g., gender presentation, skin tone, and age) and\nbenchmarks enabled by these datasets. Typically labels for these tasks come\nfrom human annotators. However, annotating attribute signals, especially skin\ntone, is a difficult and subjective task. Perceived skin tone is affected by\ntechnical factors, like lighting conditions, and social factors that shape an\nannotator's lived experience. This paper examines the subjectivity of skin tone\nannotation through a series of annotation experiments using the Monk Skin Tone\n(MST) scale, a small pool of professional photographers, and a much larger pool\nof trained crowdsourced annotators. Along with this study we release the Monk\nSkin Tone Examples (MST-E) dataset, containing 1515 images and 31 videos spread\nacross the full MST scale. MST-E is designed to help train human annotators to\nannotate MST effectively. Our study shows that annotators can reliably annotate\nskin tone in a way that aligns with an expert in the MST scale, even under\nchallenging environmental conditions. We also find evidence that annotators\nfrom different geographic regions rely on different mental models of MST\ncategories resulting in annotations that systematically vary across regions.\nGiven this, we advise practitioners to use a diverse set of annotators and a\nhigher replication count for each image when annotating skin tone for fairness\nresearch.\n", "date": "2023-05-16", "categories": []}
{"id": "2305.11348", "title": "In the Name of Fairness: Assessing the Bias in Clinical Record\n  De-identification", "abstract": "  Data sharing is crucial for open science and reproducible research, but the\nlegal sharing of clinical data requires the removal of protected health\ninformation from electronic health records. This process, known as\nde-identification, is often achieved through the use of machine learning\nalgorithms by many commercial and open-source systems. While these systems have\nshown compelling results on average, the variation in their performance across\ndifferent demographic groups has not been thoroughly examined. In this work, we\ninvestigate the bias of de-identification systems on names in clinical notes\nvia a large-scale empirical analysis. To achieve this, we create 16 name sets\nthat vary along four demographic dimensions: gender, race, name popularity, and\nthe decade of popularity. We insert these names into 100 manually curated\nclinical templates and evaluate the performance of nine public and private\nde-identification methods. Our findings reveal that there are statistically\nsignificant performance gaps along a majority of the demographic dimensions in\nmost methods. We further illustrate that de-identification quality is affected\nby polysemy in names, gender context, and clinical note characteristics. To\nmitigate the identified gaps, we propose a simple and method-agnostic solution\nby fine-tuning de-identification methods with clinical context and diverse\nnames. Overall, it is imperative to address the bias in existing methods\nimmediately so that downstream stakeholders can build high-quality systems to\nserve all demographic parties fairly.\n", "date": "2023-05-18", "categories": []}
{"id": "2401.01262", "title": "Fairness Certification for Natural Language Processing and Large\n  Language Models", "abstract": "  Natural Language Processing (NLP) plays an important role in our daily lives,\nparticularly due to the enormous progress of Large Language Models (LLM).\nHowever, NLP has many fairness-critical use cases, e.g., as an expert system in\nrecruitment or as an LLM-based tutor in education. Since NLP is based on human\nlanguage, potentially harmful biases can diffuse into NLP systems and produce\nunfair results, discriminate against minorities or generate legal issues.\nHence, it is important to develop a fairness certification for NLP approaches.\nWe follow a qualitative research approach towards a fairness certification for\nNLP. In particular, we have reviewed a large body of literature on algorithmic\nfairness, and we have conducted semi-structured expert interviews with a wide\nrange of experts from that area. We have systematically devised six fairness\ncriteria for NLP, which can be further refined into 18 sub-categories. Our\ncriteria offer a foundation for operationalizing and testing processes to\ncertify fairness, both from the perspective of the auditor and the audited\norganization.\n", "date": "2024-01-02", "categories": []}
{"id": "2401.01640", "title": "Evaluating Fairness in Self-supervised and Supervised Models for\n  Sequential Data", "abstract": "  Self-supervised learning (SSL) has become the de facto training paradigm of\nlarge models where pre-training is followed by supervised fine-tuning using\ndomain-specific data and labels. Hypothesizing that SSL models would learn more\ngeneric, hence less biased, representations, this study explores the impact of\npre-training and fine-tuning strategies on fairness (i.e., performing equally\non different demographic breakdowns). Motivated by human-centric applications\non real-world timeseries data, we interpret inductive biases on the model,\nlayer, and metric levels by systematically comparing SSL models to their\nsupervised counterparts. Our findings demonstrate that SSL has the capacity to\nachieve performance on par with supervised methods while significantly\nenhancing fairness--exhibiting up to a 27% increase in fairness with a mere 1%\nloss in performance through self-supervision. Ultimately, this work underscores\nSSL's potential in human-centric computing, particularly high-stakes,\ndata-scarce application domains like healthcare.\n", "date": "2024-01-03", "categories": []}
{"id": "2312.17443", "title": "Break Out of a Pigeonhole: A Unified Framework for Examining\n  Miscalibration, Bias, and Stereotype in Recommender Systems", "abstract": "  Despite the benefits of personalizing items and information tailored to\nusers' needs, it has been found that recommender systems tend to introduce\nbiases that favor popular items or certain categories of items, and dominant\nuser groups. In this study, we aim to characterize the systematic errors of a\nrecommendation system and how they manifest in various accountability issues,\nsuch as stereotypes, biases, and miscalibration. We propose a unified framework\nthat distinguishes the sources of prediction errors into a set of key measures\nthat quantify the various types of system-induced effects, both at the\nindividual and collective levels. Based on our measuring framework, we examine\nthe most widely adopted algorithms in the context of movie recommendation. Our\nresearch reveals three important findings: (1) Differences between algorithms:\nrecommendations generated by simpler algorithms tend to be more stereotypical\nbut less biased than those generated by more complex algorithms. (2) Disparate\nimpact on groups and individuals: system-induced biases and stereotypes have a\ndisproportionate effect on atypical users and minority groups (e.g., women and\nolder users). (3) Mitigation opportunity: using structural equation modeling,\nwe identify the interactions between user characteristics (typicality and\ndiversity), system-induced effects, and miscalibration. We further investigate\nthe possibility of mitigating system-induced effects by oversampling\nunderrepresented groups and individuals, which was found to be effective in\nreducing stereotypes and improving recommendation quality. Our research is the\nfirst systematic examination of not only system-induced effects and\nmiscalibration but also the stereotyping issue in recommender systems.\n", "date": "2023-12-29", "categories": []}
{"id": "2312.16726", "title": "FairCompass: Operationalising Fairness in Machine Learning", "abstract": "  As artificial intelligence (AI) increasingly becomes an integral part of our\nsocietal and individual activities, there is a growing imperative to develop\nresponsible AI solutions. Despite a diverse assortment of machine learning\nfairness solutions is proposed in the literature, there is reportedly a lack of\npractical implementation of these tools in real-world applications. Industry\nexperts have participated in thorough discussions on the challenges associated\nwith operationalising fairness in the development of machine learning-empowered\nsolutions, in which a shift toward human-centred approaches is promptly\nadvocated to mitigate the limitations of existing techniques. In this work, we\npropose a human-in-the-loop approach for fairness auditing, presenting a mixed\nvisual analytical system (hereafter referred to as 'FairCompass'), which\nintegrates both subgroup discovery technique and the decision tree-based schema\nfor end users. Moreover, we innovatively integrate an Exploration, Guidance and\nInformed Analysis loop, to facilitate the use of the Knowledge Generation Model\nfor Visual Analytics in FairCompass. We evaluate the effectiveness of\nFairCompass for fairness auditing in a real-world scenario, and the findings\ndemonstrate the system's potential for real-world deployability. We anticipate\nthis work will address the current gaps in research for fairness and facilitate\nthe operationalisation of fairness in machine learning systems.\n", "date": "2023-12-27", "categories": []}
{"id": "2312.08749", "title": "Mitigating Label Bias in Machine Learning: Fairness through Confident\n  Learning", "abstract": "  Discrimination can occur when the underlying unbiased labels are overwritten\nby an agent with potential bias, resulting in biased datasets that unfairly\nharm specific groups and cause classifiers to inherit these biases. In this\npaper, we demonstrate that despite only having access to the biased labels, it\nis possible to eliminate bias by filtering the fairest instances within the\nframework of confident learning. In the context of confident learning, low\nself-confidence usually indicates potential label errors; however, this is not\nalways the case. Instances, particularly those from underrepresented groups,\nmight exhibit low confidence scores for reasons other than labeling errors. To\naddress this limitation, our approach employs truncation of the confidence\nscore and extends the confidence interval of the probabilistic threshold.\nAdditionally, we incorporate with co-teaching paradigm for providing a more\nrobust and reliable selection of fair instances and effectively mitigating the\nadverse effects of biased labels. Through extensive experimentation and\nevaluation of various datasets, we demonstrate the efficacy of our approach in\npromoting fairness and reducing the impact of label bias in machine learning\nmodels.\n", "date": "2023-12-14", "categories": []}
{"id": "2312.14976", "title": "Gaussian Harmony: Attaining Fairness in Diffusion-based Face Generation\n  Models", "abstract": "  Diffusion models have achieved great progress in face generation. However,\nthese models amplify the bias in the generation process, leading to an\nimbalance in distribution of sensitive attributes such as age, gender and race.\nThis paper proposes a novel solution to this problem by balancing the facial\nattributes of the generated images. We mitigate the bias by localizing the\nmeans of the facial attributes in the latent space of the diffusion model using\nGaussian mixture models (GMM). Our motivation for choosing GMMs over other\nclustering frameworks comes from the flexible latent structure of diffusion\nmodel. Since each sampling step in diffusion models follows a Gaussian\ndistribution, we show that fitting a GMM model helps us to localize the\nsubspace responsible for generating a specific attribute. Furthermore, our\nmethod does not require retraining, we instead localize the subspace on-the-fly\nand mitigate the bias for generating a fair dataset. We evaluate our approach\non multiple face attribute datasets to demonstrate the effectiveness of our\napproach. Our results demonstrate that our approach leads to a more fair data\ngeneration in terms of representational fairness while preserving the quality\nof generated samples.\n", "date": "2023-12-21", "categories": []}
{"id": "2312.15994", "title": "Practical Bias Mitigation through Proxy Sensitive Attribute Label\n  Generation", "abstract": "  Addressing bias in the trained machine learning system often requires access\nto sensitive attributes. In practice, these attributes are not available either\ndue to legal and policy regulations or data unavailability for a given\ndemographic. Existing bias mitigation algorithms are limited in their\napplicability to real-world scenarios as they require access to sensitive\nattributes to achieve fairness. In this research work, we aim to address this\nbottleneck through our proposed unsupervised proxy-sensitive attribute label\ngeneration technique. Towards this end, we propose a two-stage approach of\nunsupervised embedding generation followed by clustering to obtain\nproxy-sensitive labels. The efficacy of our work relies on the assumption that\nbias propagates through non-sensitive attributes that are correlated to the\nsensitive attributes and, when mapped to the high dimensional latent space,\nproduces clusters of different demographic groups that exist in the data.\nExperimental results demonstrate that bias mitigation using existing algorithms\nsuch as Fair Mixup and Adversarial Debiasing yields comparable results on\nderived proxy labels when compared against using true sensitive attributes.\n", "date": "2023-12-26", "categories": []}
{"id": "2306.05059", "title": "Reconciling Predictive and Statistical Parity: A Causal Approach", "abstract": "  Since the rise of fair machine learning as a critical field of inquiry, many\ndifferent notions on how to quantify and measure discrimination have been\nproposed in the literature. Some of these notions, however, were shown to be\nmutually incompatible. Such findings make it appear that numerous different\nkinds of fairness exist, thereby making a consensus on the appropriate measure\nof fairness harder to reach, hindering the applications of these tools in\npractice. In this paper, we investigate one of these key impossibility results\nthat relates the notions of statistical and predictive parity. Specifically, we\nderive a new causal decomposition formula for the fairness measures associated\nwith predictive parity, and obtain a novel insight into how this criterion is\nrelated to statistical parity through the legal doctrines of disparate\ntreatment, disparate impact, and the notion of business necessity. Our results\nshow that through a more careful causal analysis, the notions of statistical\nand predictive parity are not really mutually exclusive, but complementary and\nspanning a spectrum of fairness notions through the concept of business\nnecessity. Finally, we demonstrate the importance of our findings on a\nreal-world example.\n", "date": "2023-06-08", "categories": []}
{"id": "2312.14299", "title": "Fairness in Submodular Maximization over a Matroid Constraint", "abstract": "  Submodular maximization over a matroid constraint is a fundamental problem\nwith various applications in machine learning. Some of these applications\ninvolve decision-making over datapoints with sensitive attributes such as\ngender or race. In such settings, it is crucial to guarantee that the selected\nsolution is fairly distributed with respect to this attribute. Recently,\nfairness has been investigated in submodular maximization under a cardinality\nconstraint in both the streaming and offline settings, however the more general\nproblem with matroid constraint has only been considered in the streaming\nsetting and only for monotone objectives. This work fills this gap. We propose\nvarious algorithms and impossibility results offering different trade-offs\nbetween quality, fairness, and generality.\n", "date": "2023-12-21", "categories": []}
{"id": "2303.00586", "title": "FAIR-Ensemble: When Fairness Naturally Emerges From Deep Ensembling", "abstract": "  Ensembling multiple Deep Neural Networks (DNNs) is a simple and effective way\nto improve top-line metrics and to outperform a larger single model. In this\nwork, we go beyond top-line metrics and instead explore the impact of\nensembling on subgroup performances. Surprisingly, we observe that even with a\nsimple homogeneous ensemble -- all the individual DNNs share the same training\nset, architecture, and design choices -- the minority group performance\ndisproportionately improves with the number of models compared to the majority\ngroup, i.e. fairness naturally emerges from ensembling. Even more surprising,\nwe find that this gain keeps occurring even when a large number of models is\nconsidered, e.g. $20$, despite the fact that the average performance of the\nensemble plateaus with fewer models. Our work establishes that simple DNN\nensembles can be a powerful tool for alleviating disparate impact from DNN\nclassifiers, thus curbing algorithmic harm. We also explore why this is the\ncase. We find that even in homogeneous ensembles, varying the sources of\nstochasticity through parameter initialization, mini-batch sampling, and\ndata-augmentation realizations, results in different fairness outcomes.\n", "date": "2023-03-01", "categories": []}
{"id": "2304.03693", "title": "Model-Agnostic Gender Debiased Image Captioning", "abstract": "  Image captioning models are known to perpetuate and amplify harmful societal\nbias in the training set. In this work, we aim to mitigate such gender bias in\nimage captioning models. While prior work has addressed this problem by forcing\nmodels to focus on people to reduce gender misclassification, it conversely\ngenerates gender-stereotypical words at the expense of predicting the correct\ngender. From this observation, we hypothesize that there are two types of\ngender bias affecting image captioning models: 1) bias that exploits context to\npredict gender, and 2) bias in the probability of generating certain (often\nstereotypical) words because of gender. To mitigate both types of gender\nbiases, we propose a framework, called LIBRA, that learns from synthetically\nbiased samples to decrease both types of biases, correcting gender\nmisclassification and changing gender-stereotypical words to more neutral ones.\nCode is available at https://github.com/rebnej/LIBRA.\n", "date": "2023-04-07", "categories": []}
{"id": "2306.03625", "title": "Fair and Robust Estimation of Heterogeneous Treatment Effects for Policy\n  Learning", "abstract": "  We propose a simple and general framework for nonparametric estimation of\nheterogeneous treatment effects under fairness constraints. Under standard\nregularity conditions, we show that the resulting estimators possess the double\nrobustness property. We use this framework to characterize the trade-off\nbetween fairness and the maximum welfare achievable by the optimal policy. We\nevaluate the methods in a simulation study and illustrate them in a real-world\ncase study.\n", "date": "2023-06-06", "categories": []}
{"id": "2312.04810", "title": "RS-Corrector: Correcting the Racial Stereotypes in Latent Diffusion\n  Models", "abstract": "  Recent text-conditioned image generation models have demonstrated an\nexceptional capacity to produce diverse and creative imagery with high visual\nquality. However, when pre-trained on billion-sized datasets randomly collected\nfrom the Internet, where potential biased human preferences exist, these models\ntend to produce images with common and recurring stereotypes, particularly for\ncertain racial groups. In this paper, we conduct an initial analysis of the\npublicly available Stable Diffusion model and its derivatives, highlighting the\npresence of racial stereotypes. These models often generate distorted or biased\nimages for certain racial groups, emphasizing stereotypical characteristics. To\naddress these issues, we propose a framework called \"RS-Corrector\", designed to\nestablish an anti-stereotypical preference in the latent space and update the\nlatent code for refined generated results. The correction process occurs during\nthe inference stage without requiring fine-tuning of the original model.\nExtensive empirical evaluations demonstrate that the introduced \\themodel\neffectively corrects the racial stereotypes of the well-trained Stable\nDiffusion model while leaving the original model unchanged.\n", "date": "2023-12-08", "categories": []}
{"id": "2312.08288", "title": "Hybrid Sample Synthesis-based Debiasing of Classifier in Limited Data\n  Setting", "abstract": "  Deep learning models are known to suffer from the problem of bias, and\nresearchers have been exploring methods to address this issue. However, most of\nthese methods require prior knowledge of the bias and are not always practical.\nIn this paper, we focus on a more practical setting with no prior information\nabout the bias. Generally, in this setting, there are a large number of\nbias-aligned samples that cause the model to produce biased predictions and a\nfew bias-conflicting samples that do not conform to the bias. If the training\ndata is limited, the influence of the bias-aligned samples may become even\nstronger on the model predictions, and we experimentally demonstrate that\nexisting debiasing techniques suffer severely in such cases. In this paper, we\nexamine the effects of unknown bias in small dataset regimes and present a\nnovel approach to mitigate this issue. The proposed approach directly addresses\nthe issue of the extremely low occurrence of bias-conflicting samples in\nlimited data settings through the synthesis of hybrid samples that can be used\nto reduce the effect of bias. We perform extensive experiments on several\nbenchmark datasets and experimentally demonstrate the effectiveness of our\nproposed approach in addressing any unknown bias in the presence of limited\ndata. Specifically, our approach outperforms the vanilla, LfF, LDD, and DebiAN\ndebiasing methods by absolute margins of 10.39%, 9.08%, 8.07%, and 9.67% when\nonly 10% of the Corrupted CIFAR-10 Type 1 dataset is available with a\nbias-conflicting sample ratio of 0.05.\n", "date": "2023-12-13", "categories": []}
{"id": "2312.12560", "title": "Comprehensive Validation on Reweighting Samples for Bias Mitigation via\n  AIF360", "abstract": "  Fairness AI aims to detect and alleviate bias across the entire AI\ndevelopment life cycle, encompassing data curation, modeling, evaluation, and\ndeployment-a pivotal aspect of ethical AI implementation. Addressing data bias,\nparticularly concerning sensitive attributes like gender and race, reweighting\nsamples proves efficient for fairness AI. This paper contributes a systematic\nexamination of reweighting samples for traditional machine learning (ML)\nmodels, employing five models for binary classification on the Adult Income and\nCOMPUS datasets with various protected attributes. The study evaluates\nprediction results using five fairness metrics, uncovering the nuanced and\nmodel-specific nature of reweighting sample effectiveness in achieving fairness\nin traditional ML models, as well as revealing the complexity of bias dynamics.\n", "date": "2023-12-19", "categories": []}
{"id": "2312.13173", "title": "Learning Fair Policies for Multi-stage Selection Problems from\n  Observational Data", "abstract": "  We consider the problem of learning fair policies for multi-stage selection\nproblems from observational data. This problem arises in several high-stakes\ndomains such as company hiring, loan approval, or bail decisions where outcomes\n(e.g., career success, loan repayment, recidivism) are only observed for those\nselected. We propose a multi-stage framework that can be augmented with various\nfairness constraints, such as demographic parity or equal opportunity. This\nproblem is a highly intractable infinite chance-constrained program involving\nthe unknown joint distribution of covariates and outcomes. Motivated by the\npotential impact of selection decisions on people's lives and livelihoods, we\npropose to focus on interpretable linear selection rules. Leveraging tools from\ncausal inference and sample average approximation, we obtain an asymptotically\nconsistent solution to this selection problem by solving a mixed binary conic\noptimization problem, which can be solved using standard off-the-shelf solvers.\nWe conduct extensive computational experiments on a variety of datasets adapted\nfrom the UCI repository on which we show that our proposed approaches can\nachieve an 11.6% improvement in precision and a 38% reduction in the measure of\nunfairness compared to the existing selection policy.\n", "date": "2023-12-20", "categories": []}
{"id": "2312.11969", "title": "GroupMixNorm Layer for Learning Fair Models", "abstract": "  Recent research has identified discriminatory behavior of automated\nprediction algorithms towards groups identified on specific protected\nattributes (e.g., gender, ethnicity, age group, etc.). When deployed in\nreal-world scenarios, such techniques may demonstrate biased predictions\nresulting in unfair outcomes. Recent literature has witnessed algorithms for\nmitigating such biased behavior mostly by adding convex surrogates of fairness\nmetrics such as demographic parity or equalized odds in the loss function,\nwhich are often not easy to estimate. This research proposes a novel\nin-processing based GroupMixNorm layer for mitigating bias from deep learning\nmodels. The GroupMixNorm layer probabilistically mixes group-level feature\nstatistics of samples across different groups based on the protected attribute.\nThe proposed method improves upon several fairness metrics with minimal impact\non overall accuracy. Analysis on benchmark tabular and image datasets\ndemonstrates the efficacy of the proposed method in achieving state-of-the-art\nperformance. Further, the experimental analysis also suggests the robustness of\nthe GroupMixNorm layer against new protected attributes during inference and\nits utility in eliminating bias from a pre-trained network.\n", "date": "2023-12-19", "categories": []}
{"id": "2312.12369", "title": "Chasing Fairness in Graphs: A GNN Architecture Perspective", "abstract": "  There has been significant progress in improving the performance of graph\nneural networks (GNNs) through enhancements in graph data, model architecture\ndesign, and training strategies. For fairness in graphs, recent studies achieve\nfair representations and predictions through either graph data pre-processing\n(e.g., node feature masking, and topology rewiring) or fair training strategies\n(e.g., regularization, adversarial debiasing, and fair contrastive learning).\nHow to achieve fairness in graphs from the model architecture perspective is\nless explored. More importantly, GNNs exhibit worse fairness performance\ncompared to multilayer perception since their model architecture (i.e.,\nneighbor aggregation) amplifies biases. To this end, we aim to achieve fairness\nvia a new GNN architecture. We propose \\textsf{F}air \\textsf{M}essage\n\\textsf{P}assing (FMP) designed within a unified optimization framework for\nGNNs. Notably, FMP \\textit{explicitly} renders sensitive attribute usage in\n\\textit{forward propagation} for node classification task using cross-entropy\nloss without data pre-processing. In FMP, the aggregation is first adopted to\nutilize neighbors' information and then the bias mitigation step explicitly\npushes demographic group node presentation centers together. In this way, FMP\nscheme can aggregate useful information from neighbors and mitigate bias to\nachieve better fairness and prediction tradeoff performance. Experiments on\nnode classification tasks demonstrate that the proposed FMP outperforms several\nbaselines in terms of fairness and accuracy on three real-world datasets. The\ncode is available in {\\url{https://github.com/zhimengj0326/FMP}}.\n", "date": "2023-12-19", "categories": []}
{"id": "2208.09147", "title": "Disentangled Representation with Causal Constraints for Counterfactual\n  Fairness", "abstract": "  Much research has been devoted to the problem of learning fair\nrepresentations; however, they do not explicitly the relationship between\nlatent representations. In many real-world applications, there may be causal\nrelationships between latent representations. Furthermore, most fair\nrepresentation learning methods focus on group-level fairness and are based on\ncorrelations, ignoring the causal relationships underlying the data. In this\nwork, we theoretically demonstrate that using the structured representations\nenable downstream predictive models to achieve counterfactual fairness, and\nthen we propose the Counterfactual Fairness Variational AutoEncoder (CF-VAE) to\nobtain structured representations with respect to domain knowledge. The\nexperimental results show that the proposed method achieves better fairness and\naccuracy performance than the benchmark fairness methods.\n", "date": "2022-08-19", "categories": []}
{"id": "2303.17743", "title": "FairGen: Towards Fair Graph Generation", "abstract": "  There have been tremendous efforts over the past decades dedicated to the\ngeneration of realistic graphs in a variety of domains, ranging from social\nnetworks to computer networks, from gene regulatory networks to online\ntransaction networks. Despite the remarkable success, the vast majority of\nthese works are unsupervised in nature and are typically trained to minimize\nthe expected graph reconstruction loss, which would result in the\nrepresentation disparity issue in the generated graphs, i.e., the protected\ngroups (often minorities) contribute less to the objective and thus suffer from\nsystematically higher errors. In this paper, we aim to tailor graph generation\nto downstream mining tasks by leveraging label information and user-preferred\nparity constraints. In particular, we start from the investigation of\nrepresentation disparity in the context of graph generative models. To mitigate\nthe disparity, we propose a fairness-aware graph generative model named\nFairGen. Our model jointly trains a label-informed graph generation module and\na fair representation learning module by progressively learning the behaviors\nof the protected and unprotected groups, from the `easy' concepts to the `hard'\nones. In addition, we propose a generic context sampling strategy for graph\ngenerative models, which is proven to be capable of fairly capturing the\ncontextual information of each group with a high probability. Experimental\nresults on seven real-world data sets, including web-based graphs, demonstrate\nthat FairGen (1) obtains performance on par with state-of-the-art graph\ngenerative models across nine network properties, (2) mitigates the\nrepresentation disparity issues in the generated graphs, and (3) substantially\nboosts the model performance by up to 17% in downstream tasks via data\naugmentation.\n", "date": "2023-03-30", "categories": []}
{"id": "2308.09437", "title": "From Hope to Safety: Unlearning Biases of Deep Models via Gradient\n  Penalization in Latent Space", "abstract": "  Deep Neural Networks are prone to learning spurious correlations embedded in\nthe training data, leading to potentially biased predictions. This poses risks\nwhen deploying these models for high-stake decision-making, such as in medical\napplications. Current methods for post-hoc model correction either require\ninput-level annotations which are only possible for spatially localized biases,\nor augment the latent feature space, thereby hoping to enforce the right\nreasons. We present a novel method for model correction on the concept level\nthat explicitly reduces model sensitivity towards biases via gradient\npenalization. When modeling biases via Concept Activation Vectors, we highlight\nthe importance of choosing robust directions, as traditional regression-based\napproaches such as Support Vector Machines tend to result in diverging\ndirections. We effectively mitigate biases in controlled and real-world\nsettings on the ISIC, Bone Age, ImageNet and CelebA datasets using VGG, ResNet\nand EfficientNet architectures. Code is available on\nhttps://github.com/frederikpahde/rrclarc.\n", "date": "2023-08-18", "categories": []}
{"id": "2309.11343", "title": "Using Property Elicitation to Understand the Impacts of Fairness\n  Regularizers", "abstract": "  Predictive algorithms are often trained by optimizing some loss function, to\nwhich regularization functions are added to impose a penalty for violating\nconstraints. As expected, the addition of such regularization functions can\nchange the minimizer of the objective. It is not well-understood which\nregularizers change the minimizer of the loss, and, when the minimizer does\nchange, how it changes. We use property elicitation to take first steps towards\nunderstanding the joint relationship between the loss and regularization\nfunctions and the optimal decision for a given problem instance. In particular,\nwe give a necessary and sufficient condition on loss and regularizer pairs for\nwhen a property changes with the addition of the regularizer, and examine some\nregularizers satisfying this condition standard in the fair machine learning\nliterature. We empirically demonstrate how algorithmic decision-making changes\nas a function of both data distribution changes and hardness of the\nconstraints.\n", "date": "2023-09-20", "categories": []}
{"id": "2312.10065", "title": "Exploring Social Bias in Downstream Applications of Text-to-Image\n  Foundation Models", "abstract": "  Text-to-image diffusion models have been adopted into key commercial\nworkflows, such as art generation and image editing. Characterising the\nimplicit social biases they exhibit, such as gender and racial stereotypes, is\na necessary first step in avoiding discriminatory outcomes. While existing\nstudies on social bias focus on image generation, the biases exhibited in\nalternate applications of diffusion-based foundation models remain\nunder-explored. We propose methods that use synthetic images to probe two\napplications of diffusion models, image editing and classification, for social\nbias. Using our methodology, we uncover meaningful and significant\ninter-sectional social biases in \\textit{Stable Diffusion}, a state-of-the-art\nopen-source text-to-image model. Our findings caution against the uninformed\nadoption of text-to-image foundation models for downstream tasks and services.\n", "date": "2023-12-05", "categories": []}
{"id": "2312.10083", "title": "The Limits of Fair Medical Imaging AI In The Wild", "abstract": "  As artificial intelligence (AI) rapidly approaches human-level performance in\nmedical imaging, it is crucial that it does not exacerbate or propagate\nhealthcare disparities. Prior research has established AI's capacity to infer\ndemographic data from chest X-rays, leading to a key concern: do models using\ndemographic shortcuts have unfair predictions across subpopulations? In this\nstudy, we conduct a thorough investigation into the extent to which medical AI\nutilizes demographic encodings, focusing on potential fairness discrepancies\nwithin both in-distribution training sets and external test sets. Our analysis\ncovers three key medical imaging disciplines: radiology, dermatology, and\nophthalmology, and incorporates data from six global chest X-ray datasets. We\nconfirm that medical imaging AI leverages demographic shortcuts in disease\nclassification. While correcting shortcuts algorithmically effectively\naddresses fairness gaps to create \"locally optimal\" models within the original\ndata distribution, this optimality is not true in new test settings.\nSurprisingly, we find that models with less encoding of demographic attributes\nare often most \"globally optimal\", exhibiting better fairness during model\nevaluation in new test environments. Our work establishes best practices for\nmedical imaging models which maintain their performance and fairness in\ndeployments beyond their initial training contexts, underscoring critical\nconsiderations for AI clinical deployments across populations and sites.\n", "date": "2023-12-11", "categories": []}
{"id": "2312.10508", "title": "TrojFair: Trojan Fairness Attacks", "abstract": "  Deep learning models have been incorporated into high-stakes sectors,\nincluding healthcare diagnosis, loan approvals, and candidate recruitment,\namong others. Consequently, any bias or unfairness in these models can harm\nthose who depend on such models. In response, many algorithms have emerged to\nensure fairness in deep learning. However, while the potential for harm is\nsubstantial, the resilience of these fair deep learning models against\nmalicious attacks has never been thoroughly explored, especially in the context\nof emerging Trojan attacks. Moving beyond prior research, we aim to fill this\nvoid by introducing \\textit{TrojFair}, a Trojan fairness attack. Unlike\nexisting attacks, TrojFair is model-agnostic and crafts a Trojaned model that\nfunctions accurately and equitably for clean inputs. However, it displays\ndiscriminatory behaviors \\text{-} producing both incorrect and unfair results\n\\text{-} for specific groups with tainted inputs containing a trigger. TrojFair\nis a stealthy Fairness attack that is resilient to existing model fairness\naudition detectors since the model for clean inputs is fair. TrojFair achieves\na target group attack success rate exceeding $88.77\\%$, with an average\naccuracy loss less than $0.44\\%$. It also maintains a high discriminative score\nbetween the target and non-target groups across various datasets and models.\n", "date": "2023-12-16", "categories": []}
{"id": "2312.08559", "title": "Fair Active Learning in Low-Data Regimes", "abstract": "  In critical machine learning applications, ensuring fairness is essential to\navoid perpetuating social inequities. In this work, we address the challenges\nof reducing bias and improving accuracy in data-scarce environments, where the\ncost of collecting labeled data prohibits the use of large, labeled datasets.\nIn such settings, active learning promises to maximize marginal accuracy gains\nof small amounts of labeled data. However, existing applications of active\nlearning for fairness fail to deliver on this, typically requiring large\nlabeled datasets, or failing to ensure the desired fairness tolerance is met on\nthe population distribution.\n  To address such limitations, we introduce an innovative active learning\nframework that combines an exploration procedure inspired by posterior sampling\nwith a fair classification subroutine. We demonstrate that this framework\nperforms effectively in very data-scarce regimes, maximizing accuracy while\nsatisfying fairness constraints with high probability. We evaluate our proposed\napproach using well-established real-world benchmark datasets and compare it\nagainst state-of-the-art methods, demonstrating its effectiveness in producing\nfair models, and improvement over existing methods.\n", "date": "2023-12-13", "categories": []}
{"id": "2312.09061", "title": "Fair Clustering: A Causal Perspective", "abstract": "  Clustering algorithms may unintentionally propagate or intensify existing\ndisparities, leading to unfair representations or biased decision-making.\nCurrent fair clustering methods rely on notions of fairness that do not capture\nany information on the underlying causal mechanisms. We show that optimising\nfor non-causal fairness notions can paradoxically induce direct discriminatory\neffects from a causal standpoint. We present a clustering approach that\nincorporates causal fairness metrics to provide a more nuanced approach to\nfairness in unsupervised learning. Our approach enables the specification of\nthe causal fairness metrics that should be minimised. We demonstrate the\nefficacy of our methodology using datasets known to harbour unfair biases.\n", "date": "2023-12-14", "categories": []}
{"id": "2309.03648", "title": "Promoting Fairness in GNNs: A Characterization of Stability", "abstract": "  The Lipschitz bound, a technique from robust statistics, can limit the\nmaximum changes in the output concerning the input, taking into account\nassociated irrelevant biased factors. It is an efficient and provable method\nfor examining the output stability of machine learning models without incurring\nadditional computation costs. Recently, Graph Neural Networks (GNNs), which\noperate on non-Euclidean data, have gained significant attention. However, no\nprevious research has investigated the GNN Lipschitz bounds to shed light on\nstabilizing model outputs, especially when working on non-Euclidean data with\ninherent biases. Given the inherent biases in common graph data used for GNN\ntraining, it poses a serious challenge to constraining the GNN output\nperturbations induced by input biases, thereby safeguarding fairness during\ntraining. Recently, despite the Lipschitz constant's use in controlling the\nstability of Euclideanneural networks, the calculation of the precise Lipschitz\nconstant remains elusive for non-Euclidean neural networks like GNNs,\nespecially within fairness contexts. To narrow this gap, we begin with the\ngeneral GNNs operating on an attributed graph, and formulate a Lipschitz bound\nto limit the changes in the output regarding biases associated with the input.\nAdditionally, we theoretically analyze how the Lipschitz constant of a GNN\nmodel could constrain the output perturbations induced by biases learned from\ndata for fairness training. We experimentally validate the Lipschitz bound's\neffectiveness in limiting biases of the model output. Finally, from a training\ndynamics perspective, we demonstrate why the theoretical Lipschitz bound can\neffectively guide the GNN training to better trade-off between accuracy and\nfairness.\n", "date": "2023-09-07", "categories": []}
{"id": "2310.18368", "title": "Muslim-Violence Bias Persists in Debiased GPT Models", "abstract": "  Abid et al. (2021) showed a tendency in GPT-3 to generate mostly violent\ncompletions when prompted about Muslims, compared with other religions. Two\npre-registered replication attempts found few violent completions and only a\nweak anti-Muslim bias in the more recent InstructGPT, fine-tuned to eliminate\nbiased and toxic outputs. However, more pre-registered experiments showed that\nusing common names associated with the religions in prompts increases\nseveral-fold the rate of violent completions, revealing a significant\nsecond-order anti-Muslim bias. ChatGPT showed a bias many times stronger\nregardless of prompt format, suggesting that the effects of debiasing were\nreduced with continued model development. Our content analysis revealed\nreligion-specific themes containing offensive stereotypes across all\nexperiments. Our results show the need for continual de-biasing of models in\nways that address both explicit and higher-order associations.\n", "date": "2023-10-25", "categories": []}
{"id": "2312.05662", "title": "Understanding the Effect of Model Compression on Social Bias in Large\n  Language Models", "abstract": "  Large Language Models (LLMs) trained with self-supervision on vast corpora of\nweb text fit to the social biases of that text. Without intervention, these\nsocial biases persist in the model's predictions in downstream tasks, leading\nto representational harm. Many strategies have been proposed to mitigate the\neffects of inappropriate social biases learned during pretraining.\nSimultaneously, methods for model compression have become increasingly popular\nto reduce the computational burden of LLMs. Despite the popularity and need for\nboth approaches, little work has been done to explore the interplay between\nthese two. We perform a carefully controlled study of the impact of model\ncompression via quantization and knowledge distillation on measures of social\nbias in LLMs. Longer pretraining and larger models led to higher social bias,\nand quantization showed a regularizer effect with its best trade-off around 20%\nof the original pretraining time.\n", "date": "2023-12-09", "categories": []}
{"id": "2302.05906", "title": "On Comparing Fair Classifiers under Data Bias", "abstract": "  In this paper, we consider a theoretical model for injecting data bias,\nnamely, under-representation and label bias (Blum & Stangl, 2019). We\nempirically study the effect of varying data biases on the accuracy and\nfairness of fair classifiers. Through extensive experiments on both synthetic\nand real-world datasets (e.g., Adult, German Credit, Bank Marketing, COMPAS),\nwe empirically audit pre-, in-, and post-processing fair classifiers from\nstandard fairness toolkits for their fairness and accuracy by injecting varying\namounts of under-representation and label bias in their training data (but not\nthe test data). Our main observations are: 1. The fairness and accuracy of many\nstandard fair classifiers degrade severely as the bias injected in their\ntraining data increases, 2. A simple logistic regression model trained on the\nright data can often outperform, in both accuracy and fairness, most fair\nclassifiers trained on biased training data, and 3. A few, simple fairness\ntechniques (e.g., reweighing, exponentiated gradients) seem to offer stable\naccuracy and fairness guarantees even when their training data is injected with\nunder-representation and label bias. Our experiments also show how to integrate\na measure of data bias risk in the existing fairness dashboards for real-world\ndeployments.\n", "date": "2023-02-12", "categories": []}
{"id": "2302.08015", "title": "Individual Fairness under Uncertainty", "abstract": "  Algorithmic fairness, the research field of making machine learning (ML)\nalgorithms fair, is an established area in ML. As ML technologies expand their\napplication domains, including ones with high societal impact, it becomes\nessential to take fairness into consideration during the building of ML\nsystems. Yet, despite its wide range of socially sensitive applications, most\nwork treats the issue of algorithmic bias as an intrinsic property of\nsupervised learning, i.e., the class label is given as a precondition. Unlike\nprior studies in fairness, we propose an individual fairness measure and a\ncorresponding algorithm that deal with the challenges of uncertainty arising\nfrom censorship in class labels, while enforcing similar individuals to be\ntreated similarly from a ranking perspective, free of the Lipschitz condition\nin the conventional individual fairness definition. We argue that this\nperspective represents a more realistic model of fairness research for\nreal-world application deployment and show how learning with such a relaxed\nprecondition draws new insights that better explains algorithmic fairness. We\nconducted experiments on four real-world datasets to evaluate our proposed\nmethod compared to other fairness models, demonstrating its superiority in\nminimizing discrimination while maintaining predictive performance with\nuncertainty present.\n", "date": "2023-02-16", "categories": []}
{"id": "2309.08836", "title": "Bias and Fairness in Chatbots: An Overview", "abstract": "  Chatbots have been studied for more than half a century. With the rapid\ndevelopment of natural language processing (NLP) technologies in recent years,\nchatbots using large language models (LLMs) have received much attention\nnowadays. Compared with traditional ones, modern chatbots are more powerful and\nhave been used in real-world applications. There are however, bias and fairness\nconcerns in modern chatbot design. Due to the huge amounts of training data,\nextremely large model sizes, and lack of interpretability, bias mitigation and\nfairness preservation of modern chatbots are challenging. Thus, a comprehensive\noverview on bias and fairness in chatbot systems is given in this paper. The\nhistory of chatbots and their categories are first reviewed. Then, bias sources\nand potential harms in applications are analyzed. Considerations in designing\nfair and unbiased chatbot systems are examined. Finally, future research\ndirections are discussed.\n", "date": "2023-09-16", "categories": []}
{"id": "2312.05551", "title": "Multi-dimensional Fair Federated Learning", "abstract": "  Federated learning (FL) has emerged as a promising collaborative and secure\nparadigm for training a model from decentralized data without compromising\nprivacy. Group fairness and client fairness are two dimensions of fairness that\nare important for FL. Standard FL can result in disproportionate disadvantages\nfor certain clients, and it still faces the challenge of treating different\ngroups equitably in a population. The problem of privately training fair FL\nmodels without compromising the generalization capability of disadvantaged\nclients remains open. In this paper, we propose a method, called mFairFL, to\naddress this problem and achieve group fairness and client fairness\nsimultaneously. mFairFL leverages differential multipliers to construct an\noptimization objective for empirical risk minimization with fairness\nconstraints. Before aggregating locally trained models, it first detects\nconflicts among their gradients, and then iteratively curates the direction and\nmagnitude of gradients to mitigate these conflicts. Theoretical analysis proves\nmFairFL facilitates the fairness in model development. The experimental\nevaluations based on three benchmark datasets show significant advantages of\nmFairFL compared to seven state-of-the-art baselines.\n", "date": "2023-12-09", "categories": []}
{"id": "2304.09424", "title": "Loss Minimization Yields Multicalibration for Large Neural Networks", "abstract": "  Multicalibration is a notion of fairness for predictors that requires them to\nprovide calibrated predictions across a large set of protected groups.\nMulticalibration is known to be a distinct goal than loss minimization, even\nfor simple predictors such as linear functions.\n  In this work, we consider the setting where the protected groups can be\nrepresented by neural networks of size $k$, and the predictors are neural\nnetworks of size $n > k$. We show that minimizing the squared loss over all\nneural nets of size $n$ implies multicalibration for all but a bounded number\nof unlucky values of $n$. We also give evidence that our bound on the number of\nunlucky values is tight, given our proof technique. Previously, results of the\nflavor that loss minimization yields multicalibration were known only for\npredictors that were near the ground truth, hence were rather limited in\napplicability. Unlike these, our results rely on the expressivity of neural\nnets and utilize the representation of the predictor.\n", "date": "2023-04-19", "categories": []}
{"id": "2312.04745", "title": "A Brief Tutorial on Sample Size Calculations for Fairness Audits", "abstract": "  In fairness audits, a standard objective is to detect whether a given\nalgorithm performs substantially differently between subgroups. Properly\npowering the statistical analysis of such audits is crucial for obtaining\ninformative fairness assessments, as it ensures a high probability of detecting\nunfairness when it exists. However, limited guidance is available on the amount\nof data necessary for a fairness audit, lacking directly applicable results\nconcerning commonly used fairness metrics. Additionally, the consideration of\nunequal subgroup sample sizes is also missing. In this tutorial, we address\nthese issues by providing guidance on how to determine the required subgroup\nsample sizes to maximize the statistical power of hypothesis tests for\ndetecting unfairness. Our findings are applicable to audits of binary\nclassification models and multiple fairness metrics derived as summaries of the\nconfusion matrix. Furthermore, we discuss other aspects of audit study designs\nthat can increase the reliability of audit results.\n", "date": "2023-12-07", "categories": []}
{"id": "2210.09943", "title": "Rethinking Bias Mitigation: Fairer Architectures Make for Fairer Face\n  Recognition", "abstract": "  Face recognition systems are widely deployed in safety-critical applications,\nincluding law enforcement, yet they exhibit bias across a range of\nsocio-demographic dimensions, such as gender and race. Conventional wisdom\ndictates that model biases arise from biased training data. As a consequence,\nprevious works on bias mitigation largely focused on pre-processing the\ntraining data, adding penalties to prevent bias from effecting the model during\ntraining, or post-processing predictions to debias them, yet these approaches\nhave shown limited success on hard problems such as face recognition. In our\nwork, we discover that biases are actually inherent to neural network\narchitectures themselves. Following this reframing, we conduct the first neural\narchitecture search for fairness, jointly with a search for hyperparameters.\nOur search outputs a suite of models which Pareto-dominate all other\nhigh-performance architectures and existing bias mitigation methods in terms of\naccuracy and fairness, often by large margins, on the two most widely used\ndatasets for face identification, CelebA and VGGFace2. Furthermore, these\nmodels generalize to other datasets and sensitive attributes. We release our\ncode, models and raw data files at https://github.com/dooleys/FR-NAS.\n", "date": "2022-10-18", "categories": []}
{"id": "2311.09428", "title": "Beyond Detection: Unveiling Fairness Vulnerabilities in Abusive Language\n  Models", "abstract": "  This work investigates the potential of undermining both fairness and\ndetection performance in abusive language detection. In a dynamic and complex\ndigital world, it is crucial to investigate the vulnerabilities of these\ndetection models to adversarial fairness attacks to improve their fairness\nrobustness. We propose a simple yet effective framework FABLE that leverages\nbackdoor attacks as they allow targeted control over the fairness and detection\nperformance. FABLE explores three types of trigger designs (i.e., rare,\nartificial, and natural triggers) and novel sampling strategies. Specifically,\nthe adversary can inject triggers into samples in the minority group with the\nfavored outcome (i.e., \"non-abusive\") and flip their labels to the unfavored\noutcome, i.e., \"abusive\". Experiments on benchmark datasets demonstrate the\neffectiveness of FABLE attacking fairness and utility in abusive language\ndetection.\n", "date": "2023-11-15", "categories": []}
{"id": "2312.03253", "title": "Seller-side Outcome Fairness in Online Marketplaces", "abstract": "  This paper aims to investigate and achieve seller-side fairness within online\nmarketplaces, where many sellers and their items are not sufficiently exposed\nto customers in an e-commerce platform. This phenomenon raises concerns\nregarding the potential loss of revenue associated with less exposed items as\nwell as less marketplace diversity. We introduce the notion of seller-side\noutcome fairness and build an optimization model to balance collected\nrecommendation rewards and the fairness metric. We then propose a\ngradient-based data-driven algorithm based on the duality and bandit theory.\nOur numerical experiments on real e-commerce data sets show that our algorithm\ncan lift seller fairness measures while not hurting metrics like collected\nGross Merchandise Value (GMV) and total purchases.\n", "date": "2023-12-06", "categories": []}
{"id": "2312.03689", "title": "Evaluating and Mitigating Discrimination in Language Model Decisions", "abstract": "  As language models (LMs) advance, interest is growing in applying them to\nhigh-stakes societal decisions, such as determining financing or housing\neligibility. However, their potential for discrimination in such contexts\nraises ethical concerns, motivating the need for better methods to evaluate\nthese risks. We present a method for proactively evaluating the potential\ndiscriminatory impact of LMs in a wide range of use cases, including\nhypothetical use cases where they have not yet been deployed. Specifically, we\nuse an LM to generate a wide array of potential prompts that decision-makers\nmay input into an LM, spanning 70 diverse decision scenarios across society,\nand systematically vary the demographic information in each prompt. Applying\nthis methodology reveals patterns of both positive and negative discrimination\nin the Claude 2.0 model in select settings when no interventions are applied.\nWhile we do not endorse or permit the use of language models to make automated\ndecisions for the high-risk use cases we study, we demonstrate techniques to\nsignificantly decrease both positive and negative discrimination through\ncareful prompt engineering, providing pathways toward safer deployment in use\ncases where they may be appropriate. Our work enables developers and\npolicymakers to anticipate, measure, and address discrimination as language\nmodel capabilities and applications continue to expand. We release our dataset\nand prompts at https://huggingface.co/datasets/Anthropic/discrim-eval\n", "date": "2023-12-06", "categories": []}
{"id": "2312.02501", "title": "Inspecting Model Fairness in Ultrasound Segmentation Tasks", "abstract": "  With the rapid expansion of machine learning and deep learning (DL),\nresearchers are increasingly employing learning-based algorithms to alleviate\ndiagnostic challenges across diverse medical tasks and applications. While\nadvancements in diagnostic precision are notable, some researchers have\nidentified a concerning trend: their models exhibit biased performance across\nsubgroups characterized by different sensitive attributes. This bias not only\ninfringes upon the rights of patients but also has the potential to lead to\nlife-altering consequences. In this paper, we inspect a series of DL\nsegmentation models using two ultrasound datasets, aiming to assess the\npresence of model unfairness in these specific tasks. Our findings reveal that\neven state-of-the-art DL algorithms demonstrate unfair behavior in ultrasound\nsegmentation tasks. These results serve as a crucial warning, underscoring the\nnecessity for careful model evaluation before their deployment in real-world\nscenarios. Such assessments are imperative to ensure ethical considerations and\nmitigate the risk of adverse impacts on patient outcomes.\n", "date": "2023-12-05", "categories": []}
{"id": "2302.03750", "title": "Linking convolutional kernel size to generalization bias in face\n  analysis CNNs", "abstract": "  Training dataset biases are by far the most scrutinized factors when\nexplaining algorithmic biases of neural networks. In contrast, hyperparameters\nrelated to the neural network architecture have largely been ignored even\nthough different network parameterizations are known to induce different\nimplicit biases over learned features. For example, convolutional kernel size\nis known to affect the frequency content of features learned in CNNs. In this\nwork, we present a causal framework for linking an architectural hyperparameter\nto out-of-distribution algorithmic bias. Our framework is experimental, in that\nwe train several versions of a network with an intervention to a specific\nhyperparameter, and measure the resulting causal effect of this choice on\nperformance bias when a particular out-of-distribution image perturbation is\napplied. In our experiments, we focused on measuring the causal relationship\nbetween convolutional kernel size and face analysis classification bias across\ndifferent subpopulations (race/gender), with respect to high-frequency image\ndetails. We show that modifying kernel size, even in one layer of a CNN,\nchanges the frequency content of learned features significantly across data\nsubgroups leading to biased generalization performance even in the presence of\na balanced dataset.\n", "date": "2023-02-07", "categories": []}
{"id": "2305.14735", "title": "Centering the Margins: Outlier-Based Identification of Harmed\n  Populations in Toxicity Detection", "abstract": "  The impact of AI models on marginalized communities has traditionally been\nmeasured by identifying performance differences between specified demographic\nsubgroups. Though this approach aims to center vulnerable groups, it risks\nobscuring patterns of harm faced by intersectional subgroups or shared across\nmultiple groups. To address this, we draw on theories of marginalization from\ndisability studies and related disciplines, which state that people farther\nfrom the norm face greater adversity, to consider the \"margins\" in the domain\nof toxicity detection. We operationalize the \"margins\" of a dataset by\nemploying outlier detection to identify text about people with demographic\nattributes distant from the \"norm\". We find that model performance is\nconsistently worse for demographic outliers, with mean squared error (MSE)\nbetween outliers and non-outliers up to 70.4% worse across toxicity types. It\nis also worse for text outliers, with a MSE up to 68.4% higher for outliers\nthan non-outliers. We also find text and demographic outliers to be\nparticularly susceptible to errors in the classification of severe toxicity and\nidentity attacks. Compared to analysis of disparities using traditional\ndemographic breakdowns, we find that our outlier analysis frequently surfaces\ngreater harms faced by a larger, more intersectional group, which suggests that\noutlier analysis is particularly beneficial for identifying harms against those\ngroups.\n", "date": "2023-05-24", "categories": []}
{"id": "2306.13440", "title": "Trading-off price for data quality to achieve fair online allocation", "abstract": "  We consider the problem of online allocation subject to a long-term fairness\npenalty. Contrary to existing works, however, we do not assume that the\ndecision-maker observes the protected attributes -- which is often unrealistic\nin practice. Instead they can purchase data that help estimate them from\nsources of different quality; and hence reduce the fairness penalty at some\ncost. We model this problem as a multi-armed bandit problem where each arm\ncorresponds to the choice of a data source, coupled with the online allocation\nproblem. We propose an algorithm that jointly solves both problems and show\nthat it has a regret bounded by $\\mathcal{O}(\\sqrt{T})$. A key difficulty is\nthat the rewards received by selecting a source are correlated by the fairness\npenalty, which leads to a need for randomization (despite a stochastic\nsetting). Our algorithm takes into account contextual information available\nbefore the source selection, and can adapt to many different fairness notions.\nWe also show that in some instances, the estimates used can be learned on the\nfly.\n", "date": "2023-06-23", "categories": []}
{"id": "2310.09219", "title": "\"Kelly is a Warm Person, Joseph is a Role Model\": Gender Biases in\n  LLM-Generated Reference Letters", "abstract": "  Large Language Models (LLMs) have recently emerged as an effective tool to\nassist individuals in writing various types of content, including professional\ndocuments such as recommendation letters. Though bringing convenience, this\napplication also introduces unprecedented fairness concerns. Model-generated\nreference letters might be directly used by users in professional scenarios. If\nunderlying biases exist in these model-constructed letters, using them without\nscrutinization could lead to direct societal harms, such as sabotaging\napplication success rates for female applicants. In light of this pressing\nissue, it is imminent and necessary to comprehensively study fairness issues\nand associated harms in this real-world use case. In this paper, we critically\nexamine gender biases in LLM-generated reference letters. Drawing inspiration\nfrom social science findings, we design evaluation methods to manifest biases\nthrough 2 dimensions: (1) biases in language style and (2) biases in lexical\ncontent. We further investigate the extent of bias propagation by analyzing the\nhallucination bias of models, a term that we define to be bias exacerbation in\nmodel-hallucinated contents. Through benchmarking evaluation on 2 popular LLMs-\nChatGPT and Alpaca, we reveal significant gender biases in LLM-generated\nrecommendation letters. Our findings not only warn against using LLMs for this\napplication without scrutinization, but also illuminate the importance of\nthoroughly studying hidden biases and harms in LLM-generated professional\ndocuments.\n", "date": "2023-10-13", "categories": []}
{"id": "2312.01509", "title": "Tackling Bias in Pre-trained Language Models: Current Trends and\n  Under-represented Societies", "abstract": "  The benefits and capabilities of pre-trained language models (LLMs) in\ncurrent and future innovations are vital to any society. However, introducing\nand using LLMs comes with biases and discrimination, resulting in concerns\nabout equality, diversity and fairness, and must be addressed. While\nunderstanding and acknowledging bias in LLMs and developing mitigation\nstrategies are crucial, the generalised assumptions towards societal needs can\nresult in disadvantages towards under-represented societies and indigenous\npopulations. Furthermore, the ongoing changes to actual and proposed amendments\nto regulations and laws worldwide also impact research capabilities in tackling\nthe bias problem. This research presents a comprehensive survey synthesising\nthe current trends and limitations in techniques used for identifying and\nmitigating bias in LLMs, where the overview of methods for tackling bias are\ngrouped into metrics, benchmark datasets, and mitigation strategies. The\nimportance and novelty of this survey are that it explores the perspective of\nunder-represented societies. We argue that current practices tackling the bias\nproblem cannot simply be 'plugged in' to address the needs of under-represented\nsocieties. We use examples from New Zealand to present requirements for\nadopting existing techniques to under-represented societies.\n", "date": "2023-12-03", "categories": []}
{"id": "2306.03775", "title": "Matched Pair Calibration for Ranking Fairness", "abstract": "  We propose a test of fairness in score-based ranking systems called matched\npair calibration. Our approach constructs a set of matched item pairs with\nminimal confounding differences between subgroups before computing an\nappropriate measure of ranking error over the set. The matching step ensures\nthat we compare subgroup outcomes between identically scored items so that\nmeasured performance differences directly imply unfairness in subgroup-level\nexposures. We show how our approach generalizes the fairness intuitions of\ncalibration from a binary classification setting to ranking and connect our\napproach to other proposals for ranking fairness measures. Moreover, our\nstrategy shows how the logic of marginal outcome tests extends to cases where\nthe analyst has access to model scores. Lastly, we provide an example of\napplying matched pair calibration to a real-word ranking data set to\ndemonstrate its efficacy in detecting ranking bias.\n", "date": "2023-06-06", "categories": []}
{"id": "2312.00434", "title": "PEFTDebias : Capturing debiasing information using PEFTs", "abstract": "  The increasing use of foundation models highlights the urgent need to address\nand eliminate implicit biases present in them that arise during pretraining. In\nthis paper, we introduce PEFTDebias, a novel approach that employs\nparameter-efficient fine-tuning (PEFT) to mitigate the biases within foundation\nmodels. PEFTDebias consists of two main phases: an upstream phase for acquiring\ndebiasing parameters along a specific bias axis, and a downstream phase where\nthese parameters are incorporated into the model and frozen during the\nfine-tuning process. By evaluating on four datasets across two bias axes namely\ngender and race, we find that downstream biases can be effectively reduced with\nPEFTs. In addition, we show that these parameters possess axis-specific\ndebiasing characteristics, enabling their effective transferability in\nmitigating biases in various downstream tasks. To ensure reproducibility, we\nrelease the code to do our experiments.\n", "date": "2023-12-01", "categories": []}
{"id": "2312.00765", "title": "Explaining Knock-on Effects of Bias Mitigation", "abstract": "  In machine learning systems, bias mitigation approaches aim to make outcomes\nfairer across privileged and unprivileged groups. Bias mitigation methods work\nin different ways and have known \"waterfall\" effects, e.g., mitigating bias at\none place may manifest bias elsewhere. In this paper, we aim to characterise\nimpacted cohorts when mitigation interventions are applied. To do so, we treat\nintervention effects as a classification task and learn an explainable\nmeta-classifier to identify cohorts that have altered outcomes. We examine a\nrange of bias mitigation strategies that work at various stages of the model\nlife cycle. We empirically demonstrate that our meta-classifier is able to\nuncover impacted cohorts. Further, we show that all tested mitigation\nstrategies negatively impact a non-trivial fraction of cases, i.e., people who\nreceive unfavourable outcomes solely on account of mitigation efforts. This is\ndespite improvement in fairness metrics. We use these results as a basis to\nargue for more careful audits of static mitigation interventions that go beyond\naggregate metrics.\n", "date": "2023-12-01", "categories": []}
{"id": "2311.18140", "title": "ROBBIE: Robust Bias Evaluation of Large Generative Language Models", "abstract": "  As generative large language models (LLMs) grow more performant and\nprevalent, we must develop comprehensive enough tools to measure and improve\ntheir fairness. Different prompt-based datasets can be used to measure social\nbias across multiple text domains and demographic axes, meaning that testing\nLLMs on more datasets can potentially help us characterize their biases more\nfully, and better ensure equal and equitable treatment of marginalized\ndemographic groups. In this work, our focus is two-fold:\n  (1) Benchmarking: a comparison of 6 different prompt-based bias and toxicity\nmetrics across 12 demographic axes and 5 families of generative LLMs. Out of\nthose 6 metrics, AdvPromptSet and HolisticBiasR are novel datasets proposed in\nthe paper. The comparison of those benchmarks gives us insights about the bias\nand toxicity of the compared models. Therefore, we explore the frequency of\ndemographic terms in common LLM pre-training corpora and how this may relate to\nmodel biases.\n  (2) Mitigation: we conduct a comprehensive study of how well 3 bias/toxicity\nmitigation techniques perform across our suite of measurements. ROBBIE aims to\nprovide insights for practitioners while deploying a model, emphasizing the\nneed to not only measure potential harms, but also understand how they arise by\ncharacterizing the data, mitigate harms once found, and balance any trade-offs.\nWe open-source our analysis code in hopes of encouraging broader measurements\nof bias in future LLMs.\n", "date": "2023-11-29", "categories": []}
{"id": "2303.17713", "title": "Mitigating Source Bias for Fairer Weak Supervision", "abstract": "  Weak supervision enables efficient development of training sets by reducing\nthe need for ground truth labels. However, the techniques that make weak\nsupervision attractive -- such as integrating any source of signal to estimate\nunknown labels -- also entail the danger that the produced pseudolabels are\nhighly biased. Surprisingly, given everyday use and the potential for increased\nbias, weak supervision has not been studied from the point of view of fairness.\nWe begin such a study, starting with the observation that even when a fair\nmodel can be built from a dataset with access to ground-truth labels, the\ncorresponding dataset labeled via weak supervision can be arbitrarily unfair.\nTo address this, we propose and empirically validate a model for source\nunfairness in weak supervision, then introduce a simple counterfactual\nfairness-based technique that can mitigate these biases. Theoretically, we show\nthat it is possible for our approach to simultaneously improve both accuracy\nand fairness -- in contrast to standard fairness approaches that suffer from\ntradeoffs. Empirically, we show that our technique improves accuracy on weak\nsupervision baselines by as much as 32\\% while reducing demographic parity gap\nby 82.5\\%. A simple extension of our method aimed at maximizing performance\nproduces state-of-the-art performance in five out of ten datasets in the WRENCH\nbenchmark.\n", "date": "2023-03-30", "categories": []}
{"id": "2311.17373", "title": "The Devil is in the Data: Learning Fair Graph Neural Networks via\n  Partial Knowledge Distillation", "abstract": "  Graph neural networks (GNNs) are being increasingly used in many high-stakes\ntasks, and as a result, there is growing attention on their fairness recently.\nGNNs have been shown to be unfair as they tend to make discriminatory decisions\ntoward certain demographic groups, divided by sensitive attributes such as\ngender and race. While recent works have been devoted to improving their\nfairness performance, they often require accessible demographic information.\nThis greatly limits their applicability in real-world scenarios due to legal\nrestrictions. To address this problem, we present a demographic-agnostic method\nto learn fair GNNs via knowledge distillation, namely FairGKD. Our work is\nmotivated by the empirical observation that training GNNs on partial data\n(i.e., only node attributes or topology data) can improve their fairness,\nalbeit at the cost of utility. To make a balanced trade-off between fairness\nand utility performance, we employ a set of fairness experts (i.e., GNNs\ntrained on different partial data) to construct the synthetic teacher, which\ndistills fairer and informative knowledge to guide the learning of the GNN\nstudent. Experiments on several benchmark datasets demonstrate that FairGKD,\nwhich does not require access to demographic information, significantly\nimproves the fairness of GNNs by a large margin while maintaining their\nutility.\n", "date": "2023-11-29", "categories": []}
{"id": "2311.16528", "title": "Utility Fairness in Contextual Dynamic Pricing with Demand Learning", "abstract": "  This paper introduces a novel contextual bandit algorithm for personalized\npricing under utility fairness constraints in scenarios with uncertain demand,\nachieving an optimal regret upper bound. Our approach, which incorporates\ndynamic pricing and demand learning, addresses the critical challenge of\nfairness in pricing strategies. We first delve into the static full-information\nsetting to formulate an optimal pricing policy as a constrained optimization\nproblem. Here, we propose an approximation algorithm for efficiently and\napproximately computing the ideal policy.\n  We also use mathematical analysis and computational studies to characterize\nthe structures of optimal contextual pricing policies subject to fairness\nconstraints, deriving simplified policies which lays the foundations of more\nin-depth research and extensions.\n  Further, we extend our study to dynamic pricing problems with demand\nlearning, establishing a non-standard regret lower bound that highlights the\ncomplexity added by fairness constraints. Our research offers a comprehensive\nanalysis of the cost of fairness and its impact on the balance between utility\nand revenue maximization. This work represents a step towards integrating\nethical considerations into algorithmic efficiency in data-driven dynamic\npricing.\n", "date": "2023-11-28", "categories": []}
{"id": "2311.14733", "title": "Thinking Outside the Box: Orthogonal Approach to Equalizing Protected\n  Attributes", "abstract": "  There is growing concern that the potential of black box AI may exacerbate\nhealth-related disparities and biases such as gender and ethnicity in clinical\ndecision-making. Biased decisions can arise from data availability and\ncollection processes, as well as from the underlying confounding effects of the\nprotected attributes themselves. This work proposes a machine learning-based\northogonal approach aiming to analyze and suppress the effect of the confounder\nthrough discriminant dimensionality reduction and orthogonalization of the\nprotected attributes against the primary attribute information. By doing so,\nthe impact of the protected attributes on disease diagnosis can be realized,\nundesirable feature correlations can be mitigated, and the model prediction\nperformance can be enhanced.\n", "date": "2023-11-21", "categories": []}
{"id": "2311.15691", "title": "Automated discovery of trade-off between utility, privacy and fairness\n  in machine learning models", "abstract": "  Machine learning models are deployed as a central component in decision\nmaking and policy operations with direct impact on individuals' lives. In order\nto act ethically and comply with government regulations, these models need to\nmake fair decisions and protect the users' privacy. However, such requirements\ncan come with decrease in models' performance compared to their potentially\nbiased, privacy-leaking counterparts. Thus the trade-off between fairness,\nprivacy and performance of ML models emerges, and practitioners need a way of\nquantifying this trade-off to enable deployment decisions. In this work we\ninterpret this trade-off as a multi-objective optimization problem, and propose\nPFairDP, a pipeline that uses Bayesian optimization for discovery of\nPareto-optimal points between fairness, privacy and utility of ML models. We\nshow how PFairDP can be used to replicate known results that were achieved\nthrough manual constraint setting process. We further demonstrate effectiveness\nof PFairDP with experiments on multiple models and datasets.\n", "date": "2023-11-27", "categories": []}
{"id": "2311.13766", "title": "A Unified Framework for Fair Spectral Clustering With Effective Graph\n  Learning", "abstract": "  We consider the problem of spectral clustering under group fairness\nconstraints, where samples from each sensitive group are approximately\nproportionally represented in each cluster. Traditional fair spectral\nclustering (FSC) methods consist of two consecutive stages, i.e., performing\nfair spectral embedding on a given graph and conducting $k$means to obtain\ndiscrete cluster labels. However, in practice, the graph is usually unknown,\nand we need to construct the underlying graph from potentially noisy data, the\nquality of which inevitably affects subsequent fair clustering performance.\nFurthermore, performing FSC through separate steps breaks the connections among\nthese steps, leading to suboptimal results. To this end, we first theoretically\nanalyze the effect of the constructed graph on FSC. Motivated by the analysis,\nwe propose a novel graph construction method with a node-adaptive graph filter\nto learn graphs from noisy data. Then, all independent stages of conventional\nFSC are integrated into a single objective function, forming an end-to-end\nframework that inputs raw data and outputs discrete cluster labels. An\nalgorithm is developed to jointly and alternately update the variables in each\nstage. Finally, we conduct extensive experiments on synthetic, benchmark, and\nreal data, which show that our model is superior to state-of-the-art fair\nclustering methods.\n", "date": "2023-11-23", "categories": []}
{"id": "2311.12447", "title": "Designing Long-term Group Fair Policies in Dynamical Systems", "abstract": "  Neglecting the effect that decisions have on individuals (and thus, on the\nunderlying data distribution) when designing algorithmic decision-making\npolicies may increase inequalities and unfairness in the long term - even if\nfairness considerations were taken in the policy design process. In this paper,\nwe propose a novel framework for achieving long-term group fairness in\ndynamical systems, in which current decisions may affect an individual's\nfeatures in the next step, and thus, future decisions. Specifically, our\nframework allows us to identify a time-independent policy that converges, if\ndeployed, to the targeted fair stationary state of the system in the long term,\nindependently of the initial data distribution. We model the system dynamics\nwith a time-homogeneous Markov chain and optimize the policy leveraging the\nMarkov chain convergence theorem to ensure unique convergence. We provide\nexamples of different targeted fair states of the system, encompassing a range\nof long-term goals for society and policymakers. Furthermore, we show how our\napproach facilitates the evaluation of different long-term targets by examining\ntheir impact on the group-conditional population distribution in the long term\nand how it evolves until convergence.\n", "date": "2023-11-21", "categories": []}
{"id": "2311.12501", "title": "Fair Polylog-Approximate Low-Cost Hierarchical Clustering", "abstract": "  Research in fair machine learning, and particularly clustering, has been\ncrucial in recent years given the many ethical controversies that modern\nintelligent systems have posed. Ahmadian et al. [2020] established the study of\nfairness in \\textit{hierarchical} clustering, a stronger, more structured\nvariant of its well-known flat counterpart, though their proposed algorithm\nthat optimizes for Dasgupta's [2016] famous cost function was highly\ntheoretical. Knittel et al. [2023] then proposed the first practical fair\napproximation for cost, however they were unable to break the\npolynomial-approximate barrier they posed as a hurdle of interest. We break\nthis barrier, proposing the first truly polylogarithmic-approximate low-cost\nfair hierarchical clustering, thus greatly bridging the gap between the best\nfair and vanilla hierarchical clustering approximations.\n", "date": "2023-11-21", "categories": []}
{"id": "2311.12684", "title": "Adversarial Reweighting Guided by Wasserstein Distance for Bias\n  Mitigation", "abstract": "  The unequal representation of different groups in a sample population can\nlead to discrimination of minority groups when machine learning models make\nautomated decisions. To address these issues, fairness-aware machine learning\njointly optimizes two (or more) metrics aiming at predictive effectiveness and\nlow unfairness. However, the inherent under-representation of minorities in the\ndata makes the disparate treatment of subpopulations less noticeable and\ndifficult to deal with during learning. In this paper, we propose a novel\nadversarial reweighting method to address such \\emph{representation bias}. To\nbalance the data distribution between the majority and the minority groups, our\napproach deemphasizes samples from the majority group. To minimize empirical\nrisk, our method prefers samples from the majority group that are close to the\nminority group as evaluated by the Wasserstein distance. Our theoretical\nanalysis shows the effectiveness of our adversarial reweighting approach.\nExperiments demonstrate that our approach mitigates bias without sacrificing\nclassification accuracy, outperforming related state-of-the-art methods on\nimage and tabular benchmark datasets.\n", "date": "2023-11-21", "categories": []}
{"id": "2311.12689", "title": "Fair Text Classification with Wasserstein Independence", "abstract": "  Group fairness is a central research topic in text classification, where\nreaching fair treatment between sensitive groups (e.g. women vs. men) remains\nan open challenge. This paper presents a novel method for mitigating biases in\nneural text classification, agnostic to the model architecture. Considering the\ndifficulty to distinguish fair from unfair information in a text encoder, we\ntake inspiration from adversarial training to induce Wasserstein independence\nbetween representations learned to predict our target label and the ones\nlearned to predict some sensitive attribute. Our approach provides two\nsignificant advantages. Firstly, it does not require annotations of sensitive\nattributes in both testing and training data. This is more suitable for\nreal-life scenarios compared to existing methods that require annotations of\nsensitive attributes at train time. Second, our approach exhibits a comparable\nor better fairness-accuracy trade-off compared to existing methods.\n", "date": "2023-11-21", "categories": []}
{"id": "2311.11911", "title": "Certification of Distributional Individual Fairness", "abstract": "  Providing formal guarantees of algorithmic fairness is of paramount\nimportance to socially responsible deployment of machine learning algorithms.\nIn this work, we study formal guarantees, i.e., certificates, for individual\nfairness (IF) of neural networks. We start by introducing a novel convex\napproximation of IF constraints that exponentially decreases the computational\ncost of providing formal guarantees of local individual fairness. We highlight\nthat prior methods are constrained by their focus on global IF certification\nand can therefore only scale to models with a few dozen hidden neurons, thus\nlimiting their practical impact. We propose to certify distributional\nindividual fairness which ensures that for a given empirical distribution and\nall distributions within a $\\gamma$-Wasserstein ball, the neural network has\nguaranteed individually fair predictions. Leveraging developments in\nquasi-convex optimization, we provide novel and efficient certified bounds on\ndistributional individual fairness and show that our method allows us to\ncertify and regularize neural networks that are several orders of magnitude\nlarger than those considered by prior works. Moreover, we study real-world\ndistribution shifts and find our bounds to be a scalable, practical, and sound\nsource of IF guarantees.\n", "date": "2023-11-20", "categories": []}
{"id": "2311.10223", "title": "Asymptotically Fair Participation in Machine Learning Models: an Optimal\n  Control Perspective", "abstract": "  The performance of state-of-the-art machine learning models often\ndeteriorates when testing on demographics that are under-represented in the\ntraining dataset. This problem has predominately been studied in a supervised\nlearning setting where the data distribution is static. However, real-world\napplications often involve distribution shifts caused by the deployed models.\nFor instance, the performance disparity against monitory users can lead to a\nhigh customer churn rate, thus the available data provided by active users are\nskewed due to the lack of minority users. This feedback effect further\nexacerbates the disparity among different demographic groups in future steps.\nTo address this issue, we propose asymptotically fair participation as a\ncondition to maintain long-term model performance over all demographic groups.\nIn this work, we aim to address the problem of achieving asymptotically fair\nparticipation via optimal control formulation. Moreover, we design a surrogate\nretention system based on existing literature on evolutionary population\ndynamics to approximate the dynamics of distribution shifts on active user\ncounts, from which the objective of achieving asymptotically fair participation\nis formulated as an optimal control problem, and the control variables are\nconsidered as the model parameters. We apply an efficient implementation of\nPontryagin's maximum principle to estimate the optimal control solution. To\nevaluate the effectiveness of the proposed method, we design a generic\nsimulation environment that simulates the population dynamics of the feedback\neffect between user retention and model performance. When we deploy the\nresulting models to the simulation environment, the optimal control solution\naccounts for long-term planning and leads to superior performance compared with\nexisting baseline methods.\n", "date": "2023-11-16", "categories": []}
{"id": "2311.10512", "title": "Causal Fairness-Guided Dataset Reweighting using Neural Networks", "abstract": "  The importance of achieving fairness in machine learning models cannot be\noverstated. Recent research has pointed out that fairness should be examined\nfrom a causal perspective, and several fairness notions based on the on Pearl's\ncausal framework have been proposed. In this paper, we construct a reweighting\nscheme of datasets to address causal fairness. Our approach aims at mitigating\nbias by considering the causal relationships among variables and incorporating\nthem into the reweighting process. The proposed method adopts two neural\nnetworks, whose structures are intentionally used to reflect the structures of\na causal graph and of an interventional graph. The two neural networks can\napproximate the causal model of the data, and the causal model of\ninterventions. Furthermore, reweighting guided by a discriminator is applied to\nachieve various fairness notions. Experiments on real-world datasets show that\nour method can achieve causal fairness on the data while remaining close to the\noriginal data for downstream tasks.\n", "date": "2023-11-17", "categories": []}
{"id": "2310.04955", "title": "Information-Theoretic Bounds on The Removal of Attribute-Specific Bias\n  From Neural Networks", "abstract": "  Ensuring a neural network is not relying on protected attributes (e.g., race,\nsex, age) for predictions is crucial in advancing fair and trustworthy AI.\nWhile several promising methods for removing attribute bias in neural networks\nhave been proposed, their limitations remain under-explored. In this work, we\nmathematically and empirically reveal an important limitation of attribute bias\nremoval methods in presence of strong bias. Specifically, we derive a general\nnon-vacuous information-theoretical upper bound on the performance of any\nattribute bias removal method in terms of the bias strength. We provide\nextensive experiments on synthetic, image, and census datasets to verify the\ntheoretical bound and its consequences in practice. Our findings show that\nexisting attribute bias removal methods are effective only when the inherent\nbias in the dataset is relatively weak, thus cautioning against the use of\nthese methods in smaller datasets where strong attribute bias can occur, and\nadvocating the need for methods that can overcome this limitation.\n", "date": "2023-10-08", "categories": []}
{"id": "2311.07141", "title": "SABAF: Removing Strong Attribute Bias from Neural Networks with\n  Adversarial Filtering", "abstract": "  Ensuring a neural network is not relying on protected attributes (e.g., race,\nsex, age) for prediction is crucial in advancing fair and trustworthy AI. While\nseveral promising methods for removing attribute bias in neural networks have\nbeen proposed, their limitations remain under-explored. To that end, in this\nwork, we mathematically and empirically reveal the limitation of existing\nattribute bias removal methods in presence of strong bias and propose a new\nmethod that can mitigate this limitation. Specifically, we first derive a\ngeneral non-vacuous information-theoretical upper bound on the performance of\nany attribute bias removal method in terms of the bias strength, revealing that\nthey are effective only when the inherent bias in the dataset is relatively\nweak. Next, we derive a necessary condition for the existence of any method\nthat can remove attribute bias regardless of the bias strength. Inspired by\nthis condition, we then propose a new method using an adversarial objective\nthat directly filters out protected attributes in the input space while\nmaximally preserving all other attributes, without requiring any specific\ntarget label. The proposed method achieves state-of-the-art performance in both\nstrong and moderate bias settings. We provide extensive experiments on\nsynthetic, image, and census datasets, to verify the derived theoretical bound\nand its consequences in practice, and evaluate the effectiveness of the\nproposed method in removing strong attribute bias.\n", "date": "2023-11-13", "categories": []}
{"id": "2311.08472", "title": "Selecting Shots for Demographic Fairness in Few-Shot Learning with Large\n  Language Models", "abstract": "  Recently, work in NLP has shifted to few-shot (in-context) learning, with\nlarge language models (LLMs) performing well across a range of tasks. However,\nwhile fairness evaluations have become a standard for supervised methods,\nlittle is known about the fairness of LLMs as prediction systems. Further,\ncommon standard methods for fairness involve access to models weights or are\napplied during finetuning, which are not applicable in few-shot learning. Do\nLLMs exhibit prediction biases when used for standard NLP tasks? In this work,\nwe explore the effect of shots, which directly affect the performance of\nmodels, on the fairness of LLMs as NLP classification systems. We consider how\ndifferent shot selection strategies, both existing and new demographically\nsensitive methods, affect model fairness across three standard fairness\ndatasets. We discuss how future work can include LLM fairness evaluations.\n", "date": "2023-11-14", "categories": []}
{"id": "2311.06513", "title": "Step by Step to Fairness: Attributing Societal Bias in Task-oriented\n  Dialogue Systems", "abstract": "  Recent works have shown considerable improvements in task-oriented dialogue\n(TOD) systems by utilizing pretrained large language models (LLMs) in an\nend-to-end manner. However, the biased behavior of each component in a TOD\nsystem and the error propagation issue in the end-to-end framework can lead to\nseriously biased TOD responses. Existing works of fairness only focus on the\ntotal bias of a system. In this paper, we propose a diagnosis method to\nattribute bias to each component of a TOD system. With the proposed attribution\nmethod, we can gain a deeper understanding of the sources of bias.\nAdditionally, researchers can mitigate biased model behavior at a more granular\nlevel. We conduct experiments to attribute the TOD system's bias toward three\ndemographic axes: gender, age, and race. Experimental results show that the\nbias of a TOD system usually comes from the response generation model.\n", "date": "2023-11-11", "categories": []}
{"id": "2304.03738", "title": "Should ChatGPT be Biased? Challenges and Risks of Bias in Large Language\n  Models", "abstract": "  As the capabilities of generative language models continue to advance, the\nimplications of biases ingrained within these models have garnered increasing\nattention from researchers, practitioners, and the broader public. This article\ninvestigates the challenges and risks associated with biases in large-scale\nlanguage models like ChatGPT. We discuss the origins of biases, stemming from,\namong others, the nature of training data, model specifications, algorithmic\nconstraints, product design, and policy decisions. We explore the ethical\nconcerns arising from the unintended consequences of biased model outputs. We\nfurther analyze the potential opportunities to mitigate biases, the\ninevitability of some biases, and the implications of deploying these models in\nvarious applications, such as virtual assistants, content generation, and\nchatbots. Finally, we review the current approaches to identify, quantify, and\nmitigate biases in language models, emphasizing the need for a\nmulti-disciplinary, collaborative effort to develop more equitable,\ntransparent, and responsible AI systems. This article aims to stimulate a\nthoughtful dialogue within the artificial intelligence community, encouraging\nresearchers and developers to reflect on the role of biases in generative\nlanguage models and the ongoing pursuit of ethical AI.\n", "date": "2023-04-07", "categories": []}
{"id": "2310.18458", "title": "Do Not Harm Protected Groups in Debiasing Language Representation Models", "abstract": "  Language Representation Models (LRMs) trained with real-world data may\ncapture and exacerbate undesired bias and cause unfair treatment of people in\nvarious demographic groups. Several techniques have been investigated for\napplying interventions to LRMs to remove bias in benchmark evaluations on, for\nexample, word embeddings. However, the negative side effects of debiasing\ninterventions are usually not revealed in the downstream tasks. We propose\nxGAP-DEBIAS, a set of evaluations on assessing the fairness of debiasing. In\nthis work, We examine four debiasing techniques on a real-world text\nclassification task and show that reducing biasing is at the cost of degrading\nperformance for all demographic groups, including those the debiasing\ntechniques aim to protect. We advocate that a debiasing technique should have\ngood downstream performance with the constraint of ensuring no harm to the\nprotected group.\n", "date": "2023-10-27", "categories": []}
{"id": "2305.19429", "title": "Adapting Fairness Interventions to Missing Values", "abstract": "  Missing values in real-world data pose a significant and unique challenge to\nalgorithmic fairness. Different demographic groups may be unequally affected by\nmissing data, and the standard procedure for handling missing values where\nfirst data is imputed, then the imputed data is used for classification -- a\nprocedure referred to as \"impute-then-classify\" -- can exacerbate\ndiscrimination. In this paper, we analyze how missing values affect algorithmic\nfairness. We first prove that training a classifier from imputed data can\nsignificantly worsen the achievable values of group fairness and average\naccuracy. This is because imputing data results in the loss of the missing\npattern of the data, which often conveys information about the predictive\nlabel. We present scalable and adaptive algorithms for fair classification with\nmissing values. These algorithms can be combined with any preexisting\nfairness-intervention algorithm to handle all possible missing patterns while\npreserving information encoded within the missing patterns. Numerical\nexperiments with state-of-the-art fairness interventions demonstrate that our\nadaptive algorithms consistently achieve higher fairness and accuracy than\nimpute-then-classify across different datasets.\n", "date": "2023-05-30", "categories": []}
{"id": "2311.05746", "title": "Bridging the Digital Divide: Performance Variation across Socio-Economic\n  Factors in Vision-Language Models", "abstract": "  Despite the impressive performance of current AI models reported across\nvarious tasks, performance reports often do not include evaluations of how\nthese models perform on the specific groups that will be impacted by these\ntechnologies. Among the minority groups under-represented in AI, data from\nlow-income households are often overlooked in data collection and model\nevaluation. We evaluate the performance of a state-of-the-art vision-language\nmodel (CLIP) on a geo-diverse dataset containing household images associated\nwith different income values (Dollar Street) and show that performance\ninequality exists among households of different income levels. Our results\nindicate that performance for the poorer groups is consistently lower than the\nwealthier groups across various topics and countries. We highlight insights\nthat can help mitigate these issues and propose actionable steps for\neconomic-level inclusive AI development. Code is available at\nhttps://github.com/MichiganNLP/Bridging_the_Digital_Divide.\n", "date": "2023-11-09", "categories": []}
{"id": "2311.06175", "title": "Search-Based Fairness Testing: An Overview", "abstract": "  Artificial Intelligence (AI) has demonstrated remarkable capabilities in\ndomains such as recruitment, finance, healthcare, and the judiciary. However,\nbiases in AI systems raise ethical and societal concerns, emphasizing the need\nfor effective fairness testing methods. This paper reviews current research on\nfairness testing, particularly its application through search-based testing.\nOur analysis highlights progress and identifies areas of improvement in\naddressing AI systems biases. Future research should focus on leveraging\nestablished search-based testing methodologies for fairness testing.\n", "date": "2023-11-10", "categories": []}
{"id": "2310.12421", "title": "Detecting and Mitigating Algorithmic Bias in Binary Classification using\n  Causal Modeling", "abstract": "  This paper proposes the use of causal modeling to detect and mitigate\nalgorithmic bias. We provide a brief description of causal modeling and a\ngeneral overview of our approach. We then use the Adult dataset, which is\navailable for download from the UC Irvine Machine Learning Repository, to\ndevelop (1) a prediction model, which is treated as a black box, and (2) a\ncausal model for bias mitigation. In this paper, we focus on gender bias and\nthe problem of binary classification. We show that gender bias in the\nprediction model is statistically significant at the 0.05 level. We demonstrate\nthe effectiveness of the causal model in mitigating gender bias by\ncross-validation. Furthermore, we show that the overall classification accuracy\nis improved slightly. Our novel approach is intuitive, easy-to-use, and can be\nimplemented using existing statistical software tools such as \"lavaan\" in R.\nHence, it enhances explainability and promotes trust.\n", "date": "2023-10-19", "categories": []}
{"id": "2311.05420", "title": "Counterfactually Fair Representation", "abstract": "  The use of machine learning models in high-stake applications (e.g.,\nhealthcare, lending, college admission) has raised growing concerns due to\npotential biases against protected social groups. Various fairness notions and\nmethods have been proposed to mitigate such biases. In this work, we focus on\nCounterfactual Fairness (CF), a fairness notion that is dependent on an\nunderlying causal graph and first proposed by Kusner \\textit{et\nal.}~\\cite{kusner2017counterfactual}; it requires that the outcome an\nindividual perceives is the same in the real world as it would be in a\n\"counterfactual\" world, in which the individual belongs to another social\ngroup. Learning fair models satisfying CF can be challenging. It was shown in\n\\cite{kusner2017counterfactual} that a sufficient condition for satisfying CF\nis to \\textbf{not} use features that are descendants of sensitive attributes in\nthe causal graph. This implies a simple method that learns CF models only using\nnon-descendants of sensitive attributes while eliminating all descendants.\nAlthough several subsequent works proposed methods that use all features for\ntraining CF models, there is no theoretical guarantee that they can satisfy CF.\nIn contrast, this work proposes a new algorithm that trains models using all\nthe available features. We theoretically and empirically show that models\ntrained with this method can satisfy CF\\footnote{The code repository for this\nwork can be found in\n\\url{https://github.com/osu-srml/CF_Representation_Learning}}.\n", "date": "2023-11-09", "categories": []}
{"id": "2311.05451", "title": "All Should Be Equal in the Eyes of Language Models: Counterfactually\n  Aware Fair Text Generation", "abstract": "  Fairness in Language Models (LMs) remains a longstanding challenge, given the\ninherent biases in training data that can be perpetuated by models and affect\nthe downstream tasks. Recent methods employ expensive retraining or attempt\ndebiasing during inference by constraining model outputs to contrast from a\nreference set of biased templates or exemplars. Regardless, they dont address\nthe primary goal of fairness to maintain equitability across different\ndemographic groups. In this work, we posit that inferencing LMs to generate\nunbiased output for one demographic under a context ensues from being aware of\noutputs for other demographics under the same context. To this end, we propose\nCounterfactually Aware Fair InferencE (CAFIE), a framework that dynamically\ncompares the model understanding of diverse demographics to generate more\nequitable sentences. We conduct an extensive empirical evaluation using base\nLMs of varying sizes and across three diverse datasets and found that CAFIE\noutperforms strong baselines. CAFIE produces fairer text and strikes the best\nbalance between fairness and language modeling capability\n", "date": "2023-11-09", "categories": []}
{"id": "2305.12495", "title": "Fair Without Leveling Down: A New Intersectional Fairness Definition", "abstract": "  In this work, we consider the problem of intersectional group fairness in the\nclassification setting, where the objective is to learn discrimination-free\nmodels in the presence of several intersecting sensitive groups. First, we\nillustrate various shortcomings of existing fairness measures commonly used to\ncapture intersectional fairness. Then, we propose a new definition called the\n$\\alpha$-Intersectional Fairness, which combines the absolute and the relative\nperformance across sensitive groups and can be seen as a generalization of the\nnotion of differential fairness. We highlight several desirable properties of\nthe proposed definition and analyze its relation to other fairness measures.\nFinally, we benchmark multiple popular in-processing fair machine learning\napproaches using our new fairness definition and show that they do not achieve\nany improvement over a simple baseline. Our results reveal that the increase in\nfairness measured by previous definitions hides a \"leveling down\" effect, i.e.,\ndegrading the best performance over groups rather than improving the worst one.\n", "date": "2023-05-21", "categories": []}
{"id": "2306.16635", "title": "Improving Fairness in Deepfake Detection", "abstract": "  Despite the development of effective deepfake detectors in recent years,\nrecent studies have demonstrated that biases in the data used to train these\ndetectors can lead to disparities in detection accuracy across different races\nand genders. This can result in different groups being unfairly targeted or\nexcluded from detection, allowing undetected deepfakes to manipulate public\nopinion and erode trust in a deepfake detection model. While existing studies\nhave focused on evaluating fairness of deepfake detectors, to the best of our\nknowledge, no method has been developed to encourage fairness in deepfake\ndetection at the algorithm level. In this work, we make the first attempt to\nimprove deepfake detection fairness by proposing novel loss functions that\nhandle both the setting where demographic information (eg, annotations of race\nand gender) is available as well as the case where this information is absent.\nFundamentally, both approaches can be used to convert many existing deepfake\ndetectors into ones that encourages fairness. Extensive experiments on four\ndeepfake datasets and five deepfake detectors demonstrate the effectiveness and\nflexibility of our approach in improving deepfake detection fairness. Our code\nis available at https://github.com/littlejuyan/DF_Fairness.\n", "date": "2023-06-29", "categories": []}
{"id": "2307.05857", "title": "FAIRO: Fairness-aware Adaptation in Sequential-Decision Making for\n  Human-in-the-Loop Systems", "abstract": "  Achieving fairness in sequential-decision making systems within\nHuman-in-the-Loop (HITL) environments is a critical concern, especially when\nmultiple humans with different behavior and expectations are affected by the\nsame adaptation decisions in the system. This human variability factor adds\nmore complexity since policies deemed fair at one point in time may become\ndiscriminatory over time due to variations in human preferences resulting from\ninter- and intra-human variability. This paper addresses the fairness problem\nfrom an equity lens, considering human behavior variability, and the changes in\nhuman preferences over time. We propose FAIRO, a novel algorithm for\nfairness-aware sequential-decision making in HITL adaptation, which\nincorporates these notions into the decision-making process. In particular,\nFAIRO decomposes this complex fairness task into adaptive sub-tasks based on\nindividual human preferences through leveraging the Options reinforcement\nlearning framework. We design FAIRO to generalize to three types of HITL\napplication setups that have the shared adaptation decision problem.\nFurthermore, we recognize that fairness-aware policies can sometimes conflict\nwith the application's utility. To address this challenge, we provide a\nfairness-utility tradeoff in FAIRO, allowing system designers to balance the\nobjectives of fairness and utility based on specific application requirements.\nExtensive evaluations of FAIRO on the three HITL applications demonstrate its\ngeneralizability and effectiveness in promoting fairness while accounting for\nhuman variability. On average, FAIRO can improve fairness compared with other\nmethods across all three applications by 35.36%.\n", "date": "2023-07-12", "categories": []}
{"id": "2311.03425", "title": "An AI-Guided Data Centric Strategy to Detect and Mitigate Biases in\n  Healthcare Datasets", "abstract": "  The adoption of diagnosis and prognostic algorithms in healthcare has led to\nconcerns about the perpetuation of bias against disadvantaged groups of\nindividuals. Deep learning methods to detect and mitigate bias have revolved\naround modifying models, optimization strategies, and threshold calibration\nwith varying levels of success. Here, we generate a data-centric,\nmodel-agnostic, task-agnostic approach to evaluate dataset bias by\ninvestigating the relationship between how easily different groups are learned\nat small sample sizes (AEquity). We then apply a systematic analysis of AEq\nvalues across subpopulations to identify and mitigate manifestations of racial\nbias in two known cases in healthcare - Chest X-rays diagnosis with deep\nconvolutional neural networks and healthcare utilization prediction with\nmultivariate logistic regression. AEq is a novel and broadly applicable metric\nthat can be applied to advance equity by diagnosing and remediating bias in\nhealthcare datasets.\n", "date": "2023-11-06", "categories": []}
{"id": "2311.03468", "title": "FinA: Fairness of Adverse Effects in Decision-Making of\n  Human-Cyber-Physical-System", "abstract": "  Ensuring fairness in decision-making systems within\nHuman-Cyber-Physical-Systems (HCPS) is a pressing concern, particularly when\ndiverse individuals, each with varying behaviors and expectations, coexist\nwithin the same application space, influenced by a shared set of control\nactions in the system. The long-term adverse effects of these actions further\npose the challenge, as historical experiences and interactions shape individual\nperceptions of fairness. This paper addresses the challenge of fairness from an\nequity perspective of adverse effects, taking into account the dynamic nature\nof human behavior and evolving preferences while recognizing the lasting impact\nof adverse effects. We formally introduce the concept of\nFairness-in-Adverse-Effects (FinA) within the HCPS context. We put forth a\ncomprehensive set of five formulations for FinA, encompassing both the\ninstantaneous and long-term aspects of adverse effects. To empirically validate\nthe effectiveness of our FinA approach, we conducted an evaluation within the\ndomain of smart homes, a pertinent HCPS application. The outcomes of our\nevaluation demonstrate that the adoption of FinA significantly enhances the\noverall perception of fairness among individuals, yielding an average\nimprovement of 66.7% when compared to the state-of-the-art method.\n", "date": "2023-11-06", "categories": []}
{"id": "2311.03532", "title": "The Fairness Stitch: Unveiling the Potential of Model Stitching in\n  Neural Network De-Biasing", "abstract": "  The pursuit of fairness in machine learning models has emerged as a critical\nresearch challenge in different applications ranging from bank loan approval to\nface detection. Despite the widespread adoption of artificial intelligence\nalgorithms across various domains, concerns persist regarding the presence of\nbiases and discrimination within these models. To address this pressing issue,\nthis study introduces a novel method called \"The Fairness Stitch (TFS)\" to\nenhance fairness in deep learning models. This method combines model stitching\nand training jointly, while incorporating fairness constraints. In this\nresearch, we assess the effectiveness of our proposed method by conducting a\ncomprehensive evaluation of two well-known datasets, CelebA and UTKFace. We\nsystematically compare the performance of our approach with the existing\nbaseline method. Our findings reveal a notable improvement in achieving a\nbalanced trade-off between fairness and performance, highlighting the promising\npotential of our method to address bias-related challenges and foster equitable\noutcomes in machine learning models. This paper poses a challenge to the\nconventional wisdom of the effectiveness of the last layer in deep learning\nmodels for de-biasing.\n", "date": "2023-11-06", "categories": []}
{"id": "2311.03714", "title": "Loss Balancing for Fair Supervised Learning", "abstract": "  Supervised learning models have been used in various domains such as lending,\ncollege admission, face recognition, natural language processing, etc. However,\nthey may inherit pre-existing biases from training data and exhibit\ndiscrimination against protected social groups. Various fairness notions have\nbeen proposed to address unfairness issues. In this work, we focus on Equalized\nLoss (EL), a fairness notion that requires the expected loss to be\n(approximately) equalized across different groups. Imposing EL on the learning\nprocess leads to a non-convex optimization problem even if the loss function is\nconvex, and the existing fair learning algorithms cannot properly be adopted to\nfind the fair predictor under the EL constraint. This paper introduces an\nalgorithm that can leverage off-the-shelf convex programming tools (e.g.,\nCVXPY) to efficiently find the global optimum of this non-convex optimization.\nIn particular, we propose the ELminimizer algorithm, which finds the optimal\nfair predictor under EL by reducing the non-convex optimization to a sequence\nof convex optimization problems. We theoretically prove that our algorithm\nfinds the global optimal solution under certain conditions. Then, we support\nour theoretical results through several empirical studies.\n", "date": "2023-11-07", "categories": []}
{"id": "2311.02243", "title": "Equal Opportunity of Coverage in Fair Regression", "abstract": "  We study fair machine learning (ML) under predictive uncertainty to enable\nreliable and trustworthy decision-making. The seminal work of ``equalized\ncoverage'' proposed an uncertainty-aware fairness notion. However, it does not\nguarantee equal coverage rates across more fine-grained groups (e.g.,\nlow-income females) conditioning on the true label and is biased in the\nassessment of uncertainty. To tackle these limitations, we propose a new\nuncertainty-aware fairness -- Equal Opportunity of Coverage (EOC) -- that aims\nto achieve two properties: (1) coverage rates for different groups with similar\noutcomes are close, and (2) the coverage rate for the entire population remains\nat a predetermined level. Further, the prediction intervals should be narrow to\nbe informative. We propose Binned Fair Quantile Regression (BFQR), a\ndistribution-free post-processing method to improve EOC with reasonable width\nfor any trained ML models. It first calibrates a hold-out set to bound\ndeviation from EOC, then leverages conformal prediction to maintain EOC on a\ntest set, meanwhile optimizing prediction interval width. Experimental results\ndemonstrate the effectiveness of our method in improving EOC. Our code is\npublicly available at https://github.com/fangxin-wang/bfqr .\n", "date": "2023-11-03", "categories": []}
{"id": "2311.03000", "title": "Strong statistical parity through fair synthetic data", "abstract": "  AI-generated synthetic data, in addition to protecting the privacy of\noriginal data sets, allows users and data consumers to tailor data to their\nneeds. This paper explores the creation of synthetic data that embodies\nFairness by Design, focusing on the statistical parity fairness definition. By\nequalizing the learned target probability distributions of the synthetic data\ngenerator across sensitive attributes, a downstream model trained on such\nsynthetic data provides fair predictions across all thresholds, that is, strong\nfair predictions even when inferring from biased, original data. This fairness\nadjustment can be either directly integrated into the sampling process of a\nsynthetic generator or added as a post-processing step. The flexibility allows\ndata consumers to create fair synthetic data and fine-tune the trade-off\nbetween accuracy and fairness without any previous assumptions on the data or\nre-training the synthetic data generator.\n", "date": "2023-11-06", "categories": []}
{"id": "2311.03311", "title": "Unraveling Downstream Gender Bias from Large Language Models: A Study on\n  AI Educational Writing Assistance", "abstract": "  Large Language Models (LLMs) are increasingly utilized in educational tasks\nsuch as providing writing suggestions to students. Despite their potential,\nLLMs are known to harbor inherent biases which may negatively impact learners.\nPrevious studies have investigated bias in models and data representations\nseparately, neglecting the potential impact of LLM bias on human writing. In\nthis paper, we investigate how bias transfers through an AI writing support\npipeline. We conduct a large-scale user study with 231 students writing\nbusiness case peer reviews in German. Students are divided into five groups\nwith different levels of writing support: one classroom group with\nfeature-based suggestions and four groups recruited from Prolific -- a control\ngroup with no assistance, two groups with suggestions from fine-tuned GPT-2 and\nGPT-3 models, and one group with suggestions from pre-trained GPT-3.5. Using\nGenBit gender bias analysis, Word Embedding Association Tests (WEAT), and\nSentence Embedding Association Test (SEAT) we evaluate the gender bias at\nvarious stages of the pipeline: in model embeddings, in suggestions generated\nby the models, and in reviews written by students. Our results demonstrate that\nthere is no significant difference in gender bias between the resulting peer\nreviews of groups with and without LLM suggestions. Our research is therefore\noptimistic about the use of AI writing support in the classroom, showcasing a\ncontext where bias in LLMs does not transfer to students' responses.\n", "date": "2023-11-06", "categories": []}
{"id": "2310.12560", "title": "Fast Model Debias with Machine Unlearning", "abstract": "  Recent discoveries have revealed that deep neural networks might behave in a\nbiased manner in many real-world scenarios. For instance, deep networks trained\non a large-scale face recognition dataset CelebA tend to predict blonde hair\nfor females and black hair for males. Such biases not only jeopardize the\nrobustness of models but also perpetuate and amplify social biases, which is\nespecially concerning for automated decision-making processes in healthcare,\nrecruitment, etc., as they could exacerbate unfair economic and social\ninequalities among different groups. Existing debiasing methods suffer from\nhigh costs in bias labeling or model re-training, while also exhibiting a\ndeficiency in terms of elucidating the origins of biases within the model. To\nthis respect, we propose a fast model debiasing framework (FMD) which offers an\nefficient approach to identify, evaluate and remove biases inherent in trained\nmodels. The FMD identifies biased attributes through an explicit counterfactual\nconcept and quantifies the influence of data samples with influence functions.\nMoreover, we design a machine unlearning-based strategy to efficiently and\neffectively remove the bias in a trained model with a small counterfactual\ndataset. Experiments on the Colored MNIST, CelebA, and Adult Income datasets\nalong with experiments with large language models demonstrate that our method\nachieves superior or competing accuracies compared with state-of-the-art\nmethods while attaining significantly fewer biases and requiring much less\ndebiasing cost. Notably, our method requires only a small external dataset and\nupdating a minimal amount of model parameters, without the requirement of\naccess to training data that may be too large or unavailable in practice.\n", "date": "2023-10-19", "categories": []}
{"id": "2311.01573", "title": "Improving Fairness using Vision-Language Driven Image Augmentation", "abstract": "  Fairness is crucial when training a deep-learning discriminative model,\nespecially in the facial domain. Models tend to correlate specific\ncharacteristics (such as age and skin color) with unrelated attributes\n(downstream tasks), resulting in biases which do not correspond to reality. It\nis common knowledge that these correlations are present in the data and are\nthen transferred to the models during training. This paper proposes a method to\nmitigate these correlations to improve fairness. To do so, we learn\ninterpretable and meaningful paths lying in the semantic space of a pre-trained\ndiffusion model (DiffAE) -- such paths being supervised by contrastive text\ndipoles. That is, we learn to edit protected characteristics (age and skin\ncolor). These paths are then applied to augment images to improve the fairness\nof a given dataset. We test the proposed method on CelebA-HQ and UTKFace on\nseveral downstream tasks with age and skin color as protected characteristics.\nAs a proxy for fairness, we compute the difference in accuracy with respect to\nthe protected characteristics. Quantitative results show how the augmented\nimages help the model improve the overall accuracy, the aforementioned metric,\nand the disparity of equal opportunity. Code is available at:\nhttps://github.com/Moreno98/Vision-Language-Bias-Control.\n", "date": "2023-11-02", "categories": []}
{"id": "2209.08762", "title": "Fairness on Synthetic Visual and Thermal Mask Images", "abstract": "  In this paper, we study performance and fairness on visual and thermal images\nand expand the assessment to masked synthetic images. Using the SpeakingFace\nand Thermal-Mask dataset, we propose a process to assess fairness on real\nimages and show how the same process can be applied to synthetic images. The\nresulting process shows a demographic parity difference of 1.59 for random\nguessing and increases to 5.0 when the recognition performance increases to a\nprecision and recall rate of 99.99\\%. We indicate that inherently biased\ndatasets can deeply impact the fairness of any biometric system. A primary\ncause of a biased dataset is the class imbalance due to the data collection\nprocess. To address imbalanced datasets, the classes with fewer samples can be\naugmented with synthetic images to generate a more balanced dataset resulting\nin less bias when training a machine learning system. For biometric-enabled\nsystems, fairness is of critical importance, while the related concept of\nEquity, Diversity, and Inclusion (EDI) is well suited for the generalization of\nfairness in biometrics, in this paper, we focus on the 3 most common\ndemographic groups age, gender, and ethnicity.\n", "date": "2022-09-19", "categories": []}
{"id": "2311.00638", "title": "FAIRLABEL: Correcting Bias in Labels", "abstract": "  There are several algorithms for measuring fairness of ML models. A\nfundamental assumption in these approaches is that the ground truth is fair or\nunbiased. In real-world datasets, however, the ground truth often contains data\nthat is a result of historical and societal biases and discrimination. Models\ntrained on these datasets will inherit and propagate the biases to the model\noutputs. We propose FAIRLABEL, an algorithm which detects and corrects biases\nin labels. The goal of FAIRLABELis to reduce the Disparate Impact (DI) across\ngroups while maintaining high accuracy in predictions. We propose metrics to\nmeasure the quality of bias correction and validate FAIRLABEL on synthetic\ndatasets and show that the label correction is correct 86.7% of the time vs.\n71.9% for a baseline model. We also apply FAIRLABEL on benchmark datasets such\nas UCI Adult, German Credit Risk, and Compas datasets and show that the\nDisparate Impact Ratio increases by as much as 54.2%.\n", "date": "2023-11-01", "categories": []}
{"id": "2212.02648", "title": "Spuriosity Rankings: Sorting Data to Measure and Mitigate Biases", "abstract": "  We present a simple but effective method to measure and mitigate model biases\ncaused by reliance on spurious cues. Instead of requiring costly changes to\none's data or model training, our method better utilizes the data one already\nhas by sorting them. Specifically, we rank images within their classes based on\nspuriosity (the degree to which common spurious cues are present), proxied via\ndeep neural features of an interpretable network. With spuriosity rankings, it\nis easy to identify minority subpopulations (i.e. low spuriosity images) and\nassess model bias as the gap in accuracy between high and low spuriosity\nimages. One can even efficiently remove a model's bias at little cost to\naccuracy by finetuning its classification head on low spuriosity images,\nresulting in fairer treatment of samples regardless of spuriosity. We\ndemonstrate our method on ImageNet, annotating $5000$ class-feature\ndependencies ($630$ of which we find to be spurious) and generating a dataset\nof $325k$ soft segmentations for these features along the way. Having computed\nspuriosity rankings via the identified spurious neural features, we assess\nbiases for $89$ diverse models and find that class-wise biases are highly\ncorrelated across models. Our results suggest that model bias due to spurious\nfeature reliance is influenced far more by what the model is trained on than\nhow it is trained.\n", "date": "2022-12-05", "categories": []}
{"id": "2308.11090", "title": "Fairness Explainability using Optimal Transport with Applications in\n  Image Classification", "abstract": "  Ensuring trust and accountability in Artificial Intelligence systems demands\nexplainability of its outcomes. Despite significant progress in Explainable AI,\nhuman biases still taint a substantial portion of its training data, raising\nconcerns about unfairness or discriminatory tendencies. Current approaches in\nthe field of Algorithmic Fairness focus on mitigating such biases in the\noutcomes of a model, but few attempts have been made to try to explain\n\\emph{why} a model is biased. To bridge this gap between the two fields, we\npropose a comprehensive approach that uses optimal transport theory to uncover\nthe causes of discrimination in Machine Learning applications, with a\nparticular emphasis on image classification. We leverage Wasserstein\nbarycenters to achieve fair predictions and introduce an extension to pinpoint\nbias-associated regions. This allows us to derive a cohesive system which uses\nthe enforced fairness to measure each features influence \\emph{on} the bias.\nTaking advantage of this interplay of enforcing and explaining fairness, our\nmethod hold significant implications for the development of trustworthy and\nunbiased AI systems, fostering transparency, accountability, and fairness in\ncritical decision-making scenarios across diverse domains.\n", "date": "2023-08-22", "categories": []}
{"id": "2310.19986", "title": "Addressing Weak Decision Boundaries in Image Classification by\n  Leveraging Web Search and Generative Models", "abstract": "  Machine learning (ML) technologies are known to be riddled with ethical and\noperational problems, however, we are witnessing an increasing thrust by\nbusinesses to deploy them in sensitive applications. One major issue among many\nis that ML models do not perform equally well for underrepresented groups. This\nputs vulnerable populations in an even disadvantaged and unfavorable position.\nWe propose an approach that leverages the power of web search and generative\nmodels to alleviate some of the shortcomings of discriminative models. We\ndemonstrate our method on an image classification problem using ImageNet's\nPeople Subtree subset, and show that it is effective in enhancing robustness\nand mitigating bias in certain classes that represent vulnerable populations\n(e.g., female doctor of color). Our new method is able to (1) identify weak\ndecision boundaries for such classes; (2) construct search queries for Google\nas well as text for generating images through DALL-E 2 and Stable Diffusion;\nand (3) show how these newly captured training samples could alleviate\npopulation bias issue. While still improving the model's overall performance\nconsiderably, we achieve a significant reduction (77.30\\%) in the model's\ngender accuracy disparity. In addition to these improvements, we observed a\nnotable enhancement in the classifier's decision boundary, as it is\ncharacterized by fewer weakspots and an increased separation between classes.\nAlthough we showcase our method on vulnerable populations in this study, the\nproposed technique is extendable to a wide range of problems and domains.\n", "date": "2023-10-30", "categories": []}
{"id": "2310.20508", "title": "Parametric Fairness with Statistical Guarantees", "abstract": "  Algorithmic fairness has gained prominence due to societal and regulatory\nconcerns about biases in Machine Learning models. Common group fairness metrics\nlike Equalized Odds for classification or Demographic Parity for both\nclassification and regression are widely used and a host of computationally\nadvantageous post-processing methods have been developed around them. However,\nthese metrics often limit users from incorporating domain knowledge. Despite\nmeeting traditional fairness criteria, they can obscure issues related to\nintersectional fairness and even replicate unwanted intra-group biases in the\nresulting fair solution. To avoid this narrow perspective, we extend the\nconcept of Demographic Parity to incorporate distributional properties in the\npredictions, allowing expert knowledge to be used in the fair solution. We\nillustrate the use of this new metric through a practical example of wages, and\ndevelop a parametric method that efficiently addresses practical challenges\nlike limited training data and constraints on total spending, offering a robust\nsolution for real-life applications.\n", "date": "2023-10-31", "categories": []}
{"id": "2207.03444", "title": "Fairness and Bias in Robot Learning", "abstract": "  Machine learning has significantly enhanced the abilities of robots, enabling\nthem to perform a wide range of tasks in human environments and adapt to our\nuncertain real world. Recent works in various machine learning domains have\nhighlighted the importance of accounting for fairness to ensure that these\nalgorithms do not reproduce human biases and consequently lead to\ndiscriminatory outcomes. With robot learning systems increasingly performing\nmore and more tasks in our everyday lives, it is crucial to understand the\ninfluence of such biases to prevent unintended behavior toward certain groups\nof people. In this work, we present the first survey on fairness in robot\nlearning from an interdisciplinary perspective spanning technical, ethical, and\nlegal challenges. We propose a taxonomy for sources of bias and the resulting\ntypes of discrimination due to them. Using examples from different robot\nlearning domains, we examine scenarios of unfair outcomes and strategies to\nmitigate them. We present early advances in the field by covering different\nfairness definitions, ethical and legal considerations, and methods for fair\nrobot learning. With this work, we aim to pave the road for groundbreaking\ndevelopments in fair robot learning.\n", "date": "2022-07-07", "categories": []}
{"id": "2306.13675", "title": "Intersectionality and Testimonial Injustice in Medical Records", "abstract": "  Detecting testimonial injustice is an essential element of addressing\ninequities and promoting inclusive healthcare practices, many of which are\nlife-critical. However, using a single demographic factor to detect testimonial\ninjustice does not fully encompass the nuanced identities that contribute to a\npatient's experience. Further, some injustices may only be evident when\nexamining the nuances that arise through the lens of intersectionality.\nIgnoring such injustices can result in poor quality of care or life-endangering\nevents. Thus, considering intersectionality could result in more accurate\nclassifications and just decisions. To illustrate this, we use real-world\nmedical data to determine whether medical records exhibit words that could lead\nto testimonial injustice, employ fairness metrics (e.g. demographic parity,\ndifferential intersectional fairness, and subgroup fairness) to assess the\nseverity to which subgroups are experiencing testimonial injustice, and analyze\nhow the intersectionality of demographic features (e.g. gender and race) make a\ndifference in uncovering testimonial injustice. From our analysis, we found\nthat with intersectionality we can better see disparities in how subgroups are\ntreated and there are differences in how someone is treated based on the\nintersection of their demographic attributes. This has not been previously\nstudied in clinical records, nor has it been proven through empirical study.\n", "date": "2023-06-20", "categories": []}
{"id": "2307.01166", "title": "Strategic Distribution Shift of Interacting Agents via Coupled Gradient\n  Flows", "abstract": "  We propose a novel framework for analyzing the dynamics of distribution shift\nin real-world systems that captures the feedback loop between learning\nalgorithms and the distributions on which they are deployed. Prior work largely\nmodels feedback-induced distribution shift as adversarial or via an overly\nsimplistic distribution-shift structure. In contrast, we propose a coupled\npartial differential equation model that captures fine-grained changes in the\ndistribution over time by accounting for complex dynamics that arise due to\nstrategic responses to algorithmic decision-making, non-local endogenous\npopulation interactions, and other exogenous sources of distribution shift. We\nconsider two common settings in machine learning: cooperative settings with\ninformation asymmetries, and competitive settings where a learner faces\nstrategic users. For both of these settings, when the algorithm retrains via\ngradient descent, we prove asymptotic convergence of the retraining procedure\nto a steady-state, both in finite and in infinite dimensions, obtaining\nexplicit rates in terms of the model parameters. To do so we derive new results\non the convergence of coupled PDEs that extends what is known on multi-species\nsystems. Empirically, we show that our approach captures well-documented forms\nof distribution shifts like polarization and disparate impacts that simpler\nmodels cannot capture.\n", "date": "2023-07-03", "categories": []}
{"id": "2310.18413", "title": "On the Fairness ROAD: Robust Optimization for Adversarial Debiasing", "abstract": "  In the field of algorithmic fairness, significant attention has been put on\ngroup fairness criteria, such as Demographic Parity and Equalized Odds.\nNevertheless, these objectives, measured as global averages, have raised\nconcerns about persistent local disparities between sensitive groups. In this\nwork, we address the problem of local fairness, which ensures that the\npredictor is unbiased not only in terms of expectations over the whole\npopulation, but also within any subregion of the feature space, unknown at\ntraining time. To enforce this objective, we introduce ROAD, a novel approach\nthat leverages the Distributionally Robust Optimization (DRO) framework within\na fair adversarial learning objective, where an adversary tries to infer the\nsensitive attribute from the predictions. Using an instance-level re-weighting\nstrategy, ROAD is designed to prioritize inputs that are likely to be locally\nunfair, i.e. where the adversary faces the least difficulty in reconstructing\nthe sensitive attribute. Numerical experiments demonstrate the effectiveness of\nour method: it achieves Pareto dominance with respect to local fairness and\naccuracy for a given global fairness level across three standard datasets, and\nalso enhances fairness generalization under distribution shift.\n", "date": "2023-10-27", "categories": []}
{"id": "2310.18555", "title": "Group Robust Classification Without Any Group Information", "abstract": "  Empirical risk minimization (ERM) is sensitive to spurious correlations in\nthe training data, which poses a significant risk when deploying systems\ntrained under this paradigm in high-stake applications. While the existing\nliterature focuses on maximizing group-balanced or worst-group accuracy,\nestimating these accuracies is hindered by costly bias annotations. This study\ncontends that current bias-unsupervised approaches to group robustness continue\nto rely on group information to achieve optimal performance. Firstly, these\nmethods implicitly assume that all group combinations are represented during\ntraining. To illustrate this, we introduce a systematic generalization task on\nthe MPI3D dataset and discover that current algorithms fail to improve the ERM\nbaseline when combinations of observed attribute values are missing. Secondly,\nbias labels are still crucial for effective model selection, restricting the\npracticality of these methods in real-world scenarios. To address these\nlimitations, we propose a revised methodology for training and validating\ndebiased models in an entirely bias-unsupervised manner. We achieve this by\nemploying pretrained self-supervised models to reliably extract bias\ninformation, which enables the integration of a logit adjustment training loss\nwith our validation criterion. Our empirical analysis on synthetic and\nreal-world tasks provides evidence that our approach overcomes the identified\nchallenges and consistently enhances robust accuracy, attaining performance\nwhich is competitive with or outperforms that of state-of-the-art methods,\nwhich, conversely, rely on bias labels for validation.\n", "date": "2023-10-28", "categories": []}
{"id": "2310.18593", "title": "Fair Streaming Principal Component Analysis: Statistical and Algorithmic\n  Viewpoint", "abstract": "  Fair Principal Component Analysis (PCA) is a problem setting where we aim to\nperform PCA while making the resulting representation fair in that the\nprojected distributions, conditional on the sensitive attributes, match one\nanother. However, existing approaches to fair PCA have two main problems:\ntheoretically, there has been no statistical foundation of fair PCA in terms of\nlearnability; practically, limited memory prevents us from using existing\napproaches, as they explicitly rely on full access to the entire data. On the\ntheoretical side, we rigorously formulate fair PCA using a new notion called\n\\emph{probably approximately fair and optimal} (PAFO) learnability. On the\npractical side, motivated by recent advances in streaming algorithms for\naddressing memory limitation, we propose a new setting called \\emph{fair\nstreaming PCA} along with a memory-efficient algorithm, fair noisy power method\n(FNPM). We then provide its {\\it statistical} guarantee in terms of\nPAFO-learnability, which is the first of its kind in fair PCA literature.\nLastly, we verify the efficacy and memory efficiency of our algorithm on\nreal-world datasets.\n", "date": "2023-10-28", "categories": []}
{"id": "2310.19297", "title": "On Measuring Fairness in Generative Models", "abstract": "  Recently, there has been increased interest in fair generative models. In\nthis work, we conduct, for the first time, an in-depth study on fairness\nmeasurement, a critical component in gauging progress on fair generative\nmodels. We make three contributions. First, we conduct a study that reveals\nthat the existing fairness measurement framework has considerable measurement\nerrors, even when highly accurate sensitive attribute (SA) classifiers are\nused. These findings cast doubts on previously reported fairness improvements.\nSecond, to address this issue, we propose CLassifier Error-Aware Measurement\n(CLEAM), a new framework which uses a statistical model to account for\ninaccuracies in SA classifiers. Our proposed CLEAM reduces measurement errors\nsignificantly, e.g., 4.98% $\\rightarrow$ 0.62% for StyleGAN2 w.r.t. Gender.\nAdditionally, CLEAM achieves this with minimal additional overhead. Third, we\nutilize CLEAM to measure fairness in important text-to-image generator and\nGANs, revealing considerable biases in these models that raise concerns about\ntheir applications. Code and more resources:\nhttps://sutd-visual-computing-group.github.io/CLEAM/.\n", "date": "2023-10-30", "categories": []}
{"id": "2310.19691", "title": "Causal Context Connects Counterfactual Fairness to Robust Prediction and\n  Group Fairness", "abstract": "  Counterfactual fairness requires that a person would have been classified in\nthe same way by an AI or other algorithmic system if they had a different\nprotected class, such as a different race or gender. This is an intuitive\nstandard, as reflected in the U.S. legal system, but its use is limited because\ncounterfactuals cannot be directly observed in real-world data. On the other\nhand, group fairness metrics (e.g., demographic parity or equalized odds) are\nless intuitive but more readily observed. In this paper, we use $\\textit{causal\ncontext}$ to bridge the gaps between counterfactual fairness, robust\nprediction, and group fairness. First, we motivate counterfactual fairness by\nshowing that there is not necessarily a fundamental trade-off between fairness\nand accuracy because, under plausible conditions, the counterfactually fair\npredictor is in fact accuracy-optimal in an unbiased target distribution.\nSecond, we develop a correspondence between the causal graph of the\ndata-generating process and which, if any, group fairness metrics are\nequivalent to counterfactual fairness. Third, we show that in three common\nfairness contexts$\\unicode{x2013}$measurement error, selection on label, and\nselection on predictors$\\unicode{x2013}$counterfactual fairness is equivalent\nto demographic parity, equalized odds, and calibration, respectively.\nCounterfactual fairness can sometimes be tested by measuring relatively simple\ngroup fairness metrics.\n", "date": "2023-10-30", "categories": []}
{"id": "2308.10397", "title": "FairMonitor: A Four-Stage Automatic Framework for Detecting Stereotypes\n  and Biases in Large Language Models", "abstract": "  Detecting stereotypes and biases in Large Language Models (LLMs) can enhance\nfairness and reduce adverse impacts on individuals or groups when these LLMs\nare applied. However, the majority of existing methods focus on measuring the\nmodel's preference towards sentences containing biases and stereotypes within\ndatasets, which lacks interpretability and cannot detect implicit biases and\nstereotypes in the real world. To address this gap, this paper introduces a\nfour-stage framework to directly evaluate stereotypes and biases in the\ngenerated content of LLMs, including direct inquiry testing, serial or adapted\nstory testing, implicit association testing, and unknown situation testing.\nAdditionally, the paper proposes multi-dimensional evaluation metrics and\nexplainable zero-shot prompts for automated evaluation. Using the education\nsector as a case study, we constructed the Edu-FairMonitor based on the\nfour-stage framework, which encompasses 12,632 open-ended questions covering\nnine sensitive factors and 26 educational scenarios. Experimental results\nreveal varying degrees of stereotypes and biases in five LLMs evaluated on\nEdu-FairMonitor. Moreover, the results of our proposed automated evaluation\nmethod have shown a high correlation with human annotations.\n", "date": "2023-08-21", "categories": []}
{"id": "2310.16506", "title": "Identifying Reasons for Bias: An Argumentation-Based Approach", "abstract": "  As algorithmic decision-making systems become more prevalent in society,\nensuring the fairness of these systems is becoming increasingly important.\nWhilst there has been substantial research in building fair algorithmic\ndecision-making systems, the majority of these methods require access to the\ntraining data, including personal characteristics, and are not transparent\nregarding which individuals are classified unfairly. In this paper, we propose\na novel model-agnostic argumentation-based method to determine why an\nindividual is classified differently in comparison to similar individuals. Our\nmethod uses a quantitative argumentation framework to represent attribute-value\npairs of an individual and of those similar to them, and uses a well-known\nsemantics to identify the attribute-value pairs in the individual contributing\nmost to their different classification. We evaluate our method on two datasets\ncommonly used in the fairness literature and illustrate its effectiveness in\nthe identification of bias.\n", "date": "2023-10-25", "categories": []}
{"id": "2211.05321", "title": "Fairness and bias correction in machine learning for depression\n  prediction: results from four study populations", "abstract": "  A significant level of stigma and inequality exists in mental healthcare,\nespecially in under-served populations. Inequalities are reflected in the data\ncollected for scientific purposes. When not properly accounted for, machine\nlearning (ML) models leart from data can reinforce these structural\ninequalities or biases. Here, we present a systematic study of bias in ML\nmodels designed to predict depression in four different case studies covering\ndifferent countries and populations. We find that standard ML approaches show\nregularly biased behaviors. We also show that mitigation techniques, both\nstandard and our own post-hoc method, can be effective in reducing the level of\nunfair bias. No single best ML model for depression prediction provides\nequality of outcomes. This emphasizes the importance of analyzing fairness\nduring model selection and transparent reporting about the impact of debiasing\ninterventions. Finally, we provide practical recommendations to develop\nbias-aware ML models for depression risk prediction.\n", "date": "2022-11-10", "categories": []}
{"id": "2310.17489", "title": "Bias in Evaluation Processes: An Optimization-Based Model", "abstract": "  Biases with respect to socially-salient attributes of individuals have been\nwell documented in evaluation processes used in settings such as admissions and\nhiring. We view such an evaluation process as a transformation of a\ndistribution of the true utility of an individual for a task to an observed\ndistribution and model it as a solution to a loss minimization problem subject\nto an information constraint. Our model has two parameters that have been\nidentified as factors leading to biases: the resource-information trade-off\nparameter in the information constraint and the risk-averseness parameter in\nthe loss function. We characterize the distributions that arise from our model\nand study the effect of the parameters on the observed distribution. The\noutputs of our model enrich the class of distributions that can be used to\ncapture variation across groups in the observed evaluations. We empirically\nvalidate our model by fitting real-world datasets and use it to study the\neffect of interventions in a downstream selection task. These results\ncontribute to an understanding of the emergence of bias in evaluation processes\nand provide tools to guide the deployment of interventions to mitigate biases.\n", "date": "2023-10-26", "categories": []}
{"id": "2310.17530", "title": "Evaluating Bias and Fairness in Gender-Neutral Pretrained\n  Vision-and-Language Models", "abstract": "  Pretrained machine learning models are known to perpetuate and even amplify\nexisting biases in data, which can result in unfair outcomes that ultimately\nimpact user experience. Therefore, it is crucial to understand the mechanisms\nbehind those prejudicial biases to ensure that model performance does not\nresult in discriminatory behaviour toward certain groups or populations. In\nthis work, we define gender bias as our case study. We quantify bias\namplification in pretraining and after fine-tuning on three families of\nvision-and-language models. We investigate the connection, if any, between the\ntwo learning stages, and evaluate how bias amplification reflects on model\nperformance. Overall, we find that bias amplification in pretraining and after\nfine-tuning are independent. We then examine the effect of continued\npretraining on gender-neutral data, finding that this reduces group\ndisparities, i.e., promotes fairness, on VQAv2 and retrieval tasks without\nsignificantly compromising task performance.\n", "date": "2023-10-26", "categories": []}
{"id": "2310.12127", "title": "A Tale of Pronouns: Interpretability Informs Gender Bias Mitigation for\n  Fairer Instruction-Tuned Machine Translation", "abstract": "  Recent instruction fine-tuned models can solve multiple NLP tasks when\nprompted to do so, with machine translation (MT) being a prominent use case.\nHowever, current research often focuses on standard performance benchmarks,\nleaving compelling fairness and ethical considerations behind. In MT, this\nmight lead to misgendered translations, resulting, among other harms, in the\nperpetuation of stereotypes and prejudices. In this work, we address this gap\nby investigating whether and to what extent such models exhibit gender bias in\nmachine translation and how we can mitigate it. Concretely, we compute\nestablished gender bias metrics on the WinoMT corpus from English to German and\nSpanish. We discover that IFT models default to male-inflected translations,\neven disregarding female occupational stereotypes. Next, using interpretability\nmethods, we unveil that models systematically overlook the pronoun indicating\nthe gender of a target occupation in misgendered translations. Finally, based\non this finding, we propose an easy-to-implement and effective bias mitigation\nsolution based on few-shot learning that leads to significantly fairer\ntranslations.\n", "date": "2023-10-18", "categories": []}
{"id": "2208.12731", "title": "Comparing Apples to Oranges: Learning Similarity Functions for Data\n  Produced by Different Distributions", "abstract": "  Similarity functions measure how comparable pairs of elements are, and play a\nkey role in a wide variety of applications, e.g., notions of Individual\nFairness abiding by the seminal paradigm of Dwork et al., as well as Clustering\nproblems. However, access to an accurate similarity function should not always\nbe considered guaranteed, and this point was even raised by Dwork et al. For\ninstance, it is reasonable to assume that when the elements to be compared are\nproduced by different distributions, or in other words belong to different\n``demographic'' groups, knowledge of their true similarity might be very\ndifficult to obtain. In this work, we present an efficient sampling framework\nthat learns these across-groups similarity functions, using only a limited\namount of experts' feedback. We show analytical results with rigorous\ntheoretical bounds, and empirically validate our algorithms via a large suite\nof experiments.\n", "date": "2022-08-26", "categories": []}
{"id": "2210.07440", "title": "InterFair: Debiasing with Natural Language Feedback for Fair\n  Interpretable Predictions", "abstract": "  Debiasing methods in NLP models traditionally focus on isolating information\nrelated to a sensitive attribute (e.g., gender or race). We instead argue that\na favorable debiasing method should use sensitive information 'fairly,' with\nexplanations, rather than blindly eliminating it. This fair balance is often\nsubjective and can be challenging to achieve algorithmically. We explore two\ninteractive setups with a frozen predictive model and show that users able to\nprovide feedback can achieve a better and fairer balance between task\nperformance and bias mitigation. In one setup, users, by interacting with test\nexamples, further decreased bias in the explanations (5-8%) while maintaining\nthe same prediction accuracy. In the other setup, human feedback was able to\ndisentangle associated bias and predictive information from the input leading\nto superior bias mitigation and improved task performance (4-5%)\nsimultaneously.\n", "date": "2022-10-14", "categories": []}
{"id": "2210.15500", "title": "COFFEE: Counterfactual Fairness for Personalized Text Generation in\n  Explainable Recommendation", "abstract": "  As language models become increasingly integrated into our digital lives,\nPersonalized Text Generation (PTG) has emerged as a pivotal component with a\nwide range of applications. However, the bias inherent in user written text,\noften used for PTG model training, can inadvertently associate different levels\nof linguistic quality with users' protected attributes. The model can inherit\nthe bias and perpetuate inequality in generating text w.r.t. users' protected\nattributes, leading to unfair treatment when serving users. In this work, we\ninvestigate fairness of PTG in the context of personalized explanation\ngeneration for recommendations. We first discuss the biases in generated\nexplanations and their fairness implications. To promote fairness, we introduce\na general framework to achieve measure-specific counterfactual fairness in\nexplanation generation. Extensive experiments and human evaluations demonstrate\nthe effectiveness of our method.\n", "date": "2022-10-14", "categories": []}
{"id": "2303.03300", "title": "Chasing Fairness Under Distribution Shift: A Model Weight Perturbation\n  Approach", "abstract": "  Fairness in machine learning has attracted increasing attention in recent\nyears. The fairness methods improving algorithmic fairness for in-distribution\ndata may not perform well under distribution shifts. In this paper, we first\ntheoretically demonstrate the inherent connection between distribution shift,\ndata perturbation, and model weight perturbation. Subsequently, we analyze the\nsufficient conditions to guarantee fairness (i.e., low demographic parity) for\nthe target dataset, including fairness for the source dataset, and low\nprediction difference between the source and target datasets for each sensitive\nattribute group. Motivated by these sufficient conditions, we propose robust\nfairness regularization (RFR) by considering the worst case within the model\nweight perturbation ball for each sensitive attribute group. We evaluate the\neffectiveness of our proposed RFR algorithm on synthetic and real distribution\nshifts across various datasets. Experimental results demonstrate that RFR\nachieves better fairness-accuracy trade-off performance compared with several\nbaselines. The source code is available at\n\\url{https://github.com/zhimengj0326/RFR_NeurIPS23}.\n", "date": "2023-03-06", "categories": []}
{"id": "2310.13990", "title": "A Novel Information-Theoretic Objective to Disentangle Representations\n  for Fair Classification", "abstract": "  One of the pursued objectives of deep learning is to provide tools that learn\nabstract representations of reality from the observation of multiple contextual\nsituations. More precisely, one wishes to extract disentangled representations\nwhich are (i) low dimensional and (ii) whose components are independent and\ncorrespond to concepts capturing the essence of the objects under consideration\n(Locatello et al., 2019b). One step towards this ambitious project consists in\nlearning disentangled representations with respect to a predefined (sensitive)\nattribute, e.g., the gender or age of the writer. Perhaps one of the main\napplication for such disentangled representations is fair classification.\nExisting methods extract the last layer of a neural network trained with a loss\nthat is composed of a cross-entropy objective and a disentanglement\nregularizer. In this work, we adopt an information-theoretic view of this\nproblem which motivates a novel family of regularizers that minimizes the\nmutual information between the latent representation and the sensitive\nattribute conditional to the target. The resulting set of losses, called\nCLINIC, is parameter free and thus, it is easier and faster to train. CLINIC\nlosses are studied through extensive numerical experiments by training over 2k\nneural networks. We demonstrate that our methods offer a better\ndisentanglement/accuracy trade-off than previous techniques, and generalize\nbetter than training with cross-entropy loss solely provided that the\ndisentanglement task is not too constraining.\n", "date": "2023-10-21", "categories": []}
{"id": "2310.14329", "title": "DiFair: A Benchmark for Disentangled Assessment of Gender Knowledge and\n  Bias", "abstract": "  Numerous debiasing techniques have been proposed to mitigate the gender bias\nthat is prevalent in pretrained language models. These are often evaluated on\ndatasets that check the extent to which the model is gender-neutral in its\npredictions. Importantly, this evaluation protocol overlooks the possible\nadverse impact of bias mitigation on useful gender knowledge. To fill this gap,\nwe propose DiFair, a manually curated dataset based on masked language modeling\nobjectives. DiFair allows us to introduce a unified metric, gender invariance\nscore, that not only quantifies a model's biased behavior, but also checks if\nuseful gender knowledge is preserved. We use DiFair as a benchmark for a number\nof widely-used pretained language models and debiasing techniques. Experimental\nresults corroborate previous findings on the existing gender biases, while also\ndemonstrating that although debiasing techniques ameliorate the issue of gender\nbias, this improvement usually comes at the price of lowering useful gender\nknowledge of the model.\n", "date": "2023-10-22", "categories": []}
{"id": "2310.14432", "title": "Fairness-aware Optimal Graph Filter Design", "abstract": "  Graphs are mathematical tools that can be used to represent complex\nreal-world interconnected systems, such as financial markets and social\nnetworks. Hence, machine learning (ML) over graphs has attracted significant\nattention recently. However, it has been demonstrated that ML over graphs\namplifies the already existing bias towards certain under-represented groups in\nvarious decision-making problems due to the information aggregation over biased\ngraph structures. Faced with this challenge, here we take a fresh look at the\nproblem of bias mitigation in graph-based learning by borrowing insights from\ngraph signal processing. Our idea is to introduce predesigned graph filters\nwithin an ML pipeline to reduce a novel unsupervised bias measure, namely the\ncorrelation between sensitive attributes and the underlying graph connectivity.\nWe show that the optimal design of said filters can be cast as a convex problem\nin the graph spectral domain. We also formulate a linear programming (LP)\nproblem informed by a theoretical bias analysis, which attains a closed-form\nsolution and leads to a more efficient fairness-aware graph filter. Finally,\nfor a design whose degrees of freedom are independent of the input graph size,\nwe minimize the bias metric over the family of polynomial graph convolutional\nfilters. Our optimal filter designs offer complementary strengths to explore\nfavorable fairness-utility-complexity tradeoffs. For performance evaluation, we\nconduct extensive and reproducible node classification experiments over\nreal-world networks. Our results show that the proposed framework leads to\nbetter fairness measures together with similar utility compared to\nstate-of-the-art fairness-aware baselines.\n", "date": "2023-10-22", "categories": []}
{"id": "2310.14527", "title": "Marginal Nodes Matter: Towards Structure Fairness in Graphs", "abstract": "  In social network, a person located at the periphery region (marginal node)\nis likely to be treated unfairly when compared with the persons at the center.\nWhile existing fairness works on graphs mainly focus on protecting sensitive\nattributes (e.g., age and gender), the fairness incurred by the graph structure\nshould also be given attention. On the other hand, the information aggregation\nmechanism of graph neural networks amplifies such structure unfairness, as\nmarginal nodes are often far away from other nodes. In this paper, we focus on\nnovel fairness incurred by the graph structure on graph neural networks, named\n\\emph{structure fairness}. Specifically, we first analyzed multiple graphs and\nobserved that marginal nodes in graphs have a worse performance of downstream\ntasks than others in graph neural networks. Motivated by the observation, we\npropose \\textbf{S}tructural \\textbf{Fair} \\textbf{G}raph \\textbf{N}eural\n\\textbf{N}etwork (SFairGNN), which combines neighborhood expansion based\nstructure debiasing with hop-aware attentive information aggregation to achieve\nstructure fairness. Our experiments show \\SFairGNN can significantly improve\nstructure fairness while maintaining overall performance in the downstream\ntasks.\n", "date": "2023-10-23", "categories": []}
{"id": "2310.14652", "title": "Invariant Feature Regularization for Fair Face Recognition", "abstract": "  Fair face recognition is all about learning invariant feature that\ngeneralizes to unseen faces in any demographic group. Unfortunately, face\ndatasets inevitably capture the imbalanced demographic attributes that are\nubiquitous in real-world observations, and the model learns biased feature that\ngeneralizes poorly in the minority group. We point out that the bias arises due\nto the confounding demographic attributes, which mislead the model to capture\nthe spurious demographic-specific feature. The confounding effect can only be\nremoved by causal intervention, which requires the confounder annotations.\nHowever, such annotations can be prohibitively expensive due to the diversity\nof the demographic attributes. To tackle this, we propose to generate diverse\ndata partitions iteratively in an unsupervised fashion. Each data partition\nacts as a self-annotated confounder, enabling our Invariant Feature\nRegularization (INV-REG) to deconfound. INV-REG is orthogonal to existing\nmethods, and combining INV-REG with two strong baselines (Arcface and CIFP)\nleads to new state-of-the-art that improves face recognition on a variety of\ndemographic groups. Code is available at\nhttps://github.com/PanasonicConnect/InvReg.\n", "date": "2023-10-23", "categories": []}
{"id": "2310.15055", "title": "Towards Conceptualization of \"Fair Explanation\": Disparate Impacts of\n  anti-Asian Hate Speech Explanations on Content Moderators", "abstract": "  Recent research at the intersection of AI explainability and fairness has\nfocused on how explanations can improve human-plus-AI task performance as\nassessed by fairness measures. We propose to characterize what constitutes an\nexplanation that is itself \"fair\" -- an explanation that does not adversely\nimpact specific populations. We formulate a novel evaluation method of \"fair\nexplanations\" using not just accuracy and label time, but also psychological\nimpact of explanations on different user groups across many metrics (mental\ndiscomfort, stereotype activation, and perceived workload). We apply this\nmethod in the context of content moderation of potential hate speech, and its\ndifferential impact on Asian vs. non-Asian proxy moderators, across explanation\napproaches (saliency map and counterfactual explanation). We find that saliency\nmaps generally perform better and show less evidence of disparate impact\n(group) and individual unfairness than counterfactual explanations.\n  Content warning: This paper contains examples of hate speech and racially\ndiscriminatory language. The authors do not support such content. Please\nconsider your risk of discomfort carefully before continuing reading!\n", "date": "2023-10-23", "categories": []}
{"id": "2310.13364", "title": "Dissecting Causal Biases", "abstract": "  Accurately measuring discrimination in machine learning-based automated\ndecision systems is required to address the vital issue of fairness between\nsubpopulations and/or individuals. Any bias in measuring discrimination can\nlead to either amplification or underestimation of the true value of\ndiscrimination. This paper focuses on a class of bias originating in the way\ntraining data is generated and/or collected. We call such class causal biases\nand use tools from the field of causality to formally define and analyze such\nbiases. Four sources of bias are considered, namely, confounding, selection,\nmeasurement, and interaction. The main contribution of this paper is to\nprovide, for each source of bias, a closed-form expression in terms of the\nmodel parameters. This makes it possible to analyze the behavior of each source\nof bias, in particular, in which cases they are absent and in which other cases\nthey are maximized. We hope that the provided characterizations help the\ncommunity better understand the sources of bias in machine learning\napplications.\n", "date": "2023-10-20", "categories": []}
{"id": "2209.09035", "title": "Fairness in Face Presentation Attack Detection", "abstract": "  Face recognition (FR) algorithms have been proven to exhibit discriminatory\nbehaviors against certain demographic and non-demographic groups, raising\nethical and legal concerns regarding their deployment in real-world scenarios.\nDespite the growing number of fairness studies in FR, the fairness of face\npresentation attack detection (PAD) has been overlooked, mainly due to the lack\nof appropriately annotated data. To avoid and mitigate the potential negative\nimpact of such behavior, it is essential to assess the fairness in face PAD and\ndevelop fair PAD models. To enable fairness analysis in face PAD, we present a\nCombined Attribute Annotated PAD Dataset (CAAD-PAD), offering seven\nhuman-annotated attribute labels. Then, we comprehensively analyze the fairness\nof PAD and its relation to the nature of the training data and the Operational\nDecision Threshold Assignment (ODTA) through a set of face PAD solutions.\nAdditionally, we propose a novel metric, the Accuracy Balanced Fairness (ABF),\nthat jointly represents both the PAD fairness and the absolute PAD performance.\nThe experimental results pointed out that female and faces with occluding\nfeatures (e.g. eyeglasses, beard, etc.) are relatively less protected than male\nand non-occlusion groups by all PAD solutions. To alleviate this observed\nunfairness, we propose a plug-and-play data augmentation method, FairSWAP, to\ndisrupt the identity/semantic information and encourage models to mine the\nattack clues. The extensive experimental results indicate that FairSWAP leads\nto better-performing and fairer face PADs in 10 out of 12 investigated cases.\n", "date": "2022-09-19", "categories": []}
{"id": "2310.12785", "title": "A Theoretical Approach to Characterize the Accuracy-Fairness Trade-off\n  Pareto Frontier", "abstract": "  While the accuracy-fairness trade-off has been frequently observed in the\nliterature of fair machine learning, rigorous theoretical analyses have been\nscarce. To demystify this long-standing challenge, this work seeks to develop a\ntheoretical framework by characterizing the shape of the accuracy-fairness\ntrade-off Pareto frontier (FairFrontier), determined by a set of all optimal\nPareto classifiers that no other classifiers can dominate. Specifically, we\nfirst demonstrate the existence of the trade-off in real-world scenarios and\nthen propose four potential categories to characterize the important properties\nof the accuracy-fairness Pareto frontier. For each category, we identify the\nnecessary conditions that lead to corresponding trade-offs. Experimental\nresults on synthetic data suggest insightful findings of the proposed\nframework: (1) When sensitive attributes can be fully interpreted by\nnon-sensitive attributes, FairFrontier is mostly continuous. (2) Accuracy can\nsuffer a \\textit{sharp} decline when over-pursuing fairness. (3) Eliminate the\ntrade-off via a two-step streamlined approach. The proposed research enables an\nin-depth understanding of the accuracy-fairness trade-off, pushing current fair\nmachine-learning research to a new frontier.\n", "date": "2023-10-19", "categories": []}
{"id": "2310.12805", "title": "Detection and Evaluation of bias-inducing Features in Machine learning", "abstract": "  The cause-to-effect analysis can help us decompose all the likely causes of a\nproblem, such as an undesirable business situation or unintended harm to the\nindividual(s). This implies that we can identify how the problems are\ninherited, rank the causes to help prioritize fixes, simplify a complex problem\nand visualize them. In the context of machine learning (ML), one can use\ncause-to-effect analysis to understand the reason for the biased behavior of\nthe system. For example, we can examine the root causes of biases by checking\neach feature for a potential cause of bias in the model. To approach this, one\ncan apply small changes to a given feature or a pair of features in the data,\nfollowing some guidelines and observing how it impacts the decision made by the\nmodel (i.e., model prediction). Therefore, we can use cause-to-effect analysis\nto identify the potential bias-inducing features, even when these features are\noriginally are unknown. This is important since most current methods require a\npre-identification of sensitive features for bias assessment and can actually\nmiss other relevant bias-inducing features, which is why systematic\nidentification of such features is necessary. Moreover, it often occurs that to\nachieve an equitable outcome, one has to take into account sensitive features\nin the model decision. Therefore, it should be up to the domain experts to\ndecide based on their knowledge of the context of a decision whether bias\ninduced by specific features is acceptable or not. In this study, we propose an\napproach for systematically identifying all bias-inducing features of a model\nto help support the decision-making of domain experts. We evaluated our\ntechnique using four well-known datasets to showcase how our contribution can\nhelp spearhead the standard procedure when developing, testing, maintaining,\nand deploying fair/equitable machine learning systems.\n", "date": "2023-10-19", "categories": []}
{"id": "2211.09925", "title": "FairMILE: Towards an Efficient Framework for Fair Graph Representation\n  Learning", "abstract": "  Graph representation learning models have demonstrated great capability in\nmany real-world applications. Nevertheless, prior research indicates that these\nmodels can learn biased representations leading to discriminatory outcomes. A\nfew works have been proposed to mitigate the bias in graph representations.\nHowever, most existing works require exceptional time and computing resources\nfor training and fine-tuning. To this end, we study the problem of efficient\nfair graph representation learning and propose a novel framework FairMILE.\nFairMILE is a multi-level paradigm that can efficiently learn graph\nrepresentations while enforcing fairness and preserving utility. It can work in\nconjunction with any unsupervised embedding approach and accommodate various\nfairness constraints. Extensive experiments across different downstream tasks\ndemonstrate that FairMILE significantly outperforms state-of-the-art baselines\nin terms of running time while achieving a superior trade-off between fairness\nand utility.\n", "date": "2022-11-17", "categories": []}
{"id": "2310.11867", "title": "Evaluating the Fairness of Discriminative Foundation Models in Computer\n  Vision", "abstract": "  We propose a novel taxonomy for bias evaluation of discriminative foundation\nmodels, such as Contrastive Language-Pretraining (CLIP), that are used for\nlabeling tasks. We then systematically evaluate existing methods for mitigating\nbias in these models with respect to our taxonomy. Specifically, we evaluate\nOpenAI's CLIP and OpenCLIP models for key applications, such as zero-shot\nclassification, image retrieval and image captioning. We categorize desired\nbehaviors based around three axes: (i) if the task concerns humans; (ii) how\nsubjective the task is (i.e., how likely it is that people from a diverse range\nof backgrounds would agree on a labeling); and (iii) the intended purpose of\nthe task and if fairness is better served by impartiality (i.e., making\ndecisions independent of the protected attributes) or representation (i.e.,\nmaking decisions to maximize diversity). Finally, we provide quantitative\nfairness evaluations for both binary-valued and multi-valued protected\nattributes over ten diverse datasets. We find that fair PCA, a post-processing\nmethod for fair representations, works very well for debiasing in most of the\naforementioned tasks while incurring only minor loss of performance. However,\ndifferent debiasing approaches vary in their effectiveness depending on the\ntask. Hence, one should choose the debiasing approach depending on the specific\nuse case.\n", "date": "2023-10-18", "categories": []}
{"id": "2310.12145", "title": "Fairer and More Accurate Tabular Models Through NAS", "abstract": "  Making models algorithmically fairer in tabular data has been long studied,\nwith techniques typically oriented towards fixes which usually take a neural\nmodel with an undesirable outcome and make changes to how the data are\ningested, what the model weights are, or how outputs are processed. We employ\nan emergent and different strategy where we consider updating the model's\narchitecture and training hyperparameters to find an entirely new model with\nbetter outcomes from the beginning of the debiasing procedure. In this work, we\npropose using multi-objective Neural Architecture Search (NAS) and\nHyperparameter Optimization (HPO) in the first application to the very\nchallenging domain of tabular data. We conduct extensive exploration of\narchitectural and hyperparameter spaces (MLP, ResNet, and FT-Transformer)\nacross diverse datasets, demonstrating the dependence of accuracy and fairness\nmetrics of model predictions on hyperparameter combinations. We show that\nmodels optimized solely for accuracy with NAS often fail to inherently address\nfairness concerns. We propose a novel approach that jointly optimizes\narchitectural and training hyperparameters in a multi-objective constraint of\nboth accuracy and fairness. We produce architectures that consistently Pareto\ndominate state-of-the-art bias mitigation methods either in fairness, accuracy\nor both, all of this while being Pareto-optimal over hyperparameters achieved\nthrough single-objective (accuracy) optimization runs. This research\nunderscores the promise of automating fairness and accuracy optimization in\ndeep learning models.\n", "date": "2023-10-18", "categories": []}
{"id": "2310.11079", "title": "Learning from Red Teaming: Gender Bias Provocation and Mitigation in\n  Large Language Models", "abstract": "  Recently, researchers have made considerable improvements in dialogue systems\nwith the progress of large language models (LLMs) such as ChatGPT and GPT-4.\nThese LLM-based chatbots encode the potential biases while retaining\ndisparities that can harm humans during interactions. The traditional biases\ninvestigation methods often rely on human-written test cases. However, these\ntest cases are usually expensive and limited. In this work, we propose a\nfirst-of-its-kind method that automatically generates test cases to detect\nLLMs' potential gender bias. We apply our method to three well-known LLMs and\nfind that the generated test cases effectively identify the presence of biases.\nTo address the biases identified, we propose a mitigation strategy that uses\nthe generated test cases as demonstrations for in-context learning to\ncircumvent the need for parameter fine-tuning. The experimental results show\nthat LLMs generate fairer responses with the proposed approach.\n", "date": "2023-10-17", "categories": []}
{"id": "2310.09373", "title": "Identifying and examining machine learning biases on Adult dataset", "abstract": "  This research delves into the reduction of machine learning model bias\nthrough Ensemble Learning. Our rigorous methodology comprehensively assesses\nbias across various categorical variables, ultimately revealing a pronounced\ngender attribute bias. The empirical evidence unveils a substantial\ngender-based wage prediction disparity: wages predicted for males, initially at\n\\$902.91, significantly decrease to \\$774.31 when the gender attribute is\nalternated to females. Notably, Kullback-Leibler divergence scores point to\ngender bias, with values exceeding 0.13, predominantly within tree-based\nmodels. Employing Ensemble Learning elucidates the quest for fairness and\ntransparency. Intriguingly, our findings reveal that the stacked model aligns\nwith individual models, confirming the resilience of model bias. This study\nunderscores ethical considerations and advocates the implementation of hybrid\nmodels for a data-driven society marked by impartiality and inclusivity.\n", "date": "2023-10-13", "categories": []}
{"id": "2310.10399", "title": "Towards Fair and Calibrated Models", "abstract": "  Recent literature has seen a significant focus on building machine learning\nmodels with specific properties such as fairness, i.e., being non-biased with\nrespect to a given set of attributes, calibration i.e., model confidence being\naligned with its predictive accuracy, and explainability, i.e., ability to be\nunderstandable to humans. While there has been work focusing on each of these\naspects individually, researchers have shied away from simultaneously\naddressing more than one of these dimensions. In this work, we address the\nproblem of building models which are both fair and calibrated. We work with a\nspecific definition of fairness, which closely matches [Biswas et. al. 2019],\nand has the nice property that Bayes optimal classifier has the maximum\npossible fairness under our definition. We show that an existing negative\nresult towards achieving a fair and calibrated model [Kleinberg et. al. 2017]\ndoes not hold for our definition of fairness. Further, we show that ensuring\ngroup-wise calibration with respect to the sensitive attributes automatically\nresults in a fair model under our definition. Using this result, we provide a\nfirst cut approach for achieving fair and calibrated models, via a simple\npost-processing technique based on temperature scaling. We then propose\nmodifications of existing calibration losses to perform group-wise calibration,\nas a way of achieving fair and calibrated models in a variety of settings.\nFinally, we perform extensive experimentation of these techniques on a diverse\nbenchmark of datasets, and present insights on the pareto-optimality of the\nresulting solutions.\n", "date": "2023-10-16", "categories": []}
{"id": "2310.10571", "title": "Emerging Challenges in Personalized Medicine: Assessing Demographic\n  Effects on Biomedical Question Answering Systems", "abstract": "  State-of-the-art question answering (QA) models exhibit a variety of social\nbiases (e.g., with respect to sex or race), generally explained by similar\nissues in their training data. However, what has been overlooked so far is that\nin the critical domain of biomedicine, any unjustified change in model output\ndue to patient demographics is problematic: it results in the unfair treatment\nof patients. Selecting only questions on biomedical topics whose answers do not\ndepend on ethnicity, sex, or sexual orientation, we ask the following research\nquestions: (RQ1) Do the answers of QA models change when being provided with\nirrelevant demographic information? (RQ2) Does the answer of RQ1 differ between\nknowledge graph (KG)-grounded and text-based QA systems? We find that\nirrelevant demographic information change up to 15% of the answers of a\nKG-grounded system and up to 23% of the answers of a text-based system,\nincluding changes that affect accuracy. We conclude that unjustified answer\nchanges caused by patient demographics are a frequent phenomenon, which raises\nfairness concerns and should be paid more attention to.\n", "date": "2023-10-16", "categories": []}
{"id": "2310.08455", "title": "Metrics for popularity bias in dynamic recommender systems", "abstract": "  Albeit the widespread application of recommender systems (RecSys) in our\ndaily lives, rather limited research has been done on quantifying unfairness\nand biases present in such systems. Prior work largely focuses on determining\nwhether a RecSys is discriminating or not but does not compute the amount of\nbias present in these systems. Biased recommendations may lead to decisions\nthat can potentially have adverse effects on individuals, sensitive user\ngroups, and society. Hence, it is important to quantify these biases for fair\nand safe commercial applications of these systems. This paper focuses on\nquantifying popularity bias that stems directly from the output of RecSys\nmodels, leading to over recommendation of popular items that are likely to be\nmisaligned with user preferences. Four metrics to quantify popularity bias in\nRescSys over time in dynamic setting across different sensitive user groups\nhave been proposed. These metrics have been demonstrated for four collaborative\nfiltering based RecSys algorithms trained on two commonly used benchmark\ndatasets in the literature. Results obtained show that the metrics proposed\nprovide a comprehensive understanding of growing disparities in treatment\nbetween sensitive groups over time when used conjointly.\n", "date": "2023-10-12", "categories": []}
{"id": "2207.07068", "title": "Bias Mitigation for Machine Learning Classifiers: A Comprehensive Survey", "abstract": "  This paper provides a comprehensive survey of bias mitigation methods for\nachieving fairness in Machine Learning (ML) models. We collect a total of 341\npublications concerning bias mitigation for ML classifiers. These methods can\nbe distinguished based on their intervention procedure (i.e., pre-processing,\nin-processing, post-processing) and the technique they apply. We investigate\nhow existing bias mitigation methods are evaluated in the literature. In\nparticular, we consider datasets, metrics and benchmarking. Based on the\ngathered insights (e.g., What is the most popular fairness metric? How many\ndatasets are used for evaluating bias mitigation methods?), we hope to support\npractitioners in making informed choices when developing and evaluating new\nbias mitigation methods.\n", "date": "2022-07-14", "categories": []}
{"id": "2310.06904", "title": "Mitigating stereotypical biases in text to image generative systems", "abstract": "  State-of-the-art generative text-to-image models are known to exhibit social\nbiases and over-represent certain groups like people of perceived lighter skin\ntones and men in their outcomes. In this work, we propose a method to mitigate\nsuch biases and ensure that the outcomes are fair across different groups of\npeople. We do this by finetuning text-to-image models on synthetic data that\nvaries in perceived skin tones and genders constructed from diverse text\nprompts. These text prompts are constructed from multiplicative combinations of\nethnicities, genders, professions, age groups, and so on, resulting in diverse\nsynthetic data. Our diversity finetuned (DFT) model improves the group fairness\nmetric by 150% for perceived skin tone and 97.7% for perceived gender. Compared\nto baselines, DFT models generate more people with perceived darker skin tone\nand more women. To foster open research, we will release all text prompts and\ncode to generate training images.\n", "date": "2023-10-10", "categories": []}
{"id": "2310.06061", "title": "Auditing Gender Analyzers on Text Data", "abstract": "  AI models have become extremely popular and accessible to the general public.\nHowever, they are continuously under the scanner due to their demonstrable\nbiases toward various sections of the society like people of color and\nnon-binary people. In this study, we audit three existing gender analyzers --\nuClassify, Readable and HackerFactor, for biases against non-binary\nindividuals. These tools are designed to predict only the cisgender binary\nlabels, which leads to discrimination against non-binary members of the\nsociety. We curate two datasets -- Reddit comments (660k) and, Tumblr posts\n(2.05M) and our experimental evaluation shows that the tools are highly\ninaccurate with the overall accuracy being ~50% on all platforms. Predictions\nfor non-binary comments on all platforms are mostly female, thus propagating\nthe societal bias that non-binary individuals are effeminate. To address this,\nwe fine-tune a BERT multi-label classifier on the two datasets in multiple\ncombinations, observe an overall performance of ~77% on the most realistically\ndeployable setting and a surprisingly higher performance of 90% for the\nnon-binary class. We also audit ChatGPT using zero-shot prompts on a small\ndataset (due to high pricing) and observe an average accuracy of 58% for Reddit\nand Tumblr combined (with overall better results for Reddit).\n  Thus, we show that existing systems, including highly advanced ones like\nChatGPT are biased, and need better audits and moderation and, that such\nsocietal biases can be addressed and alleviated through simple off-the-shelf\nmodels like BERT trained on more gender inclusive datasets.\n", "date": "2023-10-09", "categories": []}
{"id": "2310.06205", "title": "Fair Classifiers that Abstain without Harm", "abstract": "  In critical applications, it is vital for classifiers to defer\ndecision-making to humans. We propose a post-hoc method that makes existing\nclassifiers selectively abstain from predicting certain samples. Our abstaining\nclassifier is incentivized to maintain the original accuracy for each\nsub-population (i.e. no harm) while achieving a set of group fairness\ndefinitions to a user specified degree. To this end, we design an Integer\nProgramming (IP) procedure that assigns abstention decisions for each training\nsample to satisfy a set of constraints. To generalize the abstaining decisions\nto test samples, we then train a surrogate model to learn the abstaining\ndecisions based on the IP solutions in an end-to-end manner. We analyze the\nfeasibility of the IP procedure to determine the possible abstention rate for\ndifferent levels of unfairness tolerance and accuracy constraint for achieving\nno harm. To the best of our knowledge, this work is the first to identify the\ntheoretical relationships between the constraint parameters and the required\nabstention rate. Our theoretical results are important since a high abstention\nrate is often infeasible in practice due to a lack of human resources. Our\nframework outperforms existing methods in terms of fairness disparity without\nsacrificing accuracy at similar abstention rates.\n", "date": "2023-10-09", "categories": []}
{"id": "2210.05248", "title": "Self-supervised debiasing using low rank regularization", "abstract": "  Spurious correlations can cause strong biases in deep neural networks,\nimpairing generalization ability. While most existing debiasing methods require\nfull supervision on either spurious attributes or target labels, training a\ndebiased model from a limited amount of both annotations is still an open\nquestion. To address this issue, we investigate an interesting phenomenon using\nthe spectral analysis of latent representations: spuriously correlated\nattributes make neural networks inductively biased towards encoding lower\neffective rank representations. We also show that a rank regularization can\namplify this bias in a way that encourages highly correlated features.\nLeveraging these findings, we propose a self-supervised debiasing framework\npotentially compatible with unlabeled samples. Specifically, we first pretrain\na biased encoder in a self-supervised manner with the rank regularization,\nserving as a semantic bottleneck to enforce the encoder to learn the spuriously\ncorrelated attributes. This biased encoder is then used to discover and\nupweight bias-conflicting samples in a downstream task, serving as a boosting\nto effectively debias the main model. Remarkably, the proposed debiasing\nframework significantly improves the generalization performance of\nself-supervised learning baselines and, in some cases, even outperforms\nstate-of-the-art supervised debiasing approaches.\n", "date": "2022-10-11", "categories": []}
{"id": "2303.08516", "title": "Fair Off-Policy Learning from Observational Data", "abstract": "  Algorithmic decision-making in practice must be fair for legal, ethical, and\nsocietal reasons. To achieve this, prior research has contributed various\napproaches that ensure fairness in machine learning predictions, while\ncomparatively little effort has focused on fairness in decision-making,\nspecifically off-policy learning. In this paper, we propose a novel framework\nfor fair off-policy learning: we learn decision rules from observational data\nunder different notions of fairness, where we explicitly assume that\nobservational data were collected under a different potentially discriminatory\nbehavioral policy. For this, we first formalize different fairness notions for\noff-policy learning. We then propose a neural network-based framework to learn\noptimal policies under different fairness notions. We further provide\ntheoretical guarantees in the form of generalization bounds for the\nfinite-sample version of our framework. We demonstrate the effectiveness of our\nframework through extensive numerical experiments using both simulated and\nreal-world data. Altogether, our work enables algorithmic decision-making in a\nwide array of practical applications where fairness must be ensured.\n", "date": "2023-03-15", "categories": []}
{"id": "2305.10906", "title": "RobustFair: Adversarial Evaluation through Fairness Confusion Directed\n  Gradient Search", "abstract": "  Deep neural networks (DNNs) often face challenges due to their vulnerability\nto various adversarial perturbations, including false perturbations that\nundermine prediction accuracy and biased perturbations that cause biased\npredictions for similar inputs. This paper introduces a novel approach,\nRobustFair, to evaluate the accurate fairness of DNNs when subjected to these\nfalse or biased perturbations. RobustFair employs the notion of the fairness\nconfusion matrix induced in accurate fairness to identify the crucial input\nfeatures for perturbations. This matrix categorizes predictions as true fair,\ntrue biased, false fair, and false biased, and the perturbations guided by it\ncan produce a dual impact on instances and their similar counterparts to either\nundermine prediction accuracy (robustness) or cause biased predictions\n(individual fairness). RobustFair then infers the ground truth of these\ngenerated adversarial instances based on their loss function values\napproximated by the total derivative. To leverage the generated instances for\ntrustworthiness improvement, RobustFair further proposes a data augmentation\nstrategy to prioritize adversarial instances resembling the original training\nset, for data augmentation and model retraining. Notably, RobustFair excels at\ndetecting intertwined issues of robustness and individual fairness, which are\nfrequently overlooked in standard robustness and individual fairness\nevaluations. This capability empowers RobustFair to enhance both robustness and\nindividual fairness evaluations by concurrently identifying defects in either\ndomain. Empirical case studies and quantile regression analyses on benchmark\ndatasets demonstrate the effectiveness of the fairness confusion matrix guided\nperturbation for false or biased adversarial instance generation.\n", "date": "2023-05-18", "categories": []}
{"id": "2307.12065", "title": "Spectral Normalized-Cut Graph Partitioning with Fairness Constraints", "abstract": "  Normalized-cut graph partitioning aims to divide the set of nodes in a graph\ninto $k$ disjoint clusters to minimize the fraction of the total edges between\nany cluster and all other clusters. In this paper, we consider a fair variant\nof the partitioning problem wherein nodes are characterized by a categorical\nsensitive attribute (e.g., gender or race) indicating membership to different\ndemographic groups. Our goal is to ensure that each group is approximately\nproportionally represented in each cluster while minimizing the normalized cut\nvalue. To resolve this problem, we propose a two-phase spectral algorithm\ncalled FNM. In the first phase, we add an augmented Lagrangian term based on\nour fairness criteria to the objective function for obtaining a fairer spectral\nnode embedding. Then, in the second phase, we design a rounding scheme to\nproduce $k$ clusters from the fair embedding that effectively trades off\nfairness and partition quality. Through comprehensive experiments on nine\nbenchmark datasets, we demonstrate the superior performance of FNM compared\nwith three baseline methods.\n", "date": "2023-07-22", "categories": []}
{"id": "2310.05135", "title": "Are Emily and Greg Still More Employable than Lakisha and Jamal?\n  Investigating Algorithmic Hiring Bias in the Era of ChatGPT", "abstract": "  Large Language Models (LLMs) such as GPT-3.5, Bard, and Claude exhibit\napplicability across numerous tasks. One domain of interest is their use in\nalgorithmic hiring, specifically in matching resumes with job categories. Yet,\nthis introduces issues of bias on protected attributes like gender, race and\nmaternity status. The seminal work of Bertrand & Mullainathan (2003) set the\ngold-standard for identifying hiring bias via field experiments where the\nresponse rate for identical resumes that differ only in protected attributes,\ne.g., racially suggestive names such as Emily or Lakisha, is compared. We\nreplicate this experiment on state-of-art LLMs (GPT-3.5, Bard, Claude and\nLlama) to evaluate bias (or lack thereof) on gender, race, maternity status,\npregnancy status, and political affiliation. We evaluate LLMs on two tasks: (1)\nmatching resumes to job categories; and (2) summarizing resumes with employment\nrelevant information. Overall, LLMs are robust across race and gender. They\ndiffer in their performance on pregnancy status and political affiliation. We\nuse contrastive input decoding on open-source LLMs to uncover potential sources\nof bias.\n", "date": "2023-10-08", "categories": []}
{"id": "2212.07877", "title": "Manifestations of Xenophobia in AI Systems", "abstract": "  Xenophobia is one of the key drivers of marginalisation, discrimination, and\nconflict, yet many prominent machine learning (ML) fairness frameworks fail to\ncomprehensively measure or mitigate the resulting xenophobic harms. Here we aim\nto bridge this conceptual gap and help facilitate safe and ethical design of\nartificial intelligence (AI) solutions. We ground our analysis of the impact of\nxenophobia by first identifying distinct types of xenophobic harms, and then\napplying this framework across a number of prominent AI application domains,\nreviewing the potential interplay between AI and xenophobia on social media and\nrecommendation systems, healthcare, immigration, employment, as well as biases\nin large pre-trained models. These help inform our recommendations towards an\ninclusive, xenophilic design of future AI systems.\n", "date": "2022-12-15", "categories": []}
{"id": "2310.04352", "title": "Fair Feature Importance Scores for Interpreting Tree-Based Methods and\n  Surrogates", "abstract": "  Across various sectors such as healthcare, criminal justice, national\nsecurity, finance, and technology, large-scale machine learning (ML) and\nartificial intelligence (AI) systems are being deployed to make critical\ndata-driven decisions. Many have asked if we can and should trust these ML\nsystems to be making these decisions. Two critical components are prerequisites\nfor trust in ML systems: interpretability, or the ability to understand why the\nML system makes the decisions it does, and fairness, which ensures that ML\nsystems do not exhibit bias against certain individuals or groups. Both\ninterpretability and fairness are important and have separately received\nabundant attention in the ML literature, but so far, there have been very few\nmethods developed to directly interpret models with regard to their fairness.\nIn this paper, we focus on arguably the most popular type of ML interpretation:\nfeature importance scores. Inspired by the use of decision trees in knowledge\ndistillation, we propose to leverage trees as interpretable surrogates for\ncomplex black-box ML models. Specifically, we develop a novel fair feature\nimportance score for trees that can be used to interpret how each feature\ncontributes to fairness or bias in trees, tree-based ensembles, or tree-based\nsurrogates of any complex ML system. Like the popular mean decrease in impurity\nfor trees, our Fair Feature Importance Score is defined based on the mean\ndecrease (or increase) in group bias. Through simulations as well as real\nexamples on benchmark fairness datasets, we demonstrate that our Fair Feature\nImportance Score offers valid interpretations for both tree-based ensembles and\ntree-based surrogates of other ML systems.\n", "date": "2023-10-06", "categories": []}
{"id": "2310.03655", "title": "Strategic Evaluation: Subjects, Evaluators, and Society", "abstract": "  A broad current application of algorithms is in formal and quantitative\nmeasures of murky concepts -- like merit -- to make decisions. When people\nstrategically respond to these sorts of evaluations in order to gain favorable\ndecision outcomes, their behavior can be subjected to moral judgments. They may\nbe described as 'gaming the system' or 'cheating,' or (in other cases)\ninvesting 'honest effort' or 'improving.' Machine learning literature on\nstrategic behavior has tried to describe these dynamics by emphasizing the\nefforts expended by decision subjects hoping to obtain a more favorable\nassessment -- some works offer ways to preempt or prevent such manipulations,\nsome differentiate 'gaming' from 'improvement' behavior, while others aim to\nmeasure the effort burden or disparate effects of classification systems. We\nbegin from a different starting point: that the design of an evaluation itself\ncan be understood as furthering goals held by the evaluator which may be\nmisaligned with broader societal goals. To develop the idea that evaluation\nrepresents a strategic interaction in which both the evaluator and the subject\nof their evaluation are operating out of self-interest, we put forward a model\nthat represents the process of evaluation using three interacting agents: a\ndecision subject, an evaluator, and society, representing a bundle of values\nand oversight mechanisms. We highlight our model's applicability to a number of\nsocial systems where one or two players strategically undermine the others'\ninterests to advance their own. Treating evaluators as themselves strategic\nallows us to re-cast the scrutiny directed at decision subjects, towards the\nincentives that underpin institutional designs of evaluations. The moral\nstanding of strategic behaviors often depend on the moral standing of the\nevaluations and incentives that provoke such behaviors.\n", "date": "2023-10-05", "categories": []}
{"id": "2309.05148", "title": "Beyond Skin Tone: A Multidimensional Measure of Apparent Skin Color", "abstract": "  This paper strives to measure apparent skin color in computer vision, beyond\na unidimensional scale on skin tone. In their seminal paper Gender Shades,\nBuolamwini and Gebru have shown how gender classification systems can be biased\nagainst women with darker skin tones. Subsequently, fairness researchers and\npractitioners have adopted the Fitzpatrick skin type classification as a common\nmeasure to assess skin color bias in computer vision systems. While effective,\nthe Fitzpatrick scale only focuses on the skin tone ranging from light to dark.\nTowards a more comprehensive measure of skin color, we introduce the hue angle\nranging from red to yellow. When applied to images, the hue dimension reveals\nadditional biases related to skin color in both computer vision datasets and\nmodels. We then recommend multidimensional skin color scales, relying on both\nskin tone and hue, for fairness assessments.\n", "date": "2023-09-10", "categories": []}
{"id": "2310.02533", "title": "Quantifying and mitigating the impact of label errors on model disparity\n  metrics", "abstract": "  Errors in labels obtained via human annotation adversely affect a model's\nperformance. Existing approaches propose ways to mitigate the effect of label\nerror on a model's downstream accuracy, yet little is known about its impact on\na model's disparity metrics. Here we study the effect of label error on a\nmodel's disparity metrics. We empirically characterize how varying levels of\nlabel error, in both training and test data, affect these disparity metrics. We\nfind that group calibration and other metrics are sensitive to train-time and\ntest-time label error -- particularly for minority groups. This disparate\neffect persists even for models trained with noise-aware algorithms. To\nmitigate the impact of training-time label error, we present an approach to\nestimate the influence of a training input's label on a model's group disparity\nmetric. We empirically assess the proposed approach on a variety of datasets\nand find significant improvement, compared to alternative approaches, in\nidentifying training inputs that improve a model's disparity metric. We\ncomplement the approach with an automatic relabel-and-finetune scheme that\nproduces updated models with, provably, improved group calibration error.\n", "date": "2023-10-04", "categories": []}
{"id": "2310.02752", "title": "Fair Feature Selection: A Comparison of Multi-Objective Genetic\n  Algorithms", "abstract": "  Machine learning classifiers are widely used to make decisions with a major\nimpact on people's lives (e.g. accepting or denying a loan, hiring decisions,\netc). In such applications,the learned classifiers need to be both accurate and\nfair with respect to different groups of people, with different values of\nvariables such as sex and race. This paper focuses on fair feature selection\nfor classification, i.e. methods that select a feature subset aimed at\nmaximising both the accuracy and the fairness of the predictions made by a\nclassifier. More specifically, we compare two recently proposed Genetic\nAlgorithms (GAs) for fair feature selection that are based on two different\nmulti-objective optimisation approaches: (a) a Pareto dominance-based GA; and\n(b) a lexicographic optimisation-based GA, where maximising accuracy has higher\npriority than maximising fairness. Both GAs use the same measures of accuracy\nand fairness, allowing for a controlled comparison. As far as we know, this is\nthe first comparison between the Pareto and lexicographic approaches for fair\nclassification. The results show that, overall, the lexicographic GA\noutperformed the Pareto GA with respect to accuracy without degradation of the\nfairness of the learned classifiers. This is an important result because at\npresent nearly all GAs for fair classification are based on the Pareto\napproach, so these results suggest a promising new direction for research in\nthis area.\n", "date": "2023-10-04", "categories": []}
{"id": "2305.13057", "title": "Causality-Aided Trade-off Analysis for Machine Learning Fairness", "abstract": "  There has been an increasing interest in enhancing the fairness of machine\nlearning (ML). Despite the growing number of fairness-improving methods, we\nlack a systematic understanding of the trade-offs among factors considered in\nthe ML pipeline when fairness-improving methods are applied. This understanding\nis essential for developers to make informed decisions regarding the provision\nof fair ML services. Nonetheless, it is extremely difficult to analyze the\ntrade-offs when there are multiple fairness parameters and other crucial\nmetrics involved, coupled, and even in conflict with one another.\n  This paper uses causality analysis as a principled method for analyzing\ntrade-offs between fairness parameters and other crucial metrics in ML\npipelines. To ractically and effectively conduct causality analysis, we propose\na set of domain-specific optimizations to facilitate accurate causal discovery\nand a unified, novel interface for trade-off analysis based on well-established\ncausal inference methods. We conduct a comprehensive empirical study using\nthree real-world datasets on a collection of widelyused fairness-improving\ntechniques. Our study obtains actionable suggestions for users and developers\nof fair ML. We further demonstrate the versatile usage of our approach in\nselecting the optimal fairness-improving method, paving the way for more\nethical and socially responsible AI technologies.\n", "date": "2023-05-22", "categories": []}
{"id": "2306.00374", "title": "CFL: Causally Fair Language Models Through Token-level Attribute\n  Controlled Generation", "abstract": "  We propose a method to control the attributes of Language Models (LMs) for\nthe text generation task using Causal Average Treatment Effect (ATE) scores and\ncounterfactual augmentation. We explore this method, in the context of LM\ndetoxification, and propose the Causally Fair Language (CFL) architecture for\ndetoxifying pre-trained LMs in a plug-and-play manner. Our architecture is\nbased on a Structural Causal Model (SCM) that is mathematically transparent and\ncomputationally efficient as compared with many existing detoxification\ntechniques. We also propose several new metrics that aim to better understand\nthe behaviour of LMs in the context of toxic text generation. Further, we\nachieve state of the art performance for toxic degeneration, which are computed\nusing \\RTP (RTP) benchmark. Our experiments show that CFL achieves such a\ndetoxification without much impact on the model perplexity. We also show that\nCFL mitigates the unintended bias problem through experiments on the BOLD\ndataset.\n", "date": "2023-06-01", "categories": []}
{"id": "2310.01679", "title": "Estimating and Implementing Conventional Fairness Metrics With\n  Probabilistic Protected Features", "abstract": "  The vast majority of techniques to train fair models require access to the\nprotected attribute (e.g., race, gender), either at train time or in\nproduction. However, in many important applications this protected attribute is\nlargely unavailable. In this paper, we develop methods for measuring and\nreducing fairness violations in a setting with limited access to protected\nattribute labels. Specifically, we assume access to protected attribute labels\non a small subset of the dataset of interest, but only probabilistic estimates\nof protected attribute labels (e.g., via Bayesian Improved Surname Geocoding)\nfor the rest of the dataset. With this setting in mind, we propose a method to\nestimate bounds on common fairness metrics for an existing model, as well as a\nmethod for training a model to limit fairness violations by solving a\nconstrained non-convex optimization problem. Unlike similar existing\napproaches, our methods take advantage of contextual information --\nspecifically, the relationships between a model's predictions and the\nprobabilistic prediction of protected attributes, given the true protected\nattribute, and vice versa -- to provide tighter bounds on the true disparity.\nWe provide an empirical illustration of our methods using voting data. First,\nwe show our measurement method can bound the true disparity up to 5.5x tighter\nthan previous methods in these applications. Then, we demonstrate that our\ntraining technique effectively reduces disparity while incurring lesser\nfairness-accuracy trade-offs than other fair optimization methods with limited\naccess to protected attributes.\n", "date": "2023-10-02", "categories": []}
{"id": "2209.02965", "title": "Risk of Bias in Chest Radiography Deep Learning Foundation Models", "abstract": "  Purpose: To analyze a recently published chest radiography foundation model\nfor the presence of biases that could lead to subgroup performance disparities\nacross biological sex and race.\n  Materials and Methods: This retrospective study used 127,118 chest\nradiographs from 42,884 patients (mean age, 63 [SD] 17 years; 23,623 male,\n19,261 female) from the CheXpert dataset collected between October 2002 and\nJuly 2017. To determine the presence of bias in features generated by a chest\nradiography foundation model and baseline deep learning model, dimensionality\nreduction methods together with two-sample Kolmogorov-Smirnov tests were used\nto detect distribution shifts across sex and race. A comprehensive disease\ndetection performance analysis was then performed to associate any biases in\nthe features to specific disparities in classification performance across\npatient subgroups.\n  Results: Ten out of twelve pairwise comparisons across biological sex and\nrace showed statistically significant differences in the studied foundation\nmodel, compared with four significant tests in the baseline model. Significant\ndifferences were found between male and female (P < .001) and Asian and Black\npatients (P < .001) in the feature projections that primarily capture disease.\nCompared with average model performance across all subgroups, classification\nperformance on the 'no finding' label dropped between 6.8% and 7.8% for female\npatients, and performance in detecting 'pleural effusion' dropped between 10.7%\nand 11.6% for Black patients.\n  Conclusion: The studied chest radiography foundation model demonstrated\nracial and sex-related bias leading to disparate performance across patient\nsubgroups and may be unsafe for clinical applications.\n", "date": "2022-09-07", "categories": []}
{"id": "2303.08040", "title": "Beyond Demographic Parity: Redefining Equal Treatment", "abstract": "  Liberalism-oriented political philosophy reasons that all individuals should\nbe treated equally independently of their protected characteristics. Related\nwork in machine learning has translated the concept of \\emph{equal treatment}\ninto terms of \\emph{equal outcome} and measured it as \\emph{demographic parity}\n(also called \\emph{statistical parity}). Our analysis reveals that the two\nconcepts of equal outcome and equal treatment diverge; therefore, demographic\nparity does not faithfully represent the notion of \\emph{equal treatment}. We\npropose a new formalization for equal treatment by (i) considering the\ninfluence of feature values on predictions, such as computed by Shapley values\ndecomposing predictions across its features, (ii) defining distributions of\nexplanations, and (iii) comparing explanation distributions between populations\nwith different protected characteristics. We show the theoretical properties of\nour notion of equal treatment and devise a classifier two-sample test based on\nthe AUC of an equal treatment inspector. We study our formalization of equal\ntreatment on synthetic and natural data. We release \\texttt{explanationspace},\nan open-source Python package with methods and tutorials.\n", "date": "2023-03-14", "categories": []}
{"id": "2308.11819", "title": "A Counterfactual Fair Model for Longitudinal Electronic Health Records\n  via Deconfounder", "abstract": "  The fairness issue of clinical data modeling, especially on Electronic Health\nRecords (EHRs), is of utmost importance due to EHR's complex latent structure\nand potential selection bias. It is frequently necessary to mitigate health\ndisparity while keeping the model's overall accuracy in practice. However,\ntraditional methods often encounter the trade-off between accuracy and\nfairness, as they fail to capture the underlying factors beyond observed data.\nTo tackle this challenge, we propose a novel model called Fair Longitudinal\nMedical Deconfounder (FLMD) that aims to achieve both fairness and accuracy in\nlongitudinal Electronic Health Records (EHR) modeling. Drawing inspiration from\nthe deconfounder theory, FLMD employs a two-stage training process. In the\nfirst stage, FLMD captures unobserved confounders for each encounter, which\neffectively represents underlying medical factors beyond observed EHR, such as\npatient genotypes and lifestyle habits. This unobserved confounder is crucial\nfor addressing the accuracy/fairness dilemma. In the second stage, FLMD\ncombines the learned latent representation with other relevant features to make\npredictions. By incorporating appropriate fairness criteria, such as\ncounterfactual fairness, FLMD ensures that it maintains high prediction\naccuracy while simultaneously minimizing health disparities. We conducted\ncomprehensive experiments on two real-world EHR datasets to demonstrate the\neffectiveness of FLMD. Apart from the comparison of baseline methods and FLMD\nvariants in terms of fairness and accuracy, we assessed the performance of all\nmodels on disturbed/imbalanced and synthetic datasets to showcase the\nsuperiority of FLMD across different settings and provide valuable insights\ninto its capabilities.\n", "date": "2023-08-22", "categories": []}
{"id": "2309.12491", "title": "Exploring the Impact of Training Data Distribution and Subword\n  Tokenization on Gender Bias in Machine Translation", "abstract": "  We study the effect of tokenization on gender bias in machine translation, an\naspect that has been largely overlooked in previous works. Specifically, we\nfocus on the interactions between the frequency of gendered profession names in\ntraining data, their representation in the subword tokenizer's vocabulary, and\ngender bias. We observe that female and non-stereotypical gender inflections of\nprofession names (e.g., Spanish \"doctora\" for \"female doctor\") tend to be split\ninto multiple subword tokens. Our results indicate that the imbalance of gender\nforms in the model's training corpus is a major factor contributing to gender\nbias and has a greater impact than subword splitting. We show that analyzing\nsubword splits provides good estimates of gender-form imbalance in the training\ndata and can be used even when the corpus is not publicly available. We also\ndemonstrate that fine-tuning just the token embedding layer can decrease the\ngap in gender prediction accuracy between female and male forms without\nimpairing the translation quality.\n", "date": "2023-09-21", "categories": []}
{"id": "2307.08945", "title": "Mitigating Label Bias via Decoupled Confident Learning", "abstract": "  Growing concerns regarding algorithmic fairness have led to a surge in\nmethodologies to mitigate algorithmic bias. However, such methodologies largely\nassume that observed labels in training data are correct. This is problematic\nbecause bias in labels is pervasive across important domains, including\nhealthcare, hiring, and content moderation. In particular, human-generated\nlabels are prone to encoding societal biases. While the presence of labeling\nbias has been discussed conceptually, there is a lack of methodologies to\naddress this problem. We propose a pruning method -- Decoupled Confident\nLearning (DeCoLe) -- specifically designed to mitigate label bias. After\nillustrating its performance on a synthetic dataset, we apply DeCoLe in the\ncontext of hate speech detection, where label bias has been recognized as an\nimportant challenge, and show that it successfully identifies biased labels and\noutperforms competing approaches.\n", "date": "2023-07-18", "categories": []}
{"id": "2309.17197", "title": "An Investigation Into Race Bias in Random Forest Models Based on Breast\n  DCE-MRI Derived Radiomics Features", "abstract": "  Recent research has shown that artificial intelligence (AI) models can\nexhibit bias in performance when trained using data that are imbalanced by\nprotected attribute(s). Most work to date has focused on deep learning models,\nbut classical AI techniques that make use of hand-crafted features may also be\nsusceptible to such bias. In this paper we investigate the potential for race\nbias in random forest (RF) models trained using radiomics features. Our\napplication is prediction of tumour molecular subtype from dynamic contrast\nenhanced magnetic resonance imaging (DCE-MRI) of breast cancer patients. Our\nresults show that radiomics features derived from DCE-MRI data do contain\nrace-identifiable information, and that RF models can be trained to predict\nWhite and Black race from these data with 60-70% accuracy, depending on the\nsubset of features used. Furthermore, RF models trained to predict tumour\nmolecular subtype using race-imbalanced data seem to produce biased behaviour,\nexhibiting better performance on test data from the race on which they were\ntrained.\n", "date": "2023-09-29", "categories": []}
{"id": "2309.17337", "title": "Toward Operationalizing Pipeline-aware ML Fairness: A Research Agenda\n  for Developing Practical Guidelines and Tools", "abstract": "  While algorithmic fairness is a thriving area of research, in practice,\nmitigating issues of bias often gets reduced to enforcing an arbitrarily chosen\nfairness metric, either by enforcing fairness constraints during the\noptimization step, post-processing model outputs, or by manipulating the\ntraining data. Recent work has called on the ML community to take a more\nholistic approach to tackle fairness issues by systematically investigating the\nmany design choices made through the ML pipeline, and identifying interventions\nthat target the issue's root cause, as opposed to its symptoms. While we share\nthe conviction that this pipeline-based approach is the most appropriate for\ncombating algorithmic unfairness on the ground, we believe there are currently\nvery few methods of \\emph{operationalizing} this approach in practice. Drawing\non our experience as educators and practitioners, we first demonstrate that\nwithout clear guidelines and toolkits, even individuals with specialized ML\nknowledge find it challenging to hypothesize how various design choices\ninfluence model behavior. We then consult the fair-ML literature to understand\nthe progress to date toward operationalizing the pipeline-aware approach: we\nsystematically collect and organize the prior work that attempts to detect,\nmeasure, and mitigate various sources of unfairness through the ML pipeline. We\nutilize this extensive categorization of previous contributions to sketch a\nresearch agenda for the community. We hope this work serves as the stepping\nstone toward a more comprehensive set of resources for ML researchers,\npractitioners, and students interested in exploring, designing, and testing\npipeline-oriented approaches to algorithmic fairness.\n", "date": "2023-09-29", "categories": []}
{"id": "2309.17347", "title": "Demographic Parity: Mitigating Biases in Real-World Data", "abstract": "  Computer-based decision systems are widely used to automate decisions in many\naspects of everyday life, which include sensitive areas like hiring, loaning\nand even criminal sentencing. A decision pipeline heavily relies on large\nvolumes of historical real-world data for training its models. However,\nhistorical training data often contains gender, racial or other biases which\nare propagated to the trained models influencing computer-based decisions. In\nthis work, we propose a robust methodology that guarantees the removal of\nunwanted biases while maximally preserving classification utility. Our approach\ncan always achieve this in a model-independent way by deriving from real-world\ndata the asymptotic dataset that uniquely encodes demographic parity and\nrealism. As a proof-of-principle, we deduce from public census records such an\nasymptotic dataset from which synthetic samples can be generated to train\nwell-established classifiers. Benchmarking the generalization capability of\nthese classifiers trained on our synthetic data, we confirm the absence of any\nexplicit or implicit bias in the computer-aided decision.\n", "date": "2023-09-27", "categories": []}
{"id": "2309.17430", "title": "FACTS: First Amplify Correlations and Then Slice to Discover Bias", "abstract": "  Computer vision datasets frequently contain spurious correlations between\ntask-relevant labels and (easy to learn) latent task-irrelevant attributes\n(e.g. context). Models trained on such datasets learn \"shortcuts\" and\nunderperform on bias-conflicting slices of data where the correlation does not\nhold. In this work, we study the problem of identifying such slices to inform\ndownstream bias mitigation strategies. We propose First Amplify Correlations\nand Then Slice to Discover Bias (FACTS), wherein we first amplify correlations\nto fit a simple bias-aligned hypothesis via strongly regularized empirical risk\nminimization. Next, we perform correlation-aware slicing via mixture modeling\nin bias-aligned feature space to discover underperforming data slices that\ncapture distinct correlations. Despite its simplicity, our method considerably\nimproves over prior work (by as much as 35% precision@10) in correlation bias\nidentification across a range of diverse evaluation settings. Our code is\navailable at: https://github.com/yvsriram/FACTS.\n", "date": "2023-09-29", "categories": []}
{"id": "2208.06308", "title": "Developing a Philosophical Framework for Fair Machine Learning: Lessons\n  From The Case of Algorithmic Collusion", "abstract": "  Fair machine learning research has been primarily concerned with\nclassification tasks that result in discrimination. However, as machine\nlearning algorithms are applied in new contexts the harms and injustices that\nresult are qualitatively different than those presently studied. The existing\nresearch paradigm in machine learning which develops metrics and definitions of\nfairness cannot account for these qualitatively different types of injustice.\nOne example of this is the problem of algorithmic collusion and market\nfairness. The negative consequences of algorithmic collusion affect all\nconsumers, not only particular members of a protected class. Drawing on this\ncase study, I propose an ethical framework for researchers and practitioners in\nmachine learning seeking to develop and apply fairness metrics that extends to\nnew domains. This contribution ties the development of formal metrics of\nfairness to specifically scoped normative principles. This enables fairness\nmetrics to reflect different concerns from discrimination. I conclude with the\nlimitations of my proposal and discuss promising avenues for future research.\n", "date": "2022-07-05", "categories": []}
{"id": "2309.15809", "title": "Fair Canonical Correlation Analysis", "abstract": "  This paper investigates fairness and bias in Canonical Correlation Analysis\n(CCA), a widely used statistical technique for examining the relationship\nbetween two sets of variables. We present a framework that alleviates\nunfairness by minimizing the correlation disparity error associated with\nprotected attributes. Our approach enables CCA to learn global projection\nmatrices from all data points while ensuring that these matrices yield\ncomparable correlation levels to group-specific projection matrices.\nExperimental evaluation on both synthetic and real-world datasets demonstrates\nthe efficacy of our method in reducing correlation disparity error without\ncompromising CCA accuracy.\n", "date": "2023-09-27", "categories": []}
{"id": "2309.12371", "title": "Fairness Hub Technical Briefs: AUC Gap", "abstract": "  To measure bias, we encourage teams to consider using AUC Gap: the absolute\ndifference between the highest and lowest test AUC for subgroups (e.g., gender,\nrace, SES, prior knowledge). It is agnostic to the AI/ML algorithm used and it\ncaptures the disparity in model performance for any number of subgroups, which\nenables non-binary fairness assessments such as for intersectional identity\ngroups. The teams use a wide range of AI/ML models in pursuit of a common goal\nof doubling math achievement in low-income middle schools. Ensuring that the\nmodels, which are trained on datasets collected in many different contexts, do\nnot introduce or amplify biases is important for achieving the goal. We offer\nhere a versatile and easy-to-compute measure of model bias for all the teams in\norder to create a common benchmark and an analytical basis for sharing what\nstrategies have worked for different teams.\n", "date": "2023-09-20", "categories": []}
{"id": "2309.14381", "title": "Survey of Social Bias in Vision-Language Models", "abstract": "  In recent years, the rapid advancement of machine learning (ML) models,\nparticularly transformer-based pre-trained models, has revolutionized Natural\nLanguage Processing (NLP) and Computer Vision (CV) fields. However, researchers\nhave discovered that these models can inadvertently capture and reinforce\nsocial biases present in their training datasets, leading to potential social\nharms, such as uneven resource allocation and unfair representation of specific\nsocial groups. Addressing these biases and ensuring fairness in artificial\nintelligence (AI) systems has become a critical concern in the ML community.\n  The recent introduction of pre-trained vision-and-language (VL) models in the\nemerging multimodal field demands attention to the potential social biases\npresent in these models as well. Although VL models are susceptible to social\nbias, there is a limited understanding compared to the extensive discussions on\nbias in NLP and CV. This survey aims to provide researchers with a high-level\ninsight into the similarities and differences of social bias studies in\npre-trained models across NLP, CV, and VL. By examining these perspectives, the\nsurvey aims to offer valuable guidelines on how to approach and mitigate social\nbias in both unimodal and multimodal settings. The findings and recommendations\npresented here can benefit the ML community, fostering the development of\nfairer and non-biased AI models in various applications and research endeavors.\n", "date": "2023-09-24", "categories": []}
{"id": "2303.01692", "title": "Travel Demand Forecasting: A Fair AI Approach", "abstract": "  Artificial Intelligence (AI) and machine learning have been increasingly\nadopted for travel demand forecasting. The AI-based travel demand forecasting\nmodels, though generate accurate predictions, may produce prediction biases and\nraise fairness issues. Using such biased models for decision-making may lead to\ntransportation policies that exacerbate social inequalities. However, limited\nstudies have been focused on addressing the fairness issues of these models.\nTherefore, in this study, we propose a novel methodology to develop\nfairness-aware, highly-accurate travel demand forecasting models. Particularly,\nthe proposed methodology can enhance the fairness of AI models for multiple\nprotected attributes (such as race and income) simultaneously. Specifically, we\nintroduce a new fairness regularization term, which is explicitly designed to\nmeasure the correlation between prediction accuracy and multiple protected\nattributes, into the loss function of the travel demand forecasting model. We\nconduct two case studies to evaluate the performance of the proposed\nmethodology using real-world ridesourcing-trip data in Chicago, IL and Austin,\nTX, respectively. Results highlight that our proposed methodology can\neffectively enhance fairness for multiple protected attributes while preserving\nprediction accuracy. Additionally, we have compared our methodology with three\nstate-of-the-art methods that adopt the regularization term approach, and the\nresults demonstrate that our approach significantly outperforms them in both\npreserving prediction accuracy and enhancing fairness. This study can provide\ntransportation professionals with a new tool to achieve fair and accurate\ntravel demand forecasting.\n", "date": "2023-03-03", "categories": []}
{"id": "2309.13421", "title": "Penalties and Rewards for Fair Learning in Paired Kidney Exchange\n  Programs", "abstract": "  A kidney exchange program, also called a kidney paired donation program, can\nbe viewed as a repeated, dynamic trading and allocation mechanism. This\nsuggests that a dynamic algorithm for transplant exchange selection may have\nsuperior performance in comparison to the repeated use of a static algorithm.\nWe confirm this hypothesis using a full scale simulation of the Canadian Kidney\nPaired Donation Program: learning algorithms, that attempt to learn optimal\npatient-donor weights in advance via dynamic simulations, do lead to improved\noutcomes. Specifically, our learning algorithms, designed with the objective of\nfairness (that is, equity in terms of transplant accessibility across cPRA\ngroups), also lead to an increased number of transplants and shorter average\nwaiting times. Indeed, our highest performing learning algorithm improves\negalitarian fairness by 10% whilst also increasing the number of transplants by\n6% and decreasing waiting times by 24%. However, our main result is much more\nsurprising. We find that the most critical factor in determining the\nperformance of a kidney exchange program is not the judicious assignment of\npositive weights (rewards) to patient-donor pairs. Rather, the key factor in\nincreasing the number of transplants, decreasing waiting times and improving\ngroup fairness is the judicious assignment of a negative weight (penalty) to\nthe small number of non-directed donors in the kidney exchange program.\n", "date": "2023-09-23", "categories": []}
{"id": "2309.14198", "title": "(Predictable) Performance Bias in Unsupervised Anomaly Detection", "abstract": "  Background: With the ever-increasing amount of medical imaging data, the\ndemand for algorithms to assist clinicians has amplified. Unsupervised anomaly\ndetection (UAD) models promise to aid in the crucial first step of disease\ndetection. While previous studies have thoroughly explored fairness in\nsupervised models in healthcare, for UAD, this has so far been unexplored.\n  Methods: In this study, we evaluated how dataset composition regarding\nsubgroups manifests in disparate performance of UAD models along multiple\nprotected variables on three large-scale publicly available chest X-ray\ndatasets. Our experiments were validated using two state-of-the-art UAD models\nfor medical images. Finally, we introduced a novel subgroup-AUROC (sAUROC)\nmetric, which aids in quantifying fairness in machine learning.\n  Findings: Our experiments revealed empirical \"fairness laws\" (similar to\n\"scaling laws\" for Transformers) for training-dataset composition: Linear\nrelationships between anomaly detection performance within a subpopulation and\nits representation in the training data. Our study further revealed performance\ndisparities, even in the case of balanced training data, and compound effects\nthat exacerbate the drop in performance for subjects associated with multiple\nadversely affected groups.\n  Interpretation: Our study quantified the disparate performance of UAD models\nagainst certain demographic subgroups. Importantly, we showed that this\nunfairness cannot be mitigated by balanced representation alone. Instead, the\nrepresentation of some subgroups seems harder to learn by UAD models than that\nof others. The empirical fairness laws discovered in our study make disparate\nperformance in UAD models easier to estimate and aid in determining the most\ndesirable dataset composition.\n", "date": "2023-09-25", "categories": []}
{"id": "2309.12877", "title": "FairComp: Workshop on Fairness and Robustness in Machine Learning for\n  Ubiquitous Computing", "abstract": "  How can we ensure that Ubiquitous Computing (UbiComp) research outcomes are\nboth ethical and fair? While fairness in machine learning (ML) has gained\ntraction in recent years, fairness in UbiComp remains unexplored. This workshop\naims to discuss fairness in UbiComp research and its social, technical, and\nlegal implications. From a social perspective, we will examine the relationship\nbetween fairness and UbiComp research and identify pathways to ensure that\nubiquitous technologies do not cause harm or infringe on individual rights.\nFrom a technical perspective, we will initiate a discussion on data practices\nto develop bias mitigation approaches tailored to UbiComp research. From a\nlegal perspective, we will examine how new policies shape our community's work\nand future research. We aim to foster a vibrant community centered around the\ntopic of responsible UbiComp, while also charting a clear path for future\nresearch endeavours in this field.\n", "date": "2023-09-22", "categories": []}
{"id": "2309.11503", "title": "Fairness Vs. Personalization: Towards Equity in Epistemic Utility", "abstract": "  The applications of personalized recommender systems are rapidly expanding:\nencompassing social media, online shopping, search engine results, and more.\nThese systems offer a more efficient way to navigate the vast array of items\navailable. However, alongside this growth, there has been increased recognition\nof the potential for algorithmic systems to exhibit and perpetuate biases,\nrisking unfairness in personalized domains. In this work, we explicate the\ninherent tension between personalization and conventional implementations of\nfairness. As an alternative, we propose equity to achieve fairness in the\ncontext of epistemic utility. We provide a mapping between goals and practical\nimplementations and detail policy recommendations across key stakeholders to\nforge a path towards achieving fairness in personalized systems.\n", "date": "2023-09-05", "categories": []}
{"id": "2309.11682", "title": "Dr. FERMI: A Stochastic Distributionally Robust Fair Empirical Risk\n  Minimization Framework", "abstract": "  While training fair machine learning models has been studied extensively in\nrecent years, most developed methods rely on the assumption that the training\nand test data have similar distributions. In the presence of distribution\nshifts, fair models may behave unfairly on test data. There have been some\ndevelopments for fair learning robust to distribution shifts to address this\nshortcoming. However, most proposed solutions are based on the assumption of\nhaving access to the causal graph describing the interaction of different\nfeatures. Moreover, existing algorithms require full access to data and cannot\nbe used when small batches are used (stochastic/batch implementation). This\npaper proposes the first stochastic distributionally robust fairness framework\nwith convergence guarantees that do not require knowledge of the causal graph.\nMore specifically, we formulate the fair inference in the presence of the\ndistribution shift as a distributionally robust optimization problem under\n$L_p$ norm uncertainty sets with respect to the Exponential Renyi Mutual\nInformation (ERMI) as the measure of fairness violation. We then discuss how\nthe proposed method can be implemented in a stochastic fashion. We have\nevaluated the presented framework's performance and efficiency through\nextensive experiments on real datasets consisting of distribution shifts.\n", "date": "2023-09-20", "categories": []}
{"id": "2210.16242", "title": "Differential Privacy has Bounded Impact on Fairness in Classification", "abstract": "  We theoretically study the impact of differential privacy on fairness in\nclassification. We prove that, given a class of models, popular group fairness\nmeasures are pointwise Lipschitz-continuous with respect to the parameters of\nthe model. This result is a consequence of a more general statement on accuracy\nconditioned on an arbitrary event (such as membership to a sensitive group),\nwhich may be of independent interest. We use this Lipschitz property to prove a\nnon-asymptotic bound showing that, as the number of samples increases, the\nfairness level of private models gets closer to the one of their non-private\ncounterparts. This bound also highlights the importance of the confidence\nmargin of a model on the disparate impact of differential privacy.\n", "date": "2022-10-28", "categories": []}
{"id": "2304.07408", "title": "Fairness in Visual Clustering: A Novel Transformer Clustering Approach", "abstract": "  Promoting fairness for deep clustering models in unsupervised clustering\nsettings to reduce demographic bias is a challenging goal. This is because of\nthe limitation of large-scale balanced data with well-annotated labels for\nsensitive or protected attributes. In this paper, we first evaluate demographic\nbias in deep clustering models from the perspective of cluster purity, which is\nmeasured by the ratio of positive samples within a cluster to their correlation\ndegree. This measurement is adopted as an indication of demographic bias. Then,\na novel loss function is introduced to encourage a purity consistency for all\nclusters to maintain the fairness aspect of the learned clustering model.\nMoreover, we present a novel attention mechanism, Cross-attention, to measure\ncorrelations between multiple clusters, strengthening faraway positive samples\nand improving the purity of clusters during the learning process. Experimental\nresults on a large-scale dataset with numerous attribute settings have\ndemonstrated the effectiveness of the proposed approach on both clustering\naccuracy and fairness enhancement on several sensitive attributes.\n", "date": "2023-04-14", "categories": []}
{"id": "2306.10336", "title": "Fair Causal Feature Selection", "abstract": "  Fair feature selection for classification decision tasks has recently\ngarnered significant attention from researchers. However, existing fair feature\nselection algorithms fall short of providing a full explanation of the causal\nrelationship between features and sensitive attributes, potentially impacting\nthe accuracy of fair feature identification. To address this issue, we propose\na Fair Causal Feature Selection algorithm, called FairCFS. Specifically,\nFairCFS constructs a localized causal graph that identifies the Markov blankets\nof class and sensitive variables, to block the transmission of sensitive\ninformation for selecting fair causal features. Extensive experiments on seven\npublic real-world datasets validate that FairCFS has comparable accuracy\ncompared to eight state-of-the-art feature selection algorithms, while\npresenting more superior fairness.\n", "date": "2023-06-17", "categories": []}
{"id": "2309.08760", "title": "Biased Attention: Do Vision Transformers Amplify Gender Bias More than\n  Convolutional Neural Networks?", "abstract": "  Deep neural networks used in computer vision have been shown to exhibit many\nsocial biases such as gender bias. Vision Transformers (ViTs) have become\nincreasingly popular in computer vision applications, outperforming\nConvolutional Neural Networks (CNNs) in many tasks such as image\nclassification. However, given that research on mitigating bias in computer\nvision has primarily focused on CNNs, it is important to evaluate the effect of\na different network architecture on the potential for bias amplification. In\nthis paper we therefore introduce a novel metric to measure bias in\narchitectures, Accuracy Difference. We examine bias amplification when models\nbelonging to these two architectures are used as a part of large multimodal\nmodels, evaluating the different image encoders of Contrastive Language Image\nPretraining which is an important model used in many generative models such as\nDALL-E and Stable Diffusion. Our experiments demonstrate that architecture can\nplay a role in amplifying social biases due to the different techniques\nemployed by the models for feature extraction and embedding as well as their\ndifferent learning properties. This research found that ViTs amplified gender\nbias to a greater extent than CNNs\n", "date": "2023-09-15", "categories": []}
{"id": "2309.07110", "title": "Data Augmentation via Subgroup Mixup for Improving Fairness", "abstract": "  In this work, we propose data augmentation via pairwise mixup across\nsubgroups to improve group fairness. Many real-world applications of machine\nlearning systems exhibit biases across certain groups due to\nunder-representation or training data that reflects societal biases. Inspired\nby the successes of mixup for improving classification performance, we develop\na pairwise mixup scheme to augment training data and encourage fair and\naccurate decision boundaries for all subgroups. Data augmentation for group\nfairness allows us to add new samples of underrepresented groups to balance\nsubpopulations. Furthermore, our method allows us to use the generalization\nability of mixup to improve both fairness and accuracy. We compare our proposed\nmixup to existing data augmentation and bias mitigation approaches on both\nsynthetic simulations and real-world benchmark fair classification data,\ndemonstrating that we are able to achieve fair outcomes with robust if not\nimproved accuracy.\n", "date": "2023-09-13", "categories": []}
{"id": "2309.06061", "title": "Verifiable Fairness: Privacy-preserving Computation of Fairness for\n  Machine Learning Systems", "abstract": "  Fair machine learning is a thriving and vibrant research topic. In this\npaper, we propose Fairness as a Service (FaaS), a secure, verifiable and\nprivacy-preserving protocol to computes and verify the fairness of any machine\nlearning (ML) model. In the deisgn of FaaS, the data and outcomes are\nrepresented through cryptograms to ensure privacy. Also, zero knowledge proofs\nguarantee the well-formedness of the cryptograms and underlying data. FaaS is\nmodel--agnostic and can support various fairness metrics; hence, it can be used\nas a service to audit the fairness of any ML model. Our solution requires no\ntrusted third party or private channels for the computation of the fairness\nmetric. The security guarantees and commitments are implemented in a way that\nevery step is securely transparent and verifiable from the start to the end of\nthe process. The cryptograms of all input data are publicly available for\neveryone, e.g., auditors, social activists and experts, to verify the\ncorrectness of the process. We implemented FaaS to investigate performance and\ndemonstrate the successful use of FaaS for a publicly available data set with\nthousands of entries.\n", "date": "2023-09-12", "categories": []}
{"id": "2309.05227", "title": "Detecting Natural Language Biases with Prompt-based Learning", "abstract": "  In this project, we want to explore the newly emerging field of prompt\nengineering and apply it to the downstream task of detecting LM biases. More\nconcretely, we explore how to design prompts that can indicate 4 different\ntypes of biases: (1) gender, (2) race, (3) sexual orientation, and (4)\nreligion-based. Within our project, we experiment with different manually\ncrafted prompts that can draw out the subtle biases that may be present in the\nlanguage model. We apply these prompts to multiple variations of popular and\nwell-recognized models: BERT, RoBERTa, and T5 to evaluate their biases. We\nprovide a comparative analysis of these models and assess them using a two-fold\nmethod: use human judgment to decide whether model predictions are biased and\nutilize model-level judgment (through further prompts) to understand if a model\ncan self-diagnose the biases of its own prediction.\n", "date": "2023-09-11", "categories": []}
{"id": "2309.05521", "title": "Re-formalization of Individual Fairness", "abstract": "  The notion of individual fairness is a formalization of an ethical principle,\n\"Treating like cases alike,\" which has been argued such as by Aristotle. In a\nfairness-aware machine learning context, Dwork et al. firstly formalized the\nnotion. In their formalization, a similar pair of data in an unfair space\nshould be mapped to similar positions in a fair space. We propose to\nre-formalize individual fairness by the statistical independence conditioned by\nindividuals. This re-formalization has the following merits. First, our\nformalization is compatible with that of Dwork et al. Second, our formalization\nenables to combine individual fairness with the fairness notion, equalized odds\nor sufficiency, as well as statistical parity. Third, though their\nformalization implicitly assumes a pre-process approach for making fair\nprediction, our formalization is applicable to an in-process or post-process\napproach.\n", "date": "2023-09-11", "categories": []}
{"id": "2302.11137", "title": "Fairguard: Harness Logic-based Fairness Rules in Smart Cities", "abstract": "  Smart cities operate on computational predictive frameworks that collect,\naggregate, and utilize data from large-scale sensor networks. However, these\nframeworks are prone to multiple sources of data and algorithmic bias, which\noften lead to unfair prediction results. In this work, we first demonstrate\nthat bias persists at a micro-level both temporally and spatially by studying\nreal city data from Chattanooga, TN. To alleviate the issue of such bias, we\nintroduce Fairguard, a micro-level temporal logic-based approach for fair smart\ncity policy adjustment and generation in complex temporal-spatial domains. The\nFairguard framework consists of two phases: first, we develop a static\ngenerator that is able to reduce data bias based on temporal logic conditions\nby minimizing correlations between selected attributes. Then, to ensure\nfairness in predictive algorithms, we design a dynamic component to regulate\nprediction results and generate future fair predictions by harnessing logic\nrules. Evaluations show that logic-enabled static Fairguard can effectively\nreduce the biased correlations while dynamic Fairguard can guarantee fairness\non protected groups at run-time with minimal impact on overall performance.\n", "date": "2023-02-22", "categories": []}
{"id": "2305.00147", "title": "Visualizing chest X-ray dataset biases using GANs", "abstract": "  Recent work demonstrates that images from various chest X-ray datasets\ncontain visual features that are strongly correlated with protected demographic\nattributes like race and gender. This finding raises issues of fairness, since\nsome of these factors may be used by downstream algorithms for clinical\npredictions. In this work, we propose a framework, using generative adversarial\nnetworks (GANs), to visualize what features are most different between X-rays\nbelonging to two demographic subgroups.\n", "date": "2023-04-29", "categories": []}
{"id": "2305.12178", "title": "Model Debiasing via Gradient-based Explanation on Representation", "abstract": "  Machine learning systems produce biased results towards certain demographic\ngroups, known as the fairness problem. Recent approaches to tackle this problem\nlearn a latent code (i.e., representation) through disentangled representation\nlearning and then discard the latent code dimensions correlated with sensitive\nattributes (e.g., gender). Nevertheless, these approaches may suffer from\nincomplete disentanglement and overlook proxy attributes (proxies for sensitive\nattributes) when processing real-world data, especially for unstructured data,\ncausing performance degradation in fairness and loss of useful information for\ndownstream tasks. In this paper, we propose a novel fairness framework that\nperforms debiasing with regard to both sensitive attributes and proxy\nattributes, which boosts the prediction performance of downstream task models\nwithout complete disentanglement. The main idea is to, first, leverage\ngradient-based explanation to find two model focuses, 1) one focus for\npredicting sensitive attributes and 2) the other focus for predicting\ndownstream task labels, and second, use them to perturb the latent code that\nguides the training of downstream task models towards fairness and utility\ngoals. We show empirically that our framework works with both disentangled and\nnon-disentangled representation learning methods and achieves better\nfairness-accuracy trade-off on unstructured and structured datasets than\nprevious state-of-the-art approaches.\n", "date": "2023-05-20", "categories": []}
{"id": "2309.00805", "title": "Fairness Implications of Heterogeneous Treatment Effect Estimation with\n  Machine Learning Methods in Policy-making", "abstract": "  Causal machine learning methods which flexibly generate heterogeneous\ntreatment effect estimates could be very useful tools for governments trying to\nmake and implement policy. However, as the critical artificial intelligence\nliterature has shown, governments must be very careful of unintended\nconsequences when using machine learning models. One way to try and protect\nagainst unintended bad outcomes is with AI Fairness methods which seek to\ncreate machine learning models where sensitive variables like race or gender do\nnot influence outcomes. In this paper we argue that standard AI Fairness\napproaches developed for predictive machine learning are not suitable for all\ncausal machine learning applications because causal machine learning generally\n(at least so far) uses modelling to inform a human who is the ultimate\ndecision-maker while AI Fairness approaches assume a model that is making\ndecisions directly. We define these scenarios as indirect and direct\ndecision-making respectively and suggest that policy-making is best seen as a\njoint decision where the causal machine learning model usually only has\nindirect power. We lay out a definition of fairness for this scenario - a model\nthat provides the information a decision-maker needs to accurately make a value\njudgement about just policy outcomes - and argue that the complexity of causal\nmachine learning models can make this difficult to achieve. The solution here\nis not traditional AI Fairness adjustments, but careful modelling and awareness\nof some of the decision-making biases that these methods might encourage which\nwe describe.\n", "date": "2023-09-02", "categories": []}
{"id": "2309.01780", "title": "Measuring, Interpreting, and Improving Fairness of Algorithms using\n  Causal Inference and Randomized Experiments", "abstract": "  Algorithm fairness has become a central problem for the broad adoption of\nartificial intelligence. Although the past decade has witnessed an explosion of\nexcellent work studying algorithm biases, achieving fairness in real-world AI\nproduction systems has remained a challenging task. Most existing works fail to\nexcel in practical applications since either they have conflicting measurement\ntechniques and/ or heavy assumptions, or require code-access of the production\nmodels, whereas real systems demand an easy-to-implement measurement framework\nand a systematic way to correct the detected sources of bias.\n  In this paper, we leverage recent advances in causal inference and\ninterpretable machine learning to present an algorithm-agnostic framework\n(MIIF) to Measure, Interpret, and Improve the Fairness of an algorithmic\ndecision. We measure the algorithm bias using randomized experiments, which\nenables the simultaneous measurement of disparate treatment, disparate impact,\nand economic value. Furthermore, using modern interpretability techniques, we\ndevelop an explainable machine learning model which accurately interprets and\ndistills the beliefs of a blackbox algorithm. Altogether, these techniques\ncreate a simple and powerful toolset for studying algorithm fairness,\nespecially for understanding the cost of fairness in practical applications\nlike e-commerce and targeted advertising, where industry A/B testing is already\nabundant.\n", "date": "2023-09-04", "categories": []}
{"id": "2309.02160", "title": "Bias Propagation in Federated Learning", "abstract": "  We show that participating in federated learning can be detrimental to group\nfairness. In fact, the bias of a few parties against under-represented groups\n(identified by sensitive attributes such as gender or race) can propagate\nthrough the network to all the parties in the network. We analyze and explain\nbias propagation in federated learning on naturally partitioned real-world\ndatasets. Our analysis reveals that biased parties unintentionally yet\nstealthily encode their bias in a small number of model parameters, and\nthroughout the training, they steadily increase the dependence of the global\nmodel on sensitive attributes. What is important to highlight is that the\nexperienced bias in federated learning is higher than what parties would\notherwise encounter in centralized training with a model trained on the union\nof all their data. This indicates that the bias is due to the algorithm. Our\nwork calls for auditing group fairness in federated learning and designing\nlearning algorithms that are robust to bias propagation.\n", "date": "2023-09-05", "categories": []}
{"id": "2209.14613", "title": "Fair admission risk prediction with proportional multicalibration", "abstract": "  Fair calibration is a widely desirable fairness criteria in risk prediction\ncontexts. One way to measure and achieve fair calibration is with\nmulticalibration. Multicalibration constrains calibration error among\nflexibly-defined subpopulations while maintaining overall calibration. However,\nmulticalibrated models can exhibit a higher percent calibration error among\ngroups with lower base rates than groups with higher base rates. As a result,\nit is possible for a decision-maker to learn to trust or distrust model\npredictions for specific groups. To alleviate this, we propose\n\\emph{proportional multicalibration}, a criteria that constrains the percent\ncalibration error among groups and within prediction bins. We prove that\nsatisfying proportional multicalibration bounds a model's multicalibration as\nwell its \\emph{differential calibration}, a fairness criteria that directly\nmeasures how closely a model approximates sufficiency. Therefore,\nproportionally calibrated models limit the ability of decision makers to\ndistinguish between model performance on different patient groups, which may\nmake the models more trustworthy in practice. We provide an efficient algorithm\nfor post-processing risk prediction models for proportional multicalibration\nand evaluate it empirically. We conduct simulation studies and investigate a\nreal-world application of PMC-postprocessing to prediction of emergency\ndepartment patient admissions. We observe that proportional multicalibration is\na promising criteria for controlling simultaneous measures of calibration\nfairness of a model over intersectional groups with virtually no cost in terms\nof classification performance.\n", "date": "2022-09-29", "categories": []}
{"id": "2211.00980", "title": "Balancing Utility and Fairness in Submodular Maximization (Technical\n  Report)", "abstract": "  Submodular function maximization is a fundamental combinatorial optimization\nproblem with plenty of applications -- including data summarization, influence\nmaximization, and recommendation. In many of these problems, the goal is to\nfind a solution that maximizes the average utility over all users, for each of\nwhom the utility is defined by a monotone submodular function. However, when\nthe population of users is composed of several demographic groups, another\ncritical problem is whether the utility is fairly distributed across different\ngroups. Although the \\emph{utility} and \\emph{fairness} objectives are both\ndesirable, they might contradict each other, and, to the best of our knowledge,\nlittle attention has been paid to optimizing them jointly.\n  To fill this gap, we propose a new problem called \\emph{Bicriteria Submodular\nMaximization} (BSM) to balance utility and fairness. Specifically, it requires\nfinding a fixed-size solution to maximize the utility function, subject to the\nvalue of the fairness function not being below a threshold. Since BSM is\ninapproximable within any constant factor, we focus on designing efficient\ninstance-dependent approximation schemes. Our algorithmic proposal comprises\ntwo methods, with different approximation factors, obtained by converting a BSM\ninstance into other submodular optimization problem instances. Using real-world\nand synthetic datasets, we showcase applications of our proposed methods in\nthree submodular maximization problems: maximum coverage, influence\nmaximization, and facility location.\n", "date": "2022-11-02", "categories": []}
{"id": "2306.09995", "title": "Fairness in Preference-based Reinforcement Learning", "abstract": "  In this paper, we address the issue of fairness in preference-based\nreinforcement learning (PbRL) in the presence of multiple objectives. The main\nobjective is to design control policies that can optimize multiple objectives\nwhile treating each objective fairly. Toward this objective, we design a new\nfairness-induced preference-based reinforcement learning or FPbRL. The main\nidea of FPbRL is to learn vector reward functions associated with multiple\nobjectives via new welfare-based preferences rather than reward-based\npreference in PbRL, coupled with policy learning via maximizing a generalized\nGini welfare function. Finally, we provide experiment studies on three\ndifferent environments to show that the proposed FPbRL approach can achieve\nboth efficiency and equity for learning effective and fair policies.\n", "date": "2023-06-16", "categories": []}
{"id": "2307.09162", "title": "Unveiling Gender Bias in Terms of Profession Across LLMs: Analyzing and\n  Addressing Sociological Implications", "abstract": "  Gender bias in artificial intelligence (AI) and natural language processing\nhas garnered significant attention due to its potential impact on societal\nperceptions and biases. This research paper aims to analyze gender bias in\nLarge Language Models (LLMs) with a focus on multiple comparisons between GPT-2\nand GPT-3.5, some prominent language models, to better understand its\nimplications. Through a comprehensive literature review, the study examines\nexisting research on gender bias in AI language models and identifies gaps in\nthe current knowledge. The methodology involves collecting and preprocessing\ndata from GPT-2 and GPT-3.5, and employing in-depth quantitative analysis\ntechniques to evaluate gender bias in the generated text. The findings shed\nlight on gendered word associations, language usage, and biased narratives\npresent in the outputs of these Large Language Models. The discussion explores\nthe ethical implications of gender bias and its potential consequences on\nsocial perceptions and marginalized communities. Additionally, the paper\npresents strategies for reducing gender bias in LLMs, including algorithmic\napproaches and data augmentation techniques. The research highlights the\nimportance of interdisciplinary collaborations and the role of sociological\nstudies in mitigating gender bias in AI models. By addressing these issues, we\ncan pave the way for more inclusive and unbiased AI systems that have a\npositive impact on society.\n", "date": "2023-07-18", "categories": []}
{"id": "2309.00035", "title": "FACET: Fairness in Computer Vision Evaluation Benchmark", "abstract": "  Computer vision models have known performance disparities across attributes\nsuch as gender and skin tone. This means during tasks such as classification\nand detection, model performance differs for certain classes based on the\ndemographics of the people in the image. These disparities have been shown to\nexist, but until now there has not been a unified approach to measure these\ndifferences for common use-cases of computer vision models. We present a new\nbenchmark named FACET (FAirness in Computer Vision EvaluaTion), a large,\npublicly available evaluation set of 32k images for some of the most common\nvision tasks - image classification, object detection and segmentation. For\nevery image in FACET, we hired expert reviewers to manually annotate\nperson-related attributes such as perceived skin tone and hair type, manually\ndraw bounding boxes and label fine-grained person-related classes such as disk\njockey or guitarist. In addition, we use FACET to benchmark state-of-the-art\nvision models and present a deeper understanding of potential performance\ndisparities and challenges across sensitive demographic attributes. With the\nexhaustive annotations collected, we probe models using single demographics\nattributes as well as multiple attributes using an intersectional approach\n(e.g. hair color and perceived skin tone). Our results show that\nclassification, detection, segmentation, and visual grounding models exhibit\nperformance disparities across demographic attributes and intersections of\nattributes. These harms suggest that not all people represented in datasets\nreceive fair and equitable treatment in these vision tasks. We hope current and\nfuture results using our benchmark will contribute to fairer, more robust\nvision models. FACET is available publicly at https://facet.metademolab.com/\n", "date": "2023-08-31", "categories": []}
{"id": "2302.12977", "title": "Fair Attribute Completion on Graph with Missing Attributes", "abstract": "  Tackling unfairness in graph learning models is a challenging task, as the\nunfairness issues on graphs involve both attributes and topological structures.\nExisting work on fair graph learning simply assumes that attributes of all\nnodes are available for model training and then makes fair predictions. In\npractice, however, the attributes of some nodes might not be accessible due to\nmissing data or privacy concerns, which makes fair graph learning even more\nchallenging. In this paper, we propose FairAC, a fair attribute completion\nmethod, to complement missing information and learn fair node embeddings for\ngraphs with missing attributes. FairAC adopts an attention mechanism to deal\nwith the attribute missing problem and meanwhile, it mitigates two types of\nunfairness, i.e., feature unfairness from attributes and topological unfairness\ndue to attribute completion. FairAC can work on various types of homogeneous\ngraphs and generate fair embeddings for them and thus can be applied to most\ndownstream tasks to improve their fairness performance. To our best knowledge,\nFairAC is the first method that jointly addresses the graph attribution\ncompletion and graph unfairness problems. Experimental results on benchmark\ndatasets show that our method achieves better fairness performance with less\nsacrifice in accuracy, compared with the state-of-the-art methods of fair graph\nlearning. Code is available at: https://github.com/donglgcn/FairAC.\n", "date": "2023-02-25", "categories": []}
{"id": "2207.03364", "title": "Group Equality in Adaptive Submodular Maximization", "abstract": "  In this paper, we study the classic submodular maximization problem subject\nto a group equality constraint under both non-adaptive and adaptive settings.\nIt has been shown that the utility function of many machine learning\napplications, including data summarization, influence maximization in social\nnetworks, and personalized recommendation, satisfies the property of\nsubmodularity. Hence, maximizing a submodular function subject to various\nconstraints can be found at the heart of many of those applications. On a high\nlevel, submodular maximization aims to select a group of most representative\nitems (e.g., data points). However, the design of most existing algorithms does\nnot incorporate the fairness constraint, leading to under- or\nover-representation of some particular groups. This motivates us to study the\nsubmodular maximization problem with group equality, where we aim to select a\ngroup of items to maximize a (possibly non-monotone) submodular utility\nfunction subject to a group equality constraint. To this end, we develop the\nfirst constant-factor approximation algorithm for this problem. The design of\nour algorithm is robust enough to be extended to solving the submodular\nmaximization problem under a more complicated adaptive setting. Moreover, we\nfurther extend our study to incorporating a global cardinality constraint and\nother fairness notations.\n", "date": "2022-07-07", "categories": []}
{"id": "2306.09297", "title": "Fix Fairness, Don't Ruin Accuracy: Performance Aware Fairness Repair\n  using AutoML", "abstract": "  Machine learning (ML) is increasingly being used in critical decision-making\nsoftware, but incidents have raised questions about the fairness of ML\npredictions. To address this issue, new tools and methods are needed to\nmitigate bias in ML-based software. Previous studies have proposed bias\nmitigation algorithms that only work in specific situations and often result in\na loss of accuracy. Our proposed solution is a novel approach that utilizes\nautomated machine learning (AutoML) techniques to mitigate bias. Our approach\nincludes two key innovations: a novel optimization function and a\nfairness-aware search space. By improving the default optimization function of\nAutoML and incorporating fairness objectives, we are able to mitigate bias with\nlittle to no loss of accuracy. Additionally, we propose a fairness-aware search\nspace pruning method for AutoML to reduce computational cost and repair time.\nOur approach, built on the state-of-the-art Auto-Sklearn tool, is designed to\nreduce bias in real-world scenarios. In order to demonstrate the effectiveness\nof our approach, we evaluated our approach on four fairness problems and 16\ndifferent ML models, and our results show a significant improvement over the\nbaseline and existing bias mitigation techniques. Our approach, Fair-AutoML,\nsuccessfully repaired 60 out of 64 buggy cases, while existing bias mitigation\ntechniques only repaired up to 44 out of 64 cases.\n", "date": "2023-06-15", "categories": []}
{"id": "2308.14921", "title": "Gender bias and stereotypes in Large Language Models", "abstract": "  Large Language Models (LLMs) have made substantial progress in the past\nseveral months, shattering state-of-the-art benchmarks in many domains. This\npaper investigates LLMs' behavior with respect to gender stereotypes, a known\nissue for prior models. We use a simple paradigm to test the presence of gender\nbias, building on but differing from WinoBias, a commonly used gender bias\ndataset, which is likely to be included in the training data of current LLMs.\nWe test four recently published LLMs and demonstrate that they express biased\nassumptions about men and women's occupations. Our contributions in this paper\nare as follows: (a) LLMs are 3-6 times more likely to choose an occupation that\nstereotypically aligns with a person's gender; (b) these choices align with\npeople's perceptions better than with the ground truth as reflected in official\njob statistics; (c) LLMs in fact amplify the bias beyond what is reflected in\nperceptions or the ground truth; (d) LLMs ignore crucial ambiguities in\nsentence structure 95% of the time in our study items, but when explicitly\nprompted, they recognize the ambiguity; (e) LLMs provide explanations for their\nchoices that are factually inaccurate and likely obscure the true reason behind\ntheir predictions. That is, they provide rationalizations of their biased\nbehavior. This highlights a key property of these models: LLMs are trained on\nimbalanced datasets; as such, even with the recent successes of reinforcement\nlearning with human feedback, they tend to reflect those imbalances back at us.\nAs with other types of societal biases, we suggest that LLMs must be carefully\ntested to ensure that they treat minoritized individuals and communities\nequitably.\n", "date": "2023-08-28", "categories": []}
{"id": "2308.15243", "title": "Reliability Gaps Between Groups in COMPAS Dataset", "abstract": "  This paper investigates the inter-rater reliability of risk assessment\ninstruments (RAIs). The main question is whether different, socially salient\ngroups are affected differently by a lack of inter-rater reliability of RAIs,\nthat is, whether mistakes with respect to different groups affects them\ndifferently. The question is investigated with a simulation study of the COMPAS\ndataset. A controlled degree of noise is injected into the input data of a\npredictive model; the noise can be interpreted as a synthetic rater that makes\nmistakes. The main finding is that there are systematic differences in output\nreliability between groups in the COMPAS dataset. The sign of the difference\ndepends on the kind of inter-rater statistic that is used (Cohen's Kappa,\nByrt's PABAK, ICC), and in particular whether or not a correction of\npredictions prevalences of the groups is used.\n", "date": "2023-08-29", "categories": []}
{"id": "2302.08204", "title": "Counterfactual Reasoning for Bias Evaluation and Detection in a Fairness\n  under Unawareness setting", "abstract": "  Current AI regulations require discarding sensitive features (e.g., gender,\nrace, religion) in the algorithm's decision-making process to prevent unfair\noutcomes. However, even without sensitive features in the training set,\nalgorithms can persist in discrimination. Indeed, when sensitive features are\nomitted (fairness under unawareness), they could be inferred through non-linear\nrelations with the so called proxy features. In this work, we propose a way to\nreveal the potential hidden bias of a machine learning model that can persist\neven when sensitive features are discarded. This study shows that it is\npossible to unveil whether the black-box predictor is still biased by\nexploiting counterfactual reasoning. In detail, when the predictor provides a\nnegative classification outcome, our approach first builds counterfactual\nexamples for a discriminated user category to obtain a positive outcome. Then,\nthe same counterfactual samples feed an external classifier (that targets a\nsensitive feature) that reveals whether the modifications to the user\ncharacteristics needed for a positive outcome moved the individual to the\nnon-discriminated group. When this occurs, it could be a warning sign for\ndiscriminatory behavior in the decision process. Furthermore, we leverage the\ndeviation of counterfactuals from the original sample to determine which\nfeatures are proxies of specific sensitive information. Our experiments show\nthat, even if the model is trained without sensitive features, it often suffers\ndiscriminatory biases.\n", "date": "2023-02-16", "categories": []}
{"id": "2308.13730", "title": "Muffin: A Framework Toward Multi-Dimension AI Fairness by Uniting\n  Off-the-Shelf Models", "abstract": "  Model fairness (a.k.a., bias) has become one of the most critical problems in\na wide range of AI applications. An unfair model in autonomous driving may\ncause a traffic accident if corner cases (e.g., extreme weather) cannot be\nfairly regarded; or it will incur healthcare disparities if the AI model\nmisdiagnoses a certain group of people (e.g., brown and black skin). In recent\nyears, there have been emerging research works on addressing unfairness, and\nthey mainly focus on a single unfair attribute, like skin tone; however,\nreal-world data commonly have multiple attributes, among which unfairness can\nexist in more than one attribute, called 'multi-dimensional fairness'. In this\npaper, we first reveal a strong correlation between the different unfair\nattributes, i.e., optimizing fairness on one attribute will lead to the\ncollapse of others. Then, we propose a novel Multi-Dimension Fairness\nframework, namely Muffin, which includes an automatic tool to unite\noff-the-shelf models to improve the fairness on multiple attributes\nsimultaneously. Case studies on dermatology datasets with two unfair attributes\nshow that the existing approach can achieve 21.05% fairness improvement on the\nfirst attribute while it makes the second attribute unfair by 1.85%. On the\nother hand, the proposed Muffin can unite multiple models to achieve\nsimultaneously 26.32% and 20.37% fairness improvement on both attributes;\nmeanwhile, it obtains 5.58% accuracy gain.\n", "date": "2023-08-26", "categories": []}
{"id": "2308.14338", "title": "Fair Few-shot Learning with Auxiliary Sets", "abstract": "  Recently, there has been a growing interest in developing machine learning\n(ML) models that can promote fairness, i.e., eliminating biased predictions\ntowards certain populations (e.g., individuals from a specific demographic\ngroup). Most existing works learn such models based on well-designed fairness\nconstraints in optimization. Nevertheless, in many practical ML tasks, only\nvery few labeled data samples can be collected, which can lead to inferior\nfairness performance. This is because existing fairness constraints are\ndesigned to restrict the prediction disparity among different sensitive groups,\nbut with few samples, it becomes difficult to accurately measure the disparity,\nthus rendering ineffective fairness optimization. In this paper, we define the\nfairness-aware learning task with limited training samples as the \\emph{fair\nfew-shot learning} problem. To deal with this problem, we devise a novel\nframework that accumulates fairness-aware knowledge across different\nmeta-training tasks and then generalizes the learned knowledge to meta-test\ntasks. To compensate for insufficient training samples, we propose an essential\nstrategy to select and leverage an auxiliary set for each meta-test task. These\nauxiliary sets contain several labeled training samples that can enhance the\nmodel performance regarding fairness in meta-test tasks, thereby allowing for\nthe transfer of learned useful fairness-oriented knowledge to meta-test tasks.\nFurthermore, we conduct extensive experiments on three real-world datasets to\nvalidate the superiority of our framework against the state-of-the-art\nbaselines.\n", "date": "2023-08-28", "categories": []}
{"id": "2304.09270", "title": "Coarse race data conceals disparities in clinical risk score performance", "abstract": "  Healthcare data in the United States often records only a patient's coarse\nrace group: for example, both Indian and Chinese patients are typically coded\nas \"Asian.\" It is unknown, however, whether this coarse coding conceals\nmeaningful disparities in the performance of clinical risk scores across\ngranular race groups. Here we show that it does. Using data from 418K emergency\ndepartment visits, we assess clinical risk score performance disparities across\n26 granular groups for three outcomes, five risk scores, and four performance\nmetrics. Across outcomes and metrics, we show that the risk scores exhibit\nsignificant granular performance disparities within coarse race groups. In\nfact, variation in performance within coarse groups often *exceeds* the\nvariation between coarse groups. We explore why these disparities arise,\nfinding that outcome rates, feature distributions, and the relationships\nbetween features and outcomes all vary significantly across granular groups.\nOur results suggest that healthcare providers, hospital systems, and machine\nlearning researchers should strive to collect, release, and use granular race\ndata in place of coarse race data, and that existing analyses may significantly\nunderestimate racial disparities in performance.\n", "date": "2023-04-18", "categories": []}
{"id": "2308.13089", "title": "Towards a Holistic Approach: Understanding Sociodemographic Biases in\n  NLP Models using an Interdisciplinary Lens", "abstract": "  The rapid growth in the usage and applications of Natural Language Processing\n(NLP) in various sociotechnical solutions has highlighted the need for a\ncomprehensive understanding of bias and its impact on society. While research\non bias in NLP has expanded, several challenges persist that require attention.\nThese include the limited focus on sociodemographic biases beyond race and\ngender, the narrow scope of analysis predominantly centered on models, and the\ntechnocentric implementation approaches. This paper addresses these challenges\nand advocates for a more interdisciplinary approach to understanding bias in\nNLP. The work is structured into three facets, each exploring a specific aspect\nof bias in NLP.\n", "date": "2023-08-24", "categories": []}
{"id": "2308.13242", "title": "Optimizing Group-Fair Plackett-Luce Ranking Models for Relevance and\n  Ex-Post Fairness", "abstract": "  In learning-to-rank (LTR), optimizing only the relevance (or the expected\nranking utility) can cause representational harm to certain categories of\nitems. Moreover, if there is implicit bias in the relevance scores, LTR models\nmay fail to optimize for true relevance. Previous works have proposed efficient\nalgorithms to train stochastic ranking models that achieve fairness of exposure\nto the groups ex-ante (or, in expectation), which may not guarantee\nrepresentation fairness to the groups ex-post, that is, after realizing a\nranking from the stochastic ranking model. Typically, ex-post fairness is\nachieved by post-processing, but previous work does not train stochastic\nranking models that are aware of this post-processing.\n  In this paper, we propose a novel objective that maximizes expected relevance\nonly over those rankings that satisfy given representation constraints to\nensure ex-post fairness. Building upon recent work on an efficient sampler for\nex-post group-fair rankings, we propose a group-fair Plackett-Luce model and\nshow that it can be efficiently optimized for our objective in the LTR\nframework.\n  Experiments on three real-world datasets show that our group-fair algorithm\nguarantees fairness alongside usually having better relevance compared to the\nLTR baselines. In addition, our algorithm also achieves better relevance than\npost-processing baselines, which also ensures ex-post fairness. Further, when\nimplicit bias is injected into the training data, our algorithm typically\noutperforms existing LTR baselines in relevance.\n", "date": "2023-08-25", "categories": []}
{"id": "2308.08482", "title": "Benign Shortcut for Debiasing: Fair Visual Recognition via Intervention\n  with Shortcut Features", "abstract": "  Machine learning models often learn to make predictions that rely on\nsensitive social attributes like gender and race, which poses significant\nfairness risks, especially in societal applications, such as hiring, banking,\nand criminal justice. Existing work tackles this issue by minimizing the\nemployed information about social attributes in models for debiasing. However,\nthe high correlation between target task and these social attributes makes\nlearning on the target task incompatible with debiasing. Given that model bias\narises due to the learning of bias features (\\emph{i.e}., gender) that help\ntarget task optimization, we explore the following research question: \\emph{Can\nwe leverage shortcut features to replace the role of bias feature in target\ntask optimization for debiasing?} To this end, we propose \\emph{Shortcut\nDebiasing}, to first transfer the target task's learning of bias attributes\nfrom bias features to shortcut features, and then employ causal intervention to\neliminate shortcut features during inference. The key idea of \\emph{Shortcut\nDebiasing} is to design controllable shortcut features to on one hand replace\nbias features in contributing to the target task during the training stage, and\non the other hand be easily removed by intervention during the inference stage.\nThis guarantees the learning of the target task does not hinder the elimination\nof bias features. We apply \\emph{Shortcut Debiasing} to several benchmark\ndatasets, and achieve significant improvements over the state-of-the-art\ndebiasing methods in both accuracy and fairness.\n", "date": "2023-08-13", "categories": []}
{"id": "2302.08973", "title": "Measuring Equality in Machine Learning Security Defenses: A Case Study\n  in Speech Recognition", "abstract": "  Over the past decade, the machine learning security community has developed a\nmyriad of defenses for evasion attacks. An understudied question in that\ncommunity is: for whom do these defenses defend? This work considers common\napproaches to defending learned systems and how security defenses result in\nperformance inequities across different sub-populations. We outline appropriate\nparity metrics for analysis and begin to answer this question through empirical\nresults of the fairness implications of machine learning security methods. We\nfind that many methods that have been proposed can cause direct harm, like\nfalse rejection and unequal benefits from robustness training. The framework we\npropose for measuring defense equality can be applied to robustly trained\nmodels, preprocessing-based defenses, and rejection methods. We identify a set\nof datasets with a user-centered application and a reasonable computational\ncost suitable for case studies in measuring the equality of defenses. In our\ncase study of speech command recognition, we show how such adversarial training\nand augmentation have non-equal but complex protections for social subgroups\nacross gender, accent, and age in relation to user coverage. We present a\ncomparison of equality between two rejection-based defenses: randomized\nsmoothing and neural rejection, finding randomized smoothing more equitable due\nto the sampling mechanism for minority groups. This represents the first work\nexamining the disparity in the adversarial robustness in the speech domain and\nthe fairness evaluation of rejection-based defenses.\n", "date": "2023-02-17", "categories": []}
{"id": "2305.03881", "title": "Fairness in Image Search: A Study of Occupational Stereotyping in Image\n  Retrieval and its Debiasing", "abstract": "  Multi-modal search engines have experienced significant growth and widespread\nuse in recent years, making them the second most common internet use. While\nsearch engine systems offer a range of services, the image search field has\nrecently become a focal point in the information retrieval community, as the\nadage goes, \"a picture is worth a thousand words\". Although popular search\nengines like Google excel at image search accuracy and agility, there is an\nongoing debate over whether their search results can be biased in terms of\ngender, language, demographics, socio-cultural aspects, and stereotypes. This\npotential for bias can have a significant impact on individuals' perceptions\nand influence their perspectives.\n  In this paper, we present our study on bias and fairness in web search, with\na focus on keyword-based image search. We first discuss several kinds of biases\nthat exist in search systems and why it is important to mitigate them. We\nnarrow down our study to assessing and mitigating occupational stereotypes in\nimage search, which is a prevalent fairness issue in image retrieval. For the\nassessment of stereotypes, we take gender as an indicator. We explore various\nopen-source and proprietary APIs for gender identification from images. With\nthese, we examine the extent of gender bias in top-tanked image search results\nobtained for several occupational keywords. To mitigate the bias, we then\npropose a fairness-aware re-ranking algorithm that optimizes (a) relevance of\nthe search result with the keyword and (b) fairness w.r.t genders identified.\nWe experiment on 100 top-ranked images obtained for 10 occupational keywords\nand consider random re-ranking and re-ranking based on relevance as baselines.\nOur experimental results show that the fairness-aware re-ranking algorithm\nproduces rankings with better fairness scores and competitive relevance scores\nthan the baselines.\n", "date": "2023-05-06", "categories": []}
{"id": "2308.11386", "title": "Targeted Data Augmentation for bias mitigation", "abstract": "  The development of fair and ethical AI systems requires careful consideration\nof bias mitigation, an area often overlooked or ignored. In this study, we\nintroduce a novel and efficient approach for addressing biases called Targeted\nData Augmentation (TDA), which leverages classical data augmentation techniques\nto tackle the pressing issue of bias in data and models. Unlike the laborious\ntask of removing biases, our method proposes to insert biases instead,\nresulting in improved performance. To identify biases, we annotated two diverse\ndatasets: a dataset of clinical skin lesions and a dataset of male and female\nfaces. These bias annotations are published for the first time in this study,\nproviding a valuable resource for future research. Through Counterfactual Bias\nInsertion, we discovered that biases associated with the frame, ruler, and\nglasses had a significant impact on models. By randomly introducing biases\nduring training, we mitigated these biases and achieved a substantial decrease\nin bias measures, ranging from two-fold to more than 50-fold, while maintaining\na negligible increase in the error rate.\n", "date": "2023-08-22", "categories": []}
{"id": "2307.04937", "title": "Towards Fair Graph Neural Networks via Graph Counterfactual", "abstract": "  Graph neural networks have shown great ability in representation (GNNs)\nlearning on graphs, facilitating various tasks. Despite their great performance\nin modeling graphs, recent works show that GNNs tend to inherit and amplify the\nbias from training data, causing concerns of the adoption of GNNs in high-stake\nscenarios. Hence, many efforts have been taken for fairness-aware GNNs.\nHowever, most existing fair GNNs learn fair node representations by adopting\nstatistical fairness notions, which may fail to alleviate bias in the presence\nof statistical anomalies. Motivated by causal theory, there are several\nattempts utilizing graph counterfactual fairness to mitigate root causes of\nunfairness. However, these methods suffer from non-realistic counterfactuals\nobtained by perturbation or generation. In this paper, we take a causal view on\nfair graph learning problem. Guided by the casual analysis, we propose a novel\nframework CAF, which can select counterfactuals from training data to avoid\nnon-realistic counterfactuals and adopt selected counterfactuals to learn fair\nnode representations for node classification task. Extensive experiments on\nsynthetic and real-world datasets show the effectiveness of CAF. Our code is\navailable at https://github.com/TimeLovercc/CAF-GNN.\n", "date": "2023-07-10", "categories": []}
{"id": "2308.09726", "title": "Equitable Restless Multi-Armed Bandits: A General Framework Inspired By\n  Digital Health", "abstract": "  Restless multi-armed bandits (RMABs) are a popular framework for algorithmic\ndecision making in sequential settings with limited resources. RMABs are\nincreasingly being used for sensitive decisions such as in public health,\ntreatment scheduling, anti-poaching, and -- the motivation for this work --\ndigital health. For such high stakes settings, decisions must both improve\noutcomes and prevent disparities between groups (e.g., ensure health equity).\nWe study equitable objectives for RMABs (ERMABs) for the first time. We\nconsider two equity-aligned objectives from the fairness literature, minimax\nreward and max Nash welfare. We develop efficient algorithms for solving each\n-- a water filling algorithm for the former, and a greedy algorithm with\ntheoretically motivated nuance to balance disparate group sizes for the latter.\nFinally, we demonstrate across three simulation domains, including a new\ndigital health model, that our approaches can be multiple times more equitable\nthan the current state of the art without drastic sacrifices to utility. Our\nfindings underscore our work's urgency as RMABs permeate into systems that\nimpact human and wildlife outcomes. Code is available at\nhttps://github.com/google-research/socialgood/tree/equitable-rmab\n", "date": "2023-08-17", "categories": []}
{"id": "2308.10005", "title": "Partition-and-Debias: Agnostic Biases Mitigation via A Mixture of\n  Biases-Specific Experts", "abstract": "  Bias mitigation in image classification has been widely researched, and\nexisting methods have yielded notable results. However, most of these methods\nimplicitly assume that a given image contains only one type of known or unknown\nbias, failing to consider the complexities of real-world biases. We introduce a\nmore challenging scenario, agnostic biases mitigation, aiming at bias removal\nregardless of whether the type of bias or the number of types is unknown in the\ndatasets. To address this difficult task, we present the Partition-and-Debias\n(PnD) method that uses a mixture of biases-specific experts to implicitly\ndivide the bias space into multiple subspaces and a gating module to find a\nconsensus among experts to achieve debiased classification. Experiments on both\npublic and constructed benchmarks demonstrated the efficacy of the PnD. Code is\navailable at: https://github.com/Jiaxuan-Li/PnD.\n", "date": "2023-08-19", "categories": []}
{"id": "2303.06167", "title": "Overwriting Pretrained Bias with Finetuning Data", "abstract": "  Transfer learning is beneficial by allowing the expressive features of models\npretrained on large-scale datasets to be finetuned for the target task of\nsmaller, more domain-specific datasets. However, there is a concern that these\npretrained models may come with their own biases which would propagate into the\nfinetuned model. In this work, we investigate bias when conceptualized as both\nspurious correlations between the target task and a sensitive attribute as well\nas underrepresentation of a particular group in the dataset. Under both notions\nof bias, we find that (1) models finetuned on top of pretrained models can\nindeed inherit their biases, but (2) this bias can be corrected for through\nrelatively minor interventions to the finetuning dataset, and often with a\nnegligible impact to performance. Our findings imply that careful curation of\nthe finetuning dataset is important for reducing biases on a downstream task,\nand doing so can even compensate for bias in the pretrained model.\n", "date": "2023-03-10", "categories": []}
{"id": "2308.08938", "title": "Causal Adversarial Perturbations for Individual Fairness and Robustness\n  in Heterogeneous Data Spaces", "abstract": "  As responsible AI gains importance in machine learning algorithms, properties\nsuch as fairness, adversarial robustness, and causality have received\nconsiderable attention in recent years. However, despite their individual\nsignificance, there remains a critical gap in simultaneously exploring and\nintegrating these properties. In this paper, we propose a novel approach that\nexamines the relationship between individual fairness, adversarial robustness,\nand structural causal models in heterogeneous data spaces, particularly when\ndealing with discrete sensitive attributes. We use causal structural models and\nsensitive attributes to create a fair metric and apply it to measure semantic\nsimilarity among individuals. By introducing a novel causal adversarial\nperturbation and applying adversarial training, we create a new regularizer\nthat combines individual fairness, causality, and robustness in the classifier.\nOur method is evaluated on both real-world and synthetic datasets,\ndemonstrating its effectiveness in achieving an accurate classifier that\nsimultaneously exhibits fairness, adversarial robustness, and causal awareness.\n", "date": "2023-08-17", "categories": []}
{"id": "2308.09464", "title": "Data augmentation and explainability for bias discovery and mitigation\n  in deep learning", "abstract": "  This dissertation explores the impact of bias in deep neural networks and\npresents methods for reducing its influence on model performance. The first\npart begins by categorizing and describing potential sources of bias and errors\nin data and models, with a particular focus on bias in machine learning\npipelines. The next chapter outlines a taxonomy and methods of Explainable AI\nas a way to justify predictions and control and improve the model. Then, as an\nexample of a laborious manual data inspection and bias discovery process, a\nskin lesion dataset is manually examined. A Global Explanation for the Bias\nIdentification method is proposed as an alternative semi-automatic approach to\nmanual data exploration for discovering potential biases in data. Relevant\nnumerical methods and metrics are discussed for assessing the effects of the\nidentified biases on the model. Whereas identifying errors and bias is\ncritical, improving the model and reducing the number of flaws in the future is\nan absolute priority. Hence, the second part of the thesis focuses on\nmitigating the influence of bias on ML models. Three approaches are proposed\nand discussed: Style Transfer Data Augmentation, Targeted Data Augmentations,\nand Attribution Feedback. Style Transfer Data Augmentation aims to address\nshape and texture bias by merging a style of a malignant lesion with a\nconflicting shape of a benign one. Targeted Data Augmentations randomly insert\npossible biases into all images in the dataset during the training, as a way to\nmake the process random and, thus, destroy spurious correlations. Lastly,\nAttribution Feedback is used to fine-tune the model to improve its accuracy by\neliminating obvious mistakes and teaching it to ignore insignificant input\nparts via an attribution loss. The goal of these approaches is to reduce the\ninfluence of bias on machine learning models, rather than eliminate it\nentirely.\n", "date": "2023-08-18", "categories": []}
{"id": "2308.09596", "title": "Disparity, Inequality, and Accuracy Tradeoffs in Graph Neural Networks\n  for Node Classification", "abstract": "  Graph neural networks (GNNs) are increasingly used in critical human\napplications for predicting node labels in attributed graphs. Their ability to\naggregate features from nodes' neighbors for accurate classification also has\nthe capacity to exacerbate existing biases in data or to introduce new ones\ntowards members from protected demographic groups. Thus, it is imperative to\nquantify how GNNs may be biased and to what extent their harmful effects may be\nmitigated. To this end, we propose two new GNN-agnostic interventions namely,\n(i) PFR-AX which decreases the separability between nodes in protected and\nnon-protected groups, and (ii) PostProcess which updates model predictions\nbased on a blackbox policy to minimize differences between error rates across\ndemographic groups. Through a large set of experiments on four datasets, we\nframe the efficacies of our approaches (and three variants) in terms of their\nalgorithmic fairness-accuracy tradeoff and benchmark our results against three\nstrong baseline interventions on three state-of-the-art GNN models. Our results\nshow that no single intervention offers a universally optimal tradeoff, but\nPFR-AX and PostProcess provide granular control and improve model confidence\nwhen correctly predicting positive outcomes for nodes in protected groups.\n", "date": "2023-08-18", "categories": []}
{"id": "2308.09640", "title": "Revisiting Skin Tone Fairness in Dermatological Lesion Classification", "abstract": "  Addressing fairness in lesion classification from dermatological images is\ncrucial due to variations in how skin diseases manifest across skin tones.\nHowever, the absence of skin tone labels in public datasets hinders building a\nfair classifier. To date, such skin tone labels have been estimated prior to\nfairness analysis in independent studies using the Individual Typology Angle\n(ITA). Briefly, ITA calculates an angle based on pixels extracted from skin\nimages taking into account the lightness and yellow-blue tints. These angles\nare then categorised into skin tones that are subsequently used to analyse\nfairness in skin cancer classification. In this work, we review and compare\nfour ITA-based approaches of skin tone classification on the ISIC18 dataset, a\ncommon benchmark for assessing skin cancer classification fairness in the\nliterature. Our analyses reveal a high disagreement among previously published\nstudies demonstrating the risks of ITA-based skin tone estimation methods.\nMoreover, we investigate the causes of such large discrepancy among these\napproaches and find that the lack of diversity in the ISIC18 dataset limits its\nuse as a testbed for fairness analysis. Finally, we recommend further research\non robust ITA estimation and diverse dataset acquisition with skin tone\nannotation to facilitate conclusive fairness assessments of artificial\nintelligence tools in dermatology. Our code is available at\nhttps://github.com/tkalbl/RevisitingSkinToneFairness.\n", "date": "2023-08-18", "categories": []}
{"id": "2304.03646", "title": "Fairness through Aleatoric Uncertainty", "abstract": "  We propose a simple yet effective solution to tackle the often-competing\ngoals of fairness and utility in classification tasks. While fairness ensures\nthat the model's predictions are unbiased and do not discriminate against any\nparticular group or individual, utility focuses on maximizing the model's\npredictive performance. This work introduces the idea of leveraging aleatoric\nuncertainty (e.g., data ambiguity) to improve the fairness-utility trade-off.\nOur central hypothesis is that aleatoric uncertainty is a key factor for\nalgorithmic fairness and samples with low aleatoric uncertainty are modeled\nmore accurately and fairly than those with high aleatoric uncertainty. We then\npropose a principled model to improve fairness when aleatoric uncertainty is\nhigh and improve utility elsewhere. Our approach first intervenes in the data\ndistribution to better decouple aleatoric uncertainty and epistemic\nuncertainty. It then introduces a fairness-utility bi-objective loss defined\nbased on the estimated aleatoric uncertainty. Our approach is theoretically\nguaranteed to improve the fairness-utility trade-off. Experimental results on\nboth tabular and image datasets show that the proposed approach outperforms\nstate-of-the-art methods w.r.t. the fairness-utility trade-off and w.r.t. both\ngroup and individual fairness metrics. This work presents a fresh perspective\non the trade-off between utility and algorithmic fairness and opens a key\navenue for the potential of using prediction uncertainty in fair machine\nlearning.\n", "date": "2023-04-07", "categories": []}
{"id": "2308.07805", "title": "Fairness and Privacy in Federated Learning and Their Implications in\n  Healthcare", "abstract": "  Currently, many contexts exist where distributed learning is difficult or\notherwise constrained by security and communication limitations. One common\ndomain where this is a consideration is in Healthcare where data is often\ngoverned by data-use-ordinances like HIPAA. On the other hand, larger sample\nsizes and shared data models are necessary to allow models to better generalize\non account of the potential for more variability and balancing underrepresented\nclasses. Federated learning is a type of distributed learning model that allows\ndata to be trained in a decentralized manner. This, in turn, addresses data\nsecurity, privacy, and vulnerability considerations as data itself is not\nshared across a given learning network nodes. Three main challenges to\nfederated learning include node data is not independent and identically\ndistributed (iid), clients requiring high levels of communication overhead\nbetween peers, and there is the heterogeneity of different clients within a\nnetwork with respect to dataset bias and size. As the field has grown, the\nnotion of fairness in federated learning has also been introduced through novel\nimplementations. Fairness approaches differ from the standard form of federated\nlearning and also have distinct challenges and considerations for the\nhealthcare domain. This paper endeavors to outline the typical lifecycle of\nfair federated learning in research as well as provide an updated taxonomy to\naccount for the current state of fairness in implementations. Lastly, this\npaper provides added insight into the implications and challenges of\nimplementing and supporting fairness in federated learning in the healthcare\ndomain.\n", "date": "2023-08-15", "categories": []}
{"id": "2207.05302", "title": "Causal Conceptions of Fairness and their Consequences", "abstract": "  Recent work highlights the role of causality in designing equitable\ndecision-making algorithms. It is not immediately clear, however, how existing\ncausal conceptions of fairness relate to one another, or what the consequences\nare of using these definitions as design principles. Here, we first assemble\nand categorize popular causal definitions of algorithmic fairness into two\nbroad families: (1) those that constrain the effects of decisions on\ncounterfactual disparities; and (2) those that constrain the effects of legally\nprotected characteristics, like race and gender, on decisions. We then show,\nanalytically and empirically, that both families of definitions \\emph{almost\nalways} -- in a measure theoretic sense -- result in strongly Pareto dominated\ndecision policies, meaning there is an alternative, unconstrained policy\nfavored by every stakeholder with preferences drawn from a large, natural\nclass. For example, in the case of college admissions decisions, policies\nconstrained to satisfy causal fairness definitions would be disfavored by every\nstakeholder with neutral or positive preferences for both academic preparedness\nand diversity. Indeed, under a prominent definition of causal fairness, we\nprove the resulting policies require admitting all students with the same\nprobability, regardless of academic qualifications or group membership. Our\nresults highlight formal limitations and potential adverse consequences of\ncommon mathematical notions of causal fairness.\n", "date": "2022-07-12", "categories": []}
{"id": "2305.06969", "title": "A Survey on Intersectional Fairness in Machine Learning: Notions,\n  Mitigation, and Challenges", "abstract": "  The widespread adoption of Machine Learning systems, especially in more\ndecision-critical applications such as criminal sentencing and bank loans, has\nled to increased concerns about fairness implications. Algorithms and metrics\nhave been developed to mitigate and measure these discriminations. More\nrecently, works have identified a more challenging form of bias called\nintersectional bias, which encompasses multiple sensitive attributes, such as\nrace and gender, together. In this survey, we review the state-of-the-art in\nintersectional fairness. We present a taxonomy for intersectional notions of\nfairness and mitigation. Finally, we identify the key challenges and provide\nresearchers with guidelines for future directions.\n", "date": "2023-05-11", "categories": []}
{"id": "2305.07041", "title": "Fairness in Machine Learning meets with Equity in Healthcare", "abstract": "  With the growing utilization of machine learning in healthcare, there is\nincreasing potential to enhance healthcare outcomes. However, this also brings\nthe risk of perpetuating biases in data and model design that can harm certain\ndemographic groups based on factors such as age, gender, and race. This study\nproposes an artificial intelligence framework, grounded in software engineering\nprinciples, for identifying and mitigating biases in data and models while\nensuring fairness in healthcare settings. A case study is presented to\ndemonstrate how systematic biases in data can lead to amplified biases in model\npredictions, and machine learning methods are suggested to prevent such biases.\nFuture research aims to test and validate the proposed ML framework in\nreal-world clinical settings to evaluate its impact on promoting health equity.\n", "date": "2023-05-11", "categories": []}
{"id": "2308.06434", "title": "Distributionally Robust Optimization and Invariant Representation\n  Learning for Addressing Subgroup Underrepresentation: Mechanisms and\n  Limitations", "abstract": "  Spurious correlation caused by subgroup underrepresentation has received\nincreasing attention as a source of bias that can be perpetuated by deep neural\nnetworks (DNNs). Distributionally robust optimization has shown success in\naddressing this bias, although the underlying working mechanism mostly relies\non upweighting under-performing samples as surrogates for those\nunderrepresented in data. At the same time, while invariant representation\nlearning has been a powerful choice for removing nuisance-sensitive features,\nit has been little considered in settings where spurious correlations are\ncaused by significant underrepresentation of subgroups. In this paper, we take\nthe first step to better understand and improve the mechanisms for debiasing\nspurious correlation due to subgroup underrepresentation in medical image\nclassification. Through a comprehensive evaluation study, we first show that 1)\ngeneralized reweighting of under-performing samples can be problematic when\nbias is not the only cause for poor performance, while 2) naive invariant\nrepresentation learning suffers from spurious correlations itself. We then\npresent a novel approach that leverages robust optimization to facilitate the\nlearning of invariant representations at the presence of spurious correlations.\nFinetuned classifiers utilizing such representation demonstrated improved\nabilities to reduce subgroup performance disparity, while maintaining high\naverage and worst-group performance.\n", "date": "2023-08-12", "categories": []}
{"id": "2307.11379", "title": "Towards Better Fairness-Utility Trade-off: A Comprehensive\n  Measurement-Based Reinforcement Learning Framework", "abstract": "  Machine learning is widely used to make decisions with societal impact such\nas bank loan approving, criminal sentencing, and resume filtering. How to\nensure its fairness while maintaining utility is a challenging but crucial\nissue. Fairness is a complex and context-dependent concept with over 70\ndifferent measurement metrics. Since existing regulations are often vague in\nterms of which metric to use and different organizations may prefer different\nfairness metrics, it is important to have means of improving fairness\ncomprehensively. Existing mitigation techniques often target at one specific\nfairness metric and have limitations in improving multiple notions of fairness\nsimultaneously. In this work, we propose CFU (Comprehensive Fairness-Utility),\na reinforcement learning-based framework, to efficiently improve the\nfairness-utility trade-off in machine learning classifiers. A comprehensive\nmeasurement that can simultaneously consider multiple fairness notions as well\nas utility is established, and new metrics are proposed based on an in-depth\nanalysis of the relationship between different fairness metrics. The reward\nfunction of CFU is constructed with comprehensive measurement and new metrics.\nWe conduct extensive experiments to evaluate CFU on 6 tasks, 3 machine learning\nmodels, and 15 fairness-utility measurements. The results demonstrate that CFU\ncan improve the classifier on multiple fairness metrics without sacrificing its\nutility. It outperforms all state-of-the-art techniques and has witnessed a\n37.5% improvement on average.\n", "date": "2023-07-21", "categories": []}
{"id": "2308.05129", "title": "Are Sex-based Physiological Differences the Cause of Gender Bias for\n  Chest X-ray Diagnosis?", "abstract": "  While many studies have assessed the fairness of AI algorithms in the medical\nfield, the causes of differences in prediction performance are often unknown.\nThis lack of knowledge about the causes of bias hampers the efficacy of bias\nmitigation, as evidenced by the fact that simple dataset balancing still often\nperforms best in reducing performance gaps but is unable to resolve all\nperformance differences. In this work, we investigate the causes of gender bias\nin machine learning-based chest X-ray diagnosis. In particular, we explore the\nhypothesis that breast tissue leads to underexposure of the lungs and causes\nlower model performance. Methodologically, we propose a new sampling method\nwhich addresses the highly skewed distribution of recordings per patient in two\nwidely used public datasets, while at the same time reducing the impact of\nlabel errors. Our comprehensive analysis of gender differences across diseases,\ndatasets, and gender representations in the training set shows that dataset\nimbalance is not the sole cause of performance differences. Moreover, relative\ngroup performance differs strongly between datasets, indicating important\ndataset-specific factors influencing male/female group performance. Finally, we\ninvestigate the effect of breast tissue more specifically, by cropping out the\nbreasts from recordings, finding that this does not resolve the observed\nperformance gaps. In conclusion, our results indicate that dataset-specific\nfactors, not fundamental physiological differences, are the main drivers of\nmale--female performance gaps in chest X-ray analyses on widely used NIH and\nCheXpert Dataset.\n", "date": "2023-08-09", "categories": []}
{"id": "2212.04593", "title": "Towards Understanding Fairness and its Composition in Ensemble Machine\n  Learning", "abstract": "  Machine Learning (ML) software has been widely adopted in modern society,\nwith reported fairness implications for minority groups based on race, sex,\nage, etc. Many recent works have proposed methods to measure and mitigate\nalgorithmic bias in ML models. The existing approaches focus on single\nclassifier-based ML models. However, real-world ML models are often composed of\nmultiple independent or dependent learners in an ensemble (e.g., Random\nForest), where the fairness composes in a non-trivial way. How does fairness\ncompose in ensembles? What are the fairness impacts of the learners on the\nultimate fairness of the ensemble? Can fair learners result in an unfair\nensemble? Furthermore, studies have shown that hyperparameters influence the\nfairness of ML models. Ensemble hyperparameters are more complex since they\naffect how learners are combined in different categories of ensembles.\nUnderstanding the impact of ensemble hyperparameters on fairness will help\nprogrammers design fair ensembles. Today, we do not understand these fully for\ndifferent ensemble algorithms. In this paper, we comprehensively study popular\nreal-world ensembles: bagging, boosting, stacking and voting. We have developed\na benchmark of 168 ensemble models collected from Kaggle on four popular\nfairness datasets. We use existing fairness metrics to understand the\ncomposition of fairness. Our results show that ensembles can be designed to be\nfairer without using mitigation techniques. We also identify the interplay\nbetween fairness composition and data characteristics to guide fair ensemble\ndesign. Finally, our benchmark can be leveraged for further research on fair\nensembles. To the best of our knowledge, this is one of the first and largest\nstudies on fairness composition in ensembles yet presented in the literature.\n", "date": "2022-12-08", "categories": []}
{"id": "2303.17566", "title": "Non-Invasive Fairness in Learning through the Lens of Data Drift", "abstract": "  Machine Learning (ML) models are widely employed to drive many modern data\nsystems. While they are undeniably powerful tools, ML models often demonstrate\nimbalanced performance and unfair behaviors. The root of this problem often\nlies in the fact that different subpopulations commonly display divergent\ntrends: as a learning algorithm tries to identify trends in the data, it\nnaturally favors the trends of the majority groups, leading to a model that\nperforms poorly and unfairly for minority populations. Our goal is to improve\nthe fairness and trustworthiness of ML models by applying only non-invasive\ninterventions, i.e., without altering the data or the learning algorithm. We\nuse a simple but key insight: the divergence of trends between different\npopulations, and, consecutively, between a learned model and minority\npopulations, is analogous to data drift, which indicates the poor conformance\nbetween parts of the data and the trained model. We explore two strategies\n(model-splitting and reweighing) to resolve this drift, aiming to improve the\noverall conformance of models to the underlying data. Both our methods\nintroduce novel ways to employ the recently-proposed data profiling primitive\nof Conformance Constraints. Our experimental evaluation over 7 real-world\ndatasets shows that both DifFair and ConFair improve the fairness of ML models.\nWe demonstrate scenarios where DifFair has an edge, though ConFair has the\ngreatest practical impact and outperforms other baselines. Moreover, as a\nmodel-agnostic technique, ConFair stays robust when used against different\nmodels than the ones on which the weights have been learned, which is not the\ncase for other state of the art.\n", "date": "2023-03-30", "categories": []}
{"id": "2207.10384", "title": "Detecting Shortcut Learning for Fair Medical AI using Shortcut Testing", "abstract": "  Machine learning (ML) holds great promise for improving healthcare, but it is\ncritical to ensure that its use will not propagate or amplify health\ndisparities. An important step is to characterize the (un)fairness of ML models\n- their tendency to perform differently across subgroups of the population -\nand to understand its underlying mechanisms. One potential driver of\nalgorithmic unfairness, shortcut learning, arises when ML models base\npredictions on improper correlations in the training data. However, diagnosing\nthis phenomenon is difficult, especially when sensitive attributes are causally\nlinked with disease. Using multi-task learning, we propose the first method to\nassess and mitigate shortcut learning as a part of the fairness assessment of\nclinical ML systems, and demonstrate its application to clinical tasks in\nradiology and dermatology. Finally, our approach reveals instances when\nshortcutting is not responsible for unfairness, highlighting the need for a\nholistic approach to fairness mitigation in medical AI.\n", "date": "2022-07-21", "categories": []}
{"id": "2307.08496", "title": "Can We Trust Race Prediction?", "abstract": "  In the absence of sensitive race and ethnicity data, researchers, regulators,\nand firms alike turn to proxies. In this paper, I train a Bidirectional Long\nShort-Term Memory (BiLSTM) model on a novel dataset of voter registration data\nfrom all 50 US states and create an ensemble that achieves up to 36.8% higher\nout of sample (OOS) F1 scores than the best performing machine learning models\nin the literature. Additionally, I construct the most comprehensive database of\nfirst and surname distributions in the US in order to improve the coverage and\naccuracy of Bayesian Improved Surname Geocoding (BISG) and Bayesian Improved\nFirstname Surname Geocoding (BIFSG). Finally, I provide the first high-quality\nbenchmark dataset in order to fairly compare existing models and aid future\nmodel developers.\n", "date": "2023-07-17", "categories": []}
{"id": "2308.04346", "title": "Unmasking Nationality Bias: A Study of Human Perception of Nationalities\n  in AI-Generated Articles", "abstract": "  We investigate the potential for nationality biases in natural language\nprocessing (NLP) models using human evaluation methods. Biased NLP models can\nperpetuate stereotypes and lead to algorithmic discrimination, posing a\nsignificant challenge to the fairness and justice of AI systems. Our study\nemploys a two-step mixed-methods approach that includes both quantitative and\nqualitative analysis to identify and understand the impact of nationality bias\nin a text generation model. Through our human-centered quantitative analysis,\nwe measure the extent of nationality bias in articles generated by AI sources.\nWe then conduct open-ended interviews with participants, performing qualitative\ncoding and thematic analysis to understand the implications of these biases on\nhuman readers. Our findings reveal that biased NLP models tend to replicate and\namplify existing societal biases, which can translate to harm if used in a\nsociotechnical setting. The qualitative analysis from our interviews offers\ninsights into the experience readers have when encountering such articles,\nhighlighting the potential to shift a reader's perception of a country. These\nfindings emphasize the critical role of public perception in shaping AI's\nimpact on society and the need to correct biases in AI systems.\n", "date": "2023-08-08", "categories": []}
{"id": "2206.10923", "title": "FairGrad: Fairness Aware Gradient Descent", "abstract": "  We address the problem of group fairness in classification, where the\nobjective is to learn models that do not unjustly discriminate against\nsubgroups of the population. Most existing approaches are limited to simple\nbinary tasks or involve difficult to implement training mechanisms which\nreduces their practical applicability. In this paper, we propose FairGrad, a\nmethod to enforce fairness based on a re-weighting scheme that iteratively\nlearns group specific weights based on whether they are advantaged or not.\nFairGrad is easy to implement, accommodates various standard fairness\ndefinitions, and comes with minimal overhead. Furthermore, we show that it is\ncompetitive with standard baselines over various datasets including ones used\nin natural language processing and computer vision.\n  FairGrad is available as a PyPI package at -\nhttps://pypi.org/project/fairgrad\n", "date": "2022-06-22", "categories": []}
{"id": "2302.00025", "title": "On the Within-Group Fairness of Screening Classifiers", "abstract": "  Screening classifiers are increasingly used to identify qualified candidates\nin a variety of selection processes. In this context, it has been recently\nshown that, if a classifier is calibrated, one can identify the smallest set of\ncandidates which contains, in expectation, a desired number of qualified\ncandidates using a threshold decision rule. This lends support to focusing on\ncalibration as the only requirement for screening classifiers. In this paper,\nwe argue that screening policies that use calibrated classifiers may suffer\nfrom an understudied type of within-group unfairness -- they may unfairly treat\nqualified members within demographic groups of interest. Further, we argue that\nthis type of unfairness can be avoided if classifiers satisfy within-group\nmonotonicity, a natural monotonicity property within each of the groups. Then,\nwe introduce an efficient post-processing algorithm based on dynamic\nprogramming to minimally modify a given calibrated classifier so that its\nprobability estimates satisfy within-group monotonicity. We validate our\nalgorithm using US Census survey data and show that within-group monotonicity\ncan be often achieved at a small cost in terms of prediction granularity and\nshortlist size.\n", "date": "2023-01-31", "categories": []}
{"id": "2308.02680", "title": "Fair Models in Credit: Intersectional Discrimination and the\n  Amplification of Inequity", "abstract": "  The increasing usage of new data sources and machine learning (ML) technology\nin credit modeling raises concerns with regards to potentially unfair\ndecision-making that rely on protected characteristics (e.g., race, sex, age)\nor other socio-economic and demographic data. The authors demonstrate the\nimpact of such algorithmic bias in the microfinance context. Difficulties in\nassessing credit are disproportionately experienced among vulnerable groups,\nhowever, very little is known about inequities in credit allocation between\ngroups defined, not only by single, but by multiple and intersecting social\ncategories. Drawing from the intersectionality paradigm, the study examines\nintersectional horizontal inequities in credit access by gender, age, marital\nstatus, single parent status and number of children. This paper utilizes data\nfrom the Spanish microfinance market as its context to demonstrate how\npluralistic realities and intersectional identities can shape patterns of\ncredit allocation when using automated decision-making systems. With ML\ntechnology being oblivious to societal good or bad, we find that a more\nthorough examination of intersectionality can enhance the algorithmic fairness\nlens to more authentically empower action for equitable outcomes and present a\nfairer path forward. We demonstrate that while on a high-level, fairness may\nexist superficially, unfairness can exacerbate at lower levels given\ncombinatorial effects; in other words, the core fairness problem may be more\ncomplicated than current literature demonstrates. We find that in addition to\nlegally protected characteristics, sensitive attributes such as single parent\nstatus and number of children can result in imbalanced harm. We discuss the\nimplications of these findings for the financial services industry.\n", "date": "2023-08-01", "categories": []}
{"id": "2308.02898", "title": "Elucidate Gender Fairness in Singing Voice Transcription", "abstract": "  It is widely known that males and females typically possess different sound\ncharacteristics when singing, such as timbre and pitch, but it has never been\nexplored whether these gender-based characteristics lead to a performance\ndisparity in singing voice transcription (SVT), whose target includes pitch.\nSuch a disparity could cause fairness issues and severely affect the user\nexperience of downstream SVT applications. Motivated by this, we first\ndemonstrate the female superiority of SVT systems, which is observed across\ndifferent models and datasets. We find that different pitch distributions,\nrather than gender data imbalance, contribute to this disparity. To address\nthis issue, we propose using an attribute predictor to predict gender labels\nand adversarially training the SVT system to enforce the gender-invariance of\nacoustic representations. Leveraging the prior knowledge that pitch\ndistributions may contribute to the gender bias, we propose conditionally\naligning acoustic representations between demographic groups by feeding note\nevents to the attribute predictor. Empirical experiments on multiple benchmark\nSVT datasets show that our method significantly reduces gender bias (up to more\nthan 50%) with negligible degradation of overall SVT performance, on both\nin-domain and out-of-domain singing data, thus offering a better\nfairness-utility trade-off.\n", "date": "2023-08-05", "categories": []}
{"id": "2308.01948", "title": "A Multidimensional Analysis of Social Biases in Vision Transformers", "abstract": "  The embedding spaces of image models have been shown to encode a range of\nsocial biases such as racism and sexism. Here, we investigate specific factors\nthat contribute to the emergence of these biases in Vision Transformers (ViT).\nTherefore, we measure the impact of training data, model architecture, and\ntraining objectives on social biases in the learned representations of ViTs.\nOur findings indicate that counterfactual augmentation training using\ndiffusion-based image editing can mitigate biases, but does not eliminate them.\nMoreover, we find that larger models are less biased than smaller models, and\nthat models trained using discriminative objectives are less biased than those\ntrained using generative objectives. In addition, we observe inconsistencies in\nthe learned social biases. To our surprise, ViTs can exhibit opposite biases\nwhen trained on the same data set using different self-supervised objectives.\nOur findings give insights into the factors that contribute to the emergence of\nsocial biases and suggests that we could achieve substantial fairness\nimprovements based on model design choices.\n", "date": "2023-08-03", "categories": []}
{"id": "2308.02081", "title": "Target specification bias, counterfactual prediction, and algorithmic\n  fairness in healthcare", "abstract": "  Bias in applications of machine learning (ML) to healthcare is usually\nattributed to unrepresentative or incomplete data, or to underlying health\ndisparities. This article identifies a more pervasive source of bias that\naffects the clinical utility of ML-enabled prediction tools: target\nspecification bias. Target specification bias arises when the\noperationalization of the target variable does not match its definition by\ndecision makers. The mismatch is often subtle, and stems from the fact that\ndecision makers are typically interested in predicting the outcomes of\ncounterfactual, rather than actual, healthcare scenarios. Target specification\nbias persists independently of data limitations and health disparities. When\nleft uncorrected, it gives rise to an overestimation of predictive accuracy, to\ninefficient utilization of medical resources, and to suboptimal decisions that\ncan harm patients. Recent work in metrology - the science of measurement -\nsuggests ways of counteracting target specification bias and avoiding its\nharmful consequences.\n", "date": "2023-08-03", "categories": []}
{"id": "2302.02096", "title": "Matrix Estimation for Individual Fairness", "abstract": "  In recent years, multiple notions of algorithmic fairness have arisen. One\nsuch notion is individual fairness (IF), which requires that individuals who\nare similar receive similar treatment. In parallel, matrix estimation (ME) has\nemerged as a natural paradigm for handling noisy data with missing values. In\nthis work, we connect the two concepts. We show that pre-processing data using\nME can improve an algorithm's IF without sacrificing performance. Specifically,\nwe show that using a popular ME method known as singular value thresholding\n(SVT) to pre-process the data provides a strong IF guarantee under appropriate\nconditions. We then show that, under analogous conditions, SVT pre-processing\nalso yields estimates that are consistent and approximately minimax optimal. As\nsuch, the ME pre-processing step does not, under the stated conditions,\nincrease the prediction error of the base algorithm, i.e., does not impose a\nfairness-performance trade-off. We verify these results on synthetic and real\ndata.\n", "date": "2023-02-04", "categories": []}
{"id": "2307.13408", "title": "The Double-Edged Sword of Big Data and Information Technology for the\n  Disadvantaged: A Cautionary Tale from Open Banking", "abstract": "  This research article analyses and demonstrates the hidden implications for\nfairness of seemingly neutral data coupled with powerful technology, such as\nmachine learning (ML), using Open Banking as an example. Open Banking has\nignited a revolution in financial services, opening new opportunities for\ncustomer acquisition, management, retention, and risk assessment. However, the\ngranularity of transaction data holds potential for harm where unnoticed\nproxies for sensitive and prohibited characteristics may lead to indirect\ndiscrimination. Against this backdrop, we investigate the dimensions of\nfinancial vulnerability (FV), a global concern resulting from COVID-19 and\nrising inflation. Specifically, we look to understand the behavioral elements\nleading up to FV and its impact on at-risk, disadvantaged groups through the\nlens of fair interpretation. Using a unique dataset from a UK FinTech lender,\nwe demonstrate the power of fine-grained transaction data while simultaneously\ncautioning its safe usage. Three ML classifiers are compared in predicting the\nlikelihood of FV, and groups exhibiting different magnitudes and forms of FV\nare identified via clustering to highlight the effects of feature combination.\nOur results indicate that engineered features of financial behavior can be\npredictive of omitted personal information, particularly sensitive or protected\ncharacteristics, shedding light on the hidden dangers of Open Banking data. We\ndiscuss the implications and conclude fairness via unawareness is ineffective\nin this new technological environment.\n", "date": "2023-07-25", "categories": []}
{"id": "2308.00133", "title": "A Suite of Fairness Datasets for Tabular Classification", "abstract": "  There have been many papers with algorithms for improving fairness of\nmachine-learning classifiers for tabular data. Unfortunately, most use only\nvery few datasets for their experimental evaluation. We introduce a suite of\nfunctions for fetching 20 fairness datasets and providing associated fairness\nmetadata. Hopefully, these will lead to more rigorous experimental evaluations\nin future fairness-aware machine learning research.\n", "date": "2023-07-31", "categories": []}
{"id": "2308.00341", "title": "Monitoring Algorithmic Fairness under Partial Observations", "abstract": "  As AI and machine-learned software are used increasingly for making decisions\nthat affect humans, it is imperative that they remain fair and unbiased in\ntheir decisions. To complement design-time bias mitigation measures, runtime\nverification techniques have been introduced recently to monitor the\nalgorithmic fairness of deployed systems. Previous monitoring techniques assume\nfull observability of the states of the (unknown) monitored system. Moreover,\nthey can monitor only fairness properties that are specified as arithmetic\nexpressions over the probabilities of different events. In this work, we extend\nfairness monitoring to systems modeled as partially observed Markov chains\n(POMC), and to specifications containing arithmetic expressions over the\nexpected values of numerical functions on event sequences. The only assumptions\nwe make are that the underlying POMC is aperiodic and starts in the stationary\ndistribution, with a bound on its mixing time being known. These assumptions\nenable us to estimate a given property for the entire distribution of possible\nexecutions of the monitored POMC, by observing only a single execution. Our\nmonitors observe a long run of the system and, after each new observation,\noutput updated PAC-estimates of how fair or biased the system is. The monitors\nare computationally lightweight and, using a prototype implementation, we\ndemonstrate their effectiveness on several real-world examples.\n", "date": "2023-08-01", "categories": []}
{"id": "2307.15838", "title": "Holistic Survey of Privacy and Fairness in Machine Learning", "abstract": "  Privacy and fairness are two crucial pillars of responsible Artificial\nIntelligence (AI) and trustworthy Machine Learning (ML). Each objective has\nbeen independently studied in the literature with the aim of reducing utility\nloss in achieving them. Despite the significant interest attracted from both\nacademia and industry, there remains an immediate demand for more in-depth\nresearch to unravel how these two objectives can be simultaneously integrated\ninto ML models. As opposed to well-accepted trade-offs, i.e., privacy-utility\nand fairness-utility, the interrelation between privacy and fairness is not\nwell-understood. While some works suggest a trade-off between the two objective\nfunctions, there are others that demonstrate the alignment of these functions\nin certain scenarios. To fill this research gap, we provide a thorough review\nof privacy and fairness in ML, including supervised, unsupervised,\nsemi-supervised, and reinforcement learning. After examining and consolidating\nthe literature on both objectives, we present a holistic survey on the impact\nof privacy on fairness, the impact of fairness on privacy, existing\narchitectures, their interaction in application domains, and algorithms that\naim to achieve both objectives while minimizing the utility sacrificed.\nFinally, we identify research challenges in achieving privacy and fairness\nconcurrently in ML, particularly focusing on large language models.\n", "date": "2023-07-28", "categories": []}
{"id": "2307.16526", "title": "No Fair Lunch: A Causal Perspective on Dataset Bias in Machine Learning\n  for Medical Imaging", "abstract": "  As machine learning methods gain prominence within clinical decision-making,\naddressing fairness concerns becomes increasingly urgent. Despite considerable\nwork dedicated to detecting and ameliorating algorithmic bias, today's methods\nare deficient with potentially harmful consequences. Our causal perspective\nsheds new light on algorithmic bias, highlighting how different sources of\ndataset bias may appear indistinguishable yet require substantially different\nmitigation strategies. We introduce three families of causal bias mechanisms\nstemming from disparities in prevalence, presentation, and annotation. Our\ncausal analysis underscores how current mitigation methods tackle only a narrow\nand often unrealistic subset of scenarios. We provide a practical three-step\nframework for reasoning about fairness in medical imaging, supporting the\ndevelopment of safe and equitable AI prediction models.\n", "date": "2023-07-31", "categories": []}
{"id": "2212.10839", "title": "Consistent Range Approximation for Fair Predictive Modeling", "abstract": "  This paper proposes a novel framework for certifying the fairness of\npredictive models trained on biased data. It draws from query answering for\nincomplete and inconsistent databases to formulate the problem of consistent\nrange approximation (CRA) of fairness queries for a predictive model on a\ntarget population. The framework employs background knowledge of the data\ncollection process and biased data, working with or without limited statistics\nabout the target population, to compute a range of answers for fairness\nqueries. Using CRA, the framework builds predictive models that are certifiably\nfair on the target population, regardless of the availability of external data\nduring training. The framework's efficacy is demonstrated through evaluations\non real data, showing substantial improvement over existing state-of-the-art\nmethods.\n", "date": "2022-12-21", "categories": []}
{"id": "2307.15247", "title": "Is this model reliable for everyone? Testing for strong calibration", "abstract": "  In a well-calibrated risk prediction model, the average predicted probability\nis close to the true event rate for any given subgroup. Such models are\nreliable across heterogeneous populations and satisfy strong notions of\nalgorithmic fairness. However, the task of auditing a model for strong\ncalibration is well-known to be difficult -- particularly for machine learning\n(ML) algorithms -- due to the sheer number of potential subgroups. As such,\ncommon practice is to only assess calibration with respect to a few predefined\nsubgroups. Recent developments in goodness-of-fit testing offer potential\nsolutions but are not designed for settings with weak signal or where the\npoorly calibrated subgroup is small, as they either overly subdivide the data\nor fail to divide the data at all. We introduce a new testing procedure based\non the following insight: if we can reorder observations by their expected\nresiduals, there should be a change in the association between the predicted\nand observed residuals along this sequence if a poorly calibrated subgroup\nexists. This lets us reframe the problem of calibration testing into one of\nchangepoint detection, for which powerful methods already exist. We begin with\nintroducing a sample-splitting procedure where a portion of the data is used to\ntrain a suite of candidate models for predicting the residual, and the\nremaining data are used to perform a score-based cumulative sum (CUSUM) test.\nTo further improve power, we then extend this adaptive CUSUM test to\nincorporate cross-validation, while maintaining Type I error control under\nminimal assumptions. Compared to existing methods, the proposed procedure\nconsistently achieved higher power in simulation studies and more than doubled\nthe power when auditing a mortality risk prediction model.\n", "date": "2023-07-28", "categories": []}
{"id": "2307.15466", "title": "LUCID-GAN: Conditional Generative Models to Locate Unfairness", "abstract": "  Most group fairness notions detect unethical biases by computing statistical\nparity metrics on a model's output. However, this approach suffers from several\nshortcomings, such as philosophical disagreement, mutual incompatibility, and\nlack of interpretability. These shortcomings have spurred the research on\ncomplementary bias detection methods that offer additional transparency into\nthe sources of discrimination and are agnostic towards an a priori decision on\nthe definition of fairness and choice of protected features. A recent proposal\nin this direction is LUCID (Locating Unfairness through Canonical Inverse\nDesign), where canonical sets are generated by performing gradient descent on\nthe input space, revealing a model's desired input given a preferred output.\nThis information about the model's mechanisms, i.e., which feature values are\nessential to obtain specific outputs, allows exposing potential unethical\nbiases in its internal logic. Here, we present LUCID-GAN, which generates\ncanonical inputs via a conditional generative model instead of gradient-based\ninverse design. LUCID-GAN has several benefits, including that it applies to\nnon-differentiable models, ensures that canonical sets consist of realistic\ninputs, and allows to assess proxy and intersectional discrimination. We\nempirically evaluate LUCID-GAN on the UCI Adult and COMPAS data sets and show\nthat it allows for detecting unethical biases in black-box models without\nrequiring access to the training data.\n", "date": "2023-07-28", "categories": []}
{"id": "2307.14668", "title": "Bipartite Ranking Fairness through a Model Agnostic Ordering Adjustment", "abstract": "  Algorithmic fairness has been a serious concern and received lots of interest\nin machine learning community. In this paper, we focus on the bipartite ranking\nscenario, where the instances come from either the positive or negative class\nand the goal is to learn a ranking function that ranks positive instances\nhigher than negative ones. While there could be a trade-off between fairness\nand performance, we propose a model agnostic post-processing framework xOrder\nfor achieving fairness in bipartite ranking and maintaining the algorithm\nclassification performance. In particular, we optimize a weighted sum of the\nutility as identifying an optimal warping path across different protected\ngroups and solve it through a dynamic programming process. xOrder is compatible\nwith various classification models and ranking fairness metrics, including\nsupervised and unsupervised fairness metrics. In addition to binary groups,\nxOrder can be applied to multiple protected groups. We evaluate our proposed\nalgorithm on four benchmark data sets and two real-world patient electronic\nhealth record repositories. xOrder consistently achieves a better balance\nbetween the algorithm utility and ranking fairness on a variety of datasets\nwith different metrics. From the visualization of the calibrated ranking\nscores, xOrder mitigates the score distribution shifts of different groups\ncompared with baselines. Moreover, additional analytical results verify that\nxOrder achieves a robust performance when faced with fewer samples and a bigger\ndifference between training and testing ranking score distributions.\n", "date": "2023-07-27", "categories": []}
{"id": "2303.15592", "title": "Uncovering Bias in Personal Informatics", "abstract": "  Personal informatics (PI) systems, powered by smartphones and wearables,\nenable people to lead healthier lifestyles by providing meaningful and\nactionable insights that break down barriers between users and their health\ninformation. Today, such systems are used by billions of users for monitoring\nnot only physical activity and sleep but also vital signs and women's and heart\nhealth, among others. Despite their widespread usage, the processing of\nsensitive PI data may suffer from biases, which may entail practical and\nethical implications. In this work, we present the first comprehensive\nempirical and analytical study of bias in PI systems, including biases in raw\ndata and in the entire machine learning life cycle. We use the most detailed\nframework to date for exploring the different sources of bias and find that\nbiases exist both in the data generation and the model learning and\nimplementation streams. According to our results, the most affected minority\ngroups are users with health issues, such as diabetes, joint issues, and\nhypertension, and female users, whose data biases are propagated or even\namplified by learning models, while intersectional biases can also be observed.\n", "date": "2023-03-27", "categories": []}
{"id": "2307.13616", "title": "AI and ethics in insurance: a new solution to mitigate proxy\n  discrimination in risk modeling", "abstract": "  The development of Machine Learning is experiencing growing interest from the\ngeneral public, and in recent years there have been numerous press articles\nquestioning its objectivity: racism, sexism, \\dots Driven by the growing\nattention of regulators on the ethical use of data in insurance, the actuarial\ncommunity must rethink pricing and risk selection practices for fairer\ninsurance. Equity is a philosophy concept that has many different definitions\nin every jurisdiction that influence each other without currently reaching\nconsensus. In Europe, the Charter of Fundamental Rights defines guidelines on\ndiscrimination, and the use of sensitive personal data in algorithms is\nregulated. If the simple removal of the protected variables prevents any\nso-called `direct' discrimination, models are still able to `indirectly'\ndiscriminate between individuals thanks to latent interactions between\nvariables, which bring better performance (and therefore a better\nquantification of risk, segmentation of prices, and so on). After introducing\nthe key concepts related to discrimination, we illustrate the complexity of\nquantifying them. We then propose an innovative method, not yet met in the\nliterature, to reduce the risks of indirect discrimination thanks to\nmathematical concepts of linear algebra. This technique is illustrated in a\nconcrete case of risk selection in life insurance, demonstrating its simplicity\nof use and its promising performance.\n", "date": "2023-07-25", "categories": []}
{"id": "2206.02058", "title": "When Personalization Harms: Reconsidering the Use of Group Attributes in\n  Prediction", "abstract": "  Machine learning models are often personalized with categorical attributes\nthat are protected, sensitive, self-reported, or costly to acquire. In this\nwork, we show models that are personalized with group attributes can reduce\nperformance at a group level. We propose formal conditions to ensure the \"fair\nuse\" of group attributes in prediction tasks by training one additional model\n-- i.e., collective preference guarantees to ensure that each group who\nprovides personal data will receive a tailored gain in performance in return.\nWe present sufficient conditions to ensure fair use in empirical risk\nminimization and characterize failure modes that lead to fair use violations\ndue to standard practices in model development and deployment. We present a\ncomprehensive empirical study of fair use in clinical prediction tasks. Our\nresults demonstrate the prevalence of fair use violations in practice and\nillustrate simple interventions to mitigate their harm.\n", "date": "2022-06-04", "categories": []}
{"id": "2303.17555", "title": "Factoring the Matrix of Domination: A Critical Review and Reimagination\n  of Intersectionality in AI Fairness", "abstract": "  Intersectionality is a critical framework that, through inquiry and praxis,\nallows us to examine how social inequalities persist through domains of\nstructure and discipline. Given AI fairness' raison d'etre of \"fairness\", we\nargue that adopting intersectionality as an analytical framework is pivotal to\neffectively operationalizing fairness. Through a critical review of how\nintersectionality is discussed in 30 papers from the AI fairness literature, we\ndeductively and inductively: 1) map how intersectionality tenets operate within\nthe AI fairness paradigm and 2) uncover gaps between the conceptualization and\noperationalization of intersectionality. We find that researchers\noverwhelmingly reduce intersectionality to optimizing for fairness metrics over\ndemographic subgroups. They also fail to discuss their social context and when\nmentioning power, they mostly situate it only within the AI pipeline. We: 3)\noutline and assess the implications of these gaps for critical inquiry and\npraxis, and 4) provide actionable recommendations for AI fairness researchers\nto engage with intersectionality in their work by grounding it in AI\nepistemology.\n", "date": "2023-03-16", "categories": []}
{"id": "2305.15342", "title": "Is Your Model \"MADD\"? A Novel Metric to Evaluate Algorithmic Fairness\n  for Predictive Student Models", "abstract": "  Predictive student models are increasingly used in learning environments due\nto their ability to enhance educational outcomes and support stakeholders in\nmaking informed decisions. However, predictive models can be biased and produce\nunfair outcomes, leading to potential discrimination against some students and\npossible harmful long-term implications. This has prompted research on fairness\nmetrics meant to capture and quantify such biases. Nonetheless, so far,\nexisting fairness metrics used in education are predictive\nperformance-oriented, focusing on assessing biased outcomes across groups of\nstudents, without considering the behaviors of the models nor the severity of\nthe biases in the outcomes. Therefore, we propose a novel metric, the Model\nAbsolute Density Distance (MADD), to analyze models' discriminatory behaviors\nindependently from their predictive performance. We also provide a\ncomplementary visualization-based analysis to enable fine-grained human\nassessment of how the models discriminate between groups of students. We\nevaluate our approach on the common task of predicting student success in\nonline courses, using several common predictive classification models on an\nopen educational dataset. We also compare our metric to the only predictive\nperformance-oriented fairness metric developed in education, ABROCA. Results on\nthis dataset show that: (1) fair predictive performance does not guarantee fair\nmodels' behaviors and thus fair outcomes, (2) there is no direct relationship\nbetween data bias and predictive performance bias nor discriminatory behaviors\nbias, and (3) trained on the same data, models exhibit different discriminatory\nbehaviors, according to different sensitive features too. We thus recommend\nusing the MADD on models that show satisfying predictive performance, to gain a\nfiner-grained understanding on how they behave and to refine models selection\nand their usage.\n", "date": "2023-05-24", "categories": []}
{"id": "2304.09826", "title": "Fairness in AI and Its Long-Term Implications on Society", "abstract": "  Successful deployment of artificial intelligence (AI) in various settings has\nled to numerous positive outcomes for individuals and society. However, AI\nsystems have also been shown to harm parts of the population due to biased\npredictions. AI fairness focuses on mitigating such biases to ensure AI\ndecision making is not discriminatory towards certain groups. We take a closer\nlook at AI fairness and analyze how lack of AI fairness can lead to deepening\nof biases over time and act as a social stressor. More specifically, we discuss\nhow biased models can lead to more negative real-world outcomes for certain\ngroups, which may then become more prevalent by deploying new AI models trained\non increasingly biased data, resulting in a feedback loop. If the issues\npersist, they could be reinforced by interactions with other risks and have\nsevere implications on society in the form of social unrest. We examine current\nstrategies for improving AI fairness, assess their limitations in terms of\nreal-world deployment, and explore potential paths forward to ensure we reap\nAI's benefits without causing society's collapse.\n", "date": "2023-04-16", "categories": []}
{"id": "2307.01738", "title": "Mitigating Calibration Bias Without Fixed Attribute Grouping for\n  Improved Fairness in Medical Imaging Analysis", "abstract": "  Trustworthy deployment of deep learning medical imaging models into\nreal-world clinical practice requires that they be calibrated. However, models\nthat are well calibrated overall can still be poorly calibrated for a\nsub-population, potentially resulting in a clinician unwittingly making poor\ndecisions for this group based on the recommendations of the model. Although\nmethods have been shown to successfully mitigate biases across subgroups in\nterms of model accuracy, this work focuses on the open problem of mitigating\ncalibration biases in the context of medical image analysis. Our method does\nnot require subgroup attributes during training, permitting the flexibility to\nmitigate biases for different choices of sensitive attributes without\nre-training. To this end, we propose a novel two-stage method: Cluster-Focal to\nfirst identify poorly calibrated samples, cluster them into groups, and then\nintroduce group-wise focal loss to improve calibration bias. We evaluate our\nmethod on skin lesion classification with the public HAM10000 dataset, and on\npredicting future lesional activity for multiple sclerosis (MS) patients. In\naddition to considering traditional sensitive attributes (e.g. age, sex) with\ndemographic subgroups, we also consider biases among groups with different\nimage-derived attributes, such as lesion load, which are required in medical\nimage analysis. Our results demonstrate that our method effectively controls\ncalibration error in the worst-performing subgroups while preserving prediction\nperformance, and outperforming recent baselines.\n", "date": "2023-07-04", "categories": []}
{"id": "2307.08722", "title": "Certifying the Fairness of KNN in the Presence of Dataset Bias", "abstract": "  We propose a method for certifying the fairness of the classification result\nof a widely used supervised learning algorithm, the k-nearest neighbors (KNN),\nunder the assumption that the training data may have historical bias caused by\nsystematic mislabeling of samples from a protected minority group. To the best\nof our knowledge, this is the first certification method for KNN based on three\nvariants of the fairness definition: individual fairness, $\\epsilon$-fairness,\nand label-flipping fairness. We first define the fairness certification problem\nfor KNN and then propose sound approximations of the complex arithmetic\ncomputations used in the state-of-the-art KNN algorithm. This is meant to lift\nthe computation results from the concrete domain to an abstract domain, to\nreduce the computational cost. We show effectiveness of this abstract\ninterpretation based technique through experimental evaluation on six datasets\nwidely used in the fairness research literature. We also show that the method\nis accurate enough to obtain fairness certifications for a large number of test\ninputs, despite the presence of historical bias in the datasets.\n", "date": "2023-07-17", "categories": []}
{"id": "2307.09209", "title": "Automated Ableism: An Exploration of Explicit Disability Biases in\n  Sentiment and Toxicity Analysis Models", "abstract": "  We analyze sentiment analysis and toxicity detection models to detect the\npresence of explicit bias against people with disability (PWD). We employ the\nbias identification framework of Perturbation Sensitivity Analysis to examine\nconversations related to PWD on social media platforms, specifically Twitter\nand Reddit, in order to gain insight into how disability bias is disseminated\nin real-world social settings. We then create the \\textit{Bias Identification\nTest in Sentiment} (BITS) corpus to quantify explicit disability bias in any\nsentiment analysis and toxicity detection models. Our study utilizes BITS to\nuncover significant biases in four open AIaaS (AI as a Service) sentiment\nanalysis tools, namely TextBlob, VADER, Google Cloud Natural Language API,\nDistilBERT and two toxicity detection models, namely two versions of\nToxic-BERT. Our findings indicate that all of these models exhibit\nstatistically significant explicit bias against PWD.\n", "date": "2023-07-18", "categories": []}
{"id": "2302.10893", "title": "Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness", "abstract": "  Generative AI models have recently achieved astonishing results in quality\nand are consequently employed in a fast-growing number of applications.\nHowever, since they are highly data-driven, relying on billion-sized datasets\nrandomly scraped from the internet, they also suffer from degenerated and\nbiased human behavior, as we demonstrate. In fact, they may even reinforce such\nbiases. To not only uncover but also combat these undesired effects, we present\na novel strategy, called Fair Diffusion, to attenuate biases after the\ndeployment of generative text-to-image models. Specifically, we demonstrate\nshifting a bias, based on human instructions, in any direction yielding\narbitrarily new proportions for, e.g., identity groups. As our empirical\nevaluation demonstrates, this introduced control enables instructing generative\nimage models on fairness, with no data filtering and additional training\nrequired.\n", "date": "2023-02-07", "categories": []}
{"id": "2304.03935", "title": "Last-Layer Fairness Fine-tuning is Simple and Effective for Neural\n  Networks", "abstract": "  As machine learning has been deployed ubiquitously across applications in\nmodern data science, algorithmic fairness has become a great concern. Among\nthem, imposing fairness constraints during learning, i.e. in-processing fair\ntraining, has been a popular type of training method because they don't require\naccessing sensitive attributes during test time in contrast to post-processing\nmethods. While this has been extensively studied in classical machine learning\nmodels, their impact on deep neural networks remains unclear. Recent research\nhas shown that adding fairness constraints to the objective function leads to\nsevere over-fitting to fairness criteria in large models, and how to solve this\nchallenge is an important open question. To tackle this, we leverage the wisdom\nand power of pre-training and fine-tuning and develop a simple but novel\nframework to train fair neural networks in an efficient and inexpensive way --\nlast-layer fine-tuning alone can effectively promote fairness in deep neural\nnetworks. This framework offers valuable insights into representation learning\nfor training fair neural networks.\n", "date": "2023-04-08", "categories": []}
{"id": "2306.10123", "title": "Dual Node and Edge Fairness-Aware Graph Partition", "abstract": "  Fair graph partition of social networks is a crucial step toward ensuring\nfair and non-discriminatory treatments in unsupervised user analysis. Current\nfair partition methods typically consider node balance, a notion pursuing a\nproportionally balanced number of nodes from all demographic groups, but ignore\nthe bias induced by imbalanced edges in each cluster. To address this gap, we\npropose a notion edge balance to measure the proportion of edges connecting\ndifferent demographic groups in clusters. We analyze the relations between node\nbalance and edge balance, then with line graph transformations, we propose a\nco-embedding framework to learn dual node and edge fairness-aware\nrepresentations for graph partition. We validate our framework through several\nsocial network datasets and observe balanced partition in terms of both nodes\nand edges along with good utility. Moreover, we demonstrate our fair partition\ncan be used as pseudo labels to facilitate graph neural networks to behave\nfairly in node classification and link prediction tasks.\n", "date": "2023-06-16", "categories": []}
{"id": "2307.08025", "title": "Analysing Gender Bias in Text-to-Image Models using Object Detection", "abstract": "  This work presents a novel strategy to measure bias in text-to-image models.\nUsing paired prompts that specify gender and vaguely reference an object (e.g.\n\"a man/woman holding an item\") we can examine whether certain objects are\nassociated with a certain gender. In analysing results from Stable Diffusion,\nwe observed that male prompts generated objects such as ties, knives, trucks,\nbaseball bats, and bicycles more frequently. On the other hand, female prompts\nwere more likely to generate objects such as handbags, umbrellas, bowls,\nbottles, and cups. We hope that the method outlined here will be a useful tool\nfor examining bias in text-to-image models.\n", "date": "2023-07-16", "categories": []}
{"id": "2307.08232", "title": "Learning for Counterfactual Fairness from Observational Data", "abstract": "  Fairness-aware machine learning has attracted a surge of attention in many\ndomains, such as online advertising, personalized recommendation, and social\nmedia analysis in web applications. Fairness-aware machine learning aims to\neliminate biases of learning models against certain subgroups described by\ncertain protected (sensitive) attributes such as race, gender, and age. Among\nmany existing fairness notions, counterfactual fairness is a popular notion\ndefined from a causal perspective. It measures the fairness of a predictor by\ncomparing the prediction of each individual in the original world and that in\nthe counterfactual worlds in which the value of the sensitive attribute is\nmodified. A prerequisite for existing methods to achieve counterfactual\nfairness is the prior human knowledge of the causal model for the data.\nHowever, in real-world scenarios, the underlying causal model is often unknown,\nand acquiring such human knowledge could be very difficult. In these scenarios,\nit is risky to directly trust the causal models obtained from information\nsources with unknown reliability and even causal discovery methods, as\nincorrect causal models can consequently bring biases to the predictor and lead\nto unfair predictions. In this work, we address the problem of counterfactually\nfair prediction from observational data without given causal models by\nproposing a novel framework CLAIRE. Specifically, under certain general\nassumptions, CLAIRE effectively mitigates the biases from the sensitive\nattribute with a representation learning framework based on counterfactual data\naugmentation and an invariant penalty. Experiments conducted on both synthetic\nand real-world datasets validate the superiority of CLAIRE in both\ncounterfactual fairness and prediction performance.\n", "date": "2023-07-17", "categories": []}
{"id": "2307.08484", "title": "Navigating Fairness Measures and Trade-Offs", "abstract": "  In order to monitor and prevent bias in AI systems we can use a wide range of\n(statistical) fairness measures. However, it is mathematically impossible to\noptimize for all of these measures at the same time. In addition, optimizing a\nfairness measure often greatly reduces the accuracy of the system (Kozodoi et\nal, 2022). As a result, we need a substantive theory that informs us how to\nmake these decisions and for what reasons. I show that by using Rawls' notion\nof justice as fairness, we can create a basis for navigating fairness measures\nand the accuracy trade-off. In particular, this leads to a principled choice\nfocusing on both the most vulnerable groups and the type of fairness measure\nthat has the biggest impact on that group. This also helps to close part of the\ngap between philosophical accounts of distributive justice and the fairness\nliterature that has been observed (Kuppler et al, 2021) and to operationalise\nthe value of fairness.\n", "date": "2023-07-17", "categories": []}
{"id": "2307.08624", "title": "National Origin Discrimination in Deep-learning-powered Automated Resume\n  Screening", "abstract": "  Many companies and organizations have started to use some form of AIenabled\nauto mated tools to assist in their hiring process, e.g. screening resumes,\ninterviewing candi dates, performance evaluation. While those AI tools have\ngreatly improved human re source operations efficiency and provided\nconveniences to job seekers as well, there are increasing concerns on unfair\ntreatment to candidates, caused by underlying bias in AI systems. Laws around\nequal opportunity and fairness, like GDPR, CCPA, are introduced or under\ndevelopment, in attempt to regulate AI. However, it is difficult to implement\nAI regulations in practice, as technologies are constantly advancing and the\nrisk perti nent to their applications can fail to be recognized. This study\nexamined deep learning methods, a recent technology breakthrough, with focus on\ntheir application to automated resume screening. One impressive performance of\ndeep learning methods is the represen tation of individual words as\nlowdimensional numerical vectors, called word embedding, which are learned from\naggregated global wordword cooccurrence statistics from a cor pus, like\nWikipedia or Google news. The resulting word representations possess interest\ning linear substructures of the word vector space and have been widely used in\ndown stream tasks, like resume screening. However, word embedding inherits and\nreinforces the stereotyping from the training corpus, as deep learning models\nessentially learn a probability distribution of words and their relations from\nhistory data. Our study finds out that if we rely on such deeplearningpowered\nautomated resume screening tools, it may lead to decisions favoring or\ndisfavoring certain demographic groups and raise eth ical, even legal,\nconcerns. To address the issue, we developed bias mitigation method. Extensive\nexperiments on real candidate resumes are conducted to validate our study\n", "date": "2023-07-13", "categories": []}
{"id": "2307.05728", "title": "Towards A Scalable Solution for Improving Multi-Group Fairness in\n  Compositional Classification", "abstract": "  Despite the rich literature on machine learning fairness, relatively little\nattention has been paid to remediating complex systems, where the final\nprediction is the combination of multiple classifiers and where multiple groups\nare present. In this paper, we first show that natural baseline approaches for\nimproving equal opportunity fairness scale linearly with the product of the\nnumber of remediated groups and the number of remediated prediction labels,\nrendering them impractical. We then introduce two simple techniques, called\n{\\em task-overconditioning} and {\\em group-interleaving}, to achieve a constant\nscaling in this multi-group multi-label setup. Our experimental results in\nacademic and real-world environments demonstrate the effectiveness of our\nproposal at mitigation within this environment.\n", "date": "2023-07-11", "categories": []}
{"id": "2307.03306", "title": "When Fair Classification Meets Noisy Protected Attributes", "abstract": "  The operationalization of algorithmic fairness comes with several practical\nchallenges, not the least of which is the availability or reliability of\nprotected attributes in datasets. In real-world contexts, practical and legal\nimpediments may prevent the collection and use of demographic data, making it\ndifficult to ensure algorithmic fairness. While initial fairness algorithms did\nnot consider these limitations, recent proposals aim to achieve algorithmic\nfairness in classification by incorporating noisiness in protected attributes\nor not using protected attributes at all.\n  To the best of our knowledge, this is the first head-to-head study of fair\nclassification algorithms to compare attribute-reliant, noise-tolerant and\nattribute-blind algorithms along the dual axes of predictivity and fairness. We\nevaluated these algorithms via case studies on four real-world datasets and\nsynthetic perturbations. Our study reveals that attribute-blind and\nnoise-tolerant fair classifiers can potentially achieve similar level of\nperformance as attribute-reliant algorithms, even when protected attributes are\nnoisy. However, implementing them in practice requires careful nuance. Our\nstudy provides insights into the practical implications of using fair\nclassification algorithms in scenarios where protected attributes are noisy or\npartially available.\n", "date": "2023-07-06", "categories": []}
{"id": "2307.05029", "title": "FairLay-ML: Intuitive Remedies for Unfairness in Data-Driven\n  Social-Critical Algorithms", "abstract": "  This thesis explores open-sourced machine learning (ML) model explanation\ntools to understand whether these tools can allow a layman to visualize,\nunderstand, and suggest intuitive remedies to unfairness in ML-based\ndecision-support systems. Machine learning models trained on datasets biased\nagainst minority groups are increasingly used to guide life-altering social\ndecisions, prompting the urgent need to study their logic for unfairness. Due\nto this problem's impact on vast populations of the general public, it is\ncritical for the layperson -- not just subject matter experts in social justice\nor machine learning experts -- to understand the nature of unfairness within\nthese algorithms and the potential trade-offs. Existing research on fairness in\nmachine learning focuses mostly on the mathematical definitions and tools to\nunderstand and remedy unfair models, with some directly citing user-interactive\ntools as necessary for future work. This thesis presents FairLay-ML, a\nproof-of-concept GUI integrating some of the most promising tools to provide\nintuitive explanations for unfair logic in ML models by integrating existing\nresearch tools (e.g. Local Interpretable Model-Agnostic Explanations) with\nexisting ML-focused GUI (e.g. Python Streamlit). We test FairLay-ML using\nmodels of various accuracy and fairness generated by an unfairness detector\ntool, Parfait-ML, and validate our results using Themis. Our study finds that\nthe technology stack used for FairLay-ML makes it easy to install and provides\nreal-time black-box explanations of pre-trained models to users. Furthermore,\nthe explanations provided translate to actionable remedies.\n", "date": "2023-07-11", "categories": []}
{"id": "2207.07723", "title": "More Data Can Lead Us Astray: Active Data Acquisition in the Presence of\n  Label Bias", "abstract": "  An increased awareness concerning risks of algorithmic bias has driven a\nsurge of efforts around bias mitigation strategies. A vast majority of the\nproposed approaches fall under one of two categories: (1) imposing algorithmic\nfairness constraints on predictive models, and (2) collecting additional\ntraining samples. Most recently and at the intersection of these two\ncategories, methods that propose active learning under fairness constraints\nhave been developed. However, proposed bias mitigation strategies typically\noverlook the bias presented in the observed labels. In this work, we study\nfairness considerations of active data collection strategies in the presence of\nlabel bias. We first present an overview of different types of label bias in\nthe context of supervised learning systems. We then empirically show that, when\noverlooking label bias, collecting more data can aggravate bias, and imposing\nfairness constraints that rely on the observed labels in the data collection\nprocess may not address the problem. Our results illustrate the unintended\nconsequences of deploying a model that attempts to mitigate a single type of\nbias while neglecting others, emphasizing the importance of explicitly\ndifferentiating between the types of bias that fairness-aware algorithms aim to\naddress, and highlighting the risks of neglecting label bias during data\ncollection.\n", "date": "2022-07-15", "categories": []}
{"id": "2307.03929", "title": "Fairness-Aware Graph Neural Networks: A Survey", "abstract": "  Graph Neural Networks (GNNs) have become increasingly important due to their\nrepresentational power and state-of-the-art predictive performance on many\nfundamental learning tasks. Despite this success, GNNs suffer from fairness\nissues that arise as a result of the underlying graph data and the fundamental\naggregation mechanism that lies at the heart of the large class of GNN models.\nIn this article, we examine and categorize fairness techniques for improving\nthe fairness of GNNs. Previous work on fair GNN models and techniques are\ndiscussed in terms of whether they focus on improving fairness during a\npreprocessing step, during training, or in a post-processing phase.\nFurthermore, we discuss how such techniques can be used together whenever\nappropriate, and highlight the advantages and intuition as well. We also\nintroduce an intuitive taxonomy for fairness evaluation metrics including\ngraph-level fairness, neighborhood-level fairness, embedding-level fairness,\nand prediction-level fairness metrics. In addition, graph datasets that are\nuseful for benchmarking the fairness of GNN models are summarized succinctly.\nFinally, we highlight key open problems and challenges that remain to be\naddressed.\n", "date": "2023-07-08", "categories": []}
{"id": "2307.04105", "title": "Towards Assumption-free Bias Mitigation", "abstract": "  Despite the impressive prediction ability, machine learning models show\ndiscrimination towards certain demographics and suffer from unfair prediction\nbehaviors. To alleviate the discrimination, extensive studies focus on\neliminating the unequal distribution of sensitive attributes via multiple\napproaches. However, due to privacy concerns, sensitive attributes are often\neither unavailable or missing in real-world scenarios. Therefore, several\nexisting works alleviate the bias without sensitive attributes. Those studies\nface challenges, either in inaccurate predictions of sensitive attributes or\nthe need to mitigate unequal distribution of manually defined non-sensitive\nattributes related to bias. The latter requires strong assumptions about the\ncorrelation between sensitive and non-sensitive attributes. As data\ndistribution and task goals vary, the strong assumption on non-sensitive\nattributes may not be valid and require domain expertise. In this work, we\npropose an assumption-free framework to detect the related attributes\nautomatically by modeling feature interaction for bias mitigation. The proposed\nframework aims to mitigate the unfair impact of identified biased feature\ninteractions. Experimental results on four real-world datasets demonstrate that\nour proposed framework can significantly alleviate unfair prediction behaviors\nby considering biased feature interactions.\n", "date": "2023-07-09", "categories": []}
{"id": "2307.04138", "title": "On The Impact of Machine Learning Randomness on Group Fairness", "abstract": "  Statistical measures for group fairness in machine learning reflect the gap\nin performance of algorithms across different groups. These measures, however,\nexhibit a high variance between different training instances, which makes them\nunreliable for empirical evaluation of fairness. What causes this high\nvariance? We investigate the impact on group fairness of different sources of\nrandomness in training neural networks. We show that the variance in group\nfairness measures is rooted in the high volatility of the learning process on\nunder-represented groups. Further, we recognize the dominant source of\nrandomness as the stochasticity of data order during training. Based on these\nfindings, we show how one can control group-level accuracy (i.e., model\nfairness), with high efficiency and negligible impact on the model's overall\nperformance, by simply changing the data order for a single epoch.\n", "date": "2023-07-09", "categories": []}
{"id": "2307.04303", "title": "Learning to Generate Equitable Text in Dialogue from Biased Training\n  Data", "abstract": "  The ingrained principles of fairness in a dialogue system's decision-making\nprocess and generated responses are crucial for user engagement, satisfaction,\nand task achievement. Absence of equitable and inclusive principles can hinder\nthe formation of common ground, which in turn negatively impacts the overall\nperformance of the system. For example, misusing pronouns in a user interaction\nmay cause ambiguity about the intended subject. Yet, there is no comprehensive\nstudy of equitable text generation in dialogue. Aptly, in this work, we use\ntheories of computational learning to study this problem. We provide formal\ndefinitions of equity in text generation, and further, prove formal connections\nbetween learning human-likeness and learning equity: algorithms for improving\nequity ultimately reduce to algorithms for improving human-likeness (on\naugmented data). With this insight, we also formulate reasonable conditions\nunder which text generation algorithms can learn to generate equitable text\nwithout any modifications to the biased training data on which they learn. To\nexemplify our theory in practice, we look at a group of algorithms for the\nGuessWhat?! visual dialogue game and, using this example, test our theory\nempirically. Our theory accurately predicts relative-performance of multiple\nalgorithms in generating equitable text as measured by both human and automated\nevaluation.\n", "date": "2023-07-10", "categories": []}
{"id": "2301.00719", "title": "Detection of Groups with Biased Representation in Ranking", "abstract": "  Real-life tools for decision-making in many critical domains are based on\nranking results. With the increasing awareness of algorithmic fairness, recent\nworks have presented measures for fairness in ranking. Many of those\ndefinitions consider the representation of different ``protected groups'', in\nthe top-$k$ ranked items, for any reasonable $k$. Given the protected groups,\nconfirming algorithmic fairness is a simple task. However, the groups'\ndefinitions may be unknown in advance. In this paper, we study the problem of\ndetecting groups with biased representation in the top-$k$ ranked items,\neliminating the need to pre-define protected groups. The number of such groups\npossible can be exponential, making the problem hard. We propose efficient\nsearch algorithms for two different fairness measures: global representation\nbounds, and proportional representation. Then we propose a method to explain\nthe bias in the representations of groups utilizing the notion of Shapley\nvalues. We conclude with an experimental study, showing the scalability of our\napproach and demonstrating the usefulness of the proposed algorithms.\n", "date": "2022-12-30", "categories": []}
{"id": "2307.03360", "title": "Evaluating Biased Attitude Associations of Language Models in an\n  Intersectional Context", "abstract": "  Language models are trained on large-scale corpora that embed implicit biases\ndocumented in psychology. Valence associations (pleasantness/unpleasantness) of\nsocial groups determine the biased attitudes towards groups and concepts in\nsocial cognition. Building on this established literature, we quantify how\nsocial groups are valenced in English language models using a sentence template\nthat provides an intersectional context. We study biases related to age,\neducation, gender, height, intelligence, literacy, race, religion, sex, sexual\norientation, social class, and weight. We present a concept projection approach\nto capture the valence subspace through contextualized word embeddings of\nlanguage models. Adapting the projection-based approach to embedding\nassociation tests that quantify bias, we find that language models exhibit the\nmost biased attitudes against gender identity, social class, and sexual\norientation signals in language. We find that the largest and better-performing\nmodel that we study is also more biased as it effectively captures bias\nembedded in sociocultural data. We validate the bias evaluation method by\noverperforming on an intrinsic valence evaluation task. The approach enables us\nto measure complex intersectional biases as they are known to manifest in the\noutputs and applications of language models that perpetuate historical biases.\nMoreover, our approach contributes to design justice as it studies the\nassociations of groups underrepresented in language such as transgender and\nhomosexual individuals.\n", "date": "2023-07-07", "categories": []}
{"id": "2305.08283", "title": "From Pretraining Data to Language Models to Downstream Tasks: Tracking\n  the Trails of Political Biases Leading to Unfair NLP Models", "abstract": "  Language models (LMs) are pretrained on diverse data sources, including news,\ndiscussion forums, books, and online encyclopedias. A significant portion of\nthis data includes opinions and perspectives which, on one hand, celebrate\ndemocracy and diversity of ideas, and on the other hand are inherently socially\nbiased. Our work develops new methods to (1) measure political biases in LMs\ntrained on such corpora, along social and economic axes, and (2) measure the\nfairness of downstream NLP models trained on top of politically biased LMs. We\nfocus on hate speech and misinformation detection, aiming to empirically\nquantify the effects of political (social, economic) biases in pretraining data\non the fairness of high-stakes social-oriented tasks. Our findings reveal that\npretrained LMs do have political leanings that reinforce the polarization\npresent in pretraining corpora, propagating social biases into hate speech\npredictions and misinformation detectors. We discuss the implications of our\nfindings for NLP research and propose future directions to mitigate unfairness.\n", "date": "2023-05-15", "categories": []}
{"id": "2307.02726", "title": "Through the Fairness Lens: Experimental Analysis and Evaluation of\n  Entity Matching", "abstract": "  Entity matching (EM) is a challenging problem studied by different\ncommunities for over half a century. Algorithmic fairness has also become a\ntimely topic to address machine bias and its societal impacts. Despite\nextensive research on these two topics, little attention has been paid to the\nfairness of entity matching.\n  Towards addressing this gap, we perform an extensive experimental evaluation\nof a variety of EM techniques in this paper. We generated two social datasets\nfrom publicly available datasets for the purpose of auditing EM through the\nlens of fairness. Our findings underscore potential unfairness under two common\nconditions in real-world societies: (i) when some demographic groups are\noverrepresented, and (ii) when names are more similar in some groups compared\nto others. Among our many findings, it is noteworthy to mention that while\nvarious fairness definitions are valuable for different settings, due to EM's\nclass imbalance nature, measures such as positive predictive value parity and\ntrue positive rate parity are, in general, more capable of revealing EM\nunfairness.\n", "date": "2023-07-06", "categories": []}
{"id": "2212.02090", "title": "Breaking the Spurious Causality of Conditional Generation via Fairness\n  Intervention with Corrective Sampling", "abstract": "  To capture the relationship between samples and labels, conditional\ngenerative models often inherit spurious correlations from the training\ndataset. This can result in label-conditional distributions that are imbalanced\nwith respect to another latent attribute. To mitigate this issue, which we call\nspurious causality of conditional generation, we propose a general two-step\nstrategy. (a) Fairness Intervention (FI): emphasize the minority samples that\nare hard to generate due to the spurious correlation in the training dataset.\n(b) Corrective Sampling (CS): explicitly filter the generated samples and\nensure that they follow the desired latent attribute distribution. We have\ndesigned the fairness intervention to work for various degrees of supervision\non the spurious attribute, including unsupervised, weakly-supervised, and\nsemi-supervised scenarios. Our experimental results demonstrate that FICS can\neffectively resolve spurious causality of conditional generation across various\ndatasets.\n", "date": "2022-12-05", "categories": []}
{"id": "2303.08325", "title": "FairAdaBN: Mitigating unfairness with adaptive batch normalization and\n  its application to dermatological disease classification", "abstract": "  Deep learning is becoming increasingly ubiquitous in medical research and\napplications while involving sensitive information and even critical diagnosis\ndecisions. Researchers observe a significant performance disparity among\nsubgroups with different demographic attributes, which is called model\nunfairness, and put lots of effort into carefully designing elegant\narchitectures to address unfairness, which poses heavy training burden, brings\npoor generalization, and reveals the trade-off between model performance and\nfairness. To tackle these issues, we propose FairAdaBN by making batch\nnormalization adaptive to sensitive attribute. This simple but effective design\ncan be adopted to several classification backbones that are originally unaware\nof fairness. Additionally, we derive a novel loss function that restrains\nstatistical parity between subgroups on mini-batches, encouraging the model to\nconverge with considerable fairness. In order to evaluate the trade-off between\nmodel performance and fairness, we propose a new metric, named\nFairness-Accuracy Trade-off Efficiency (FATE), to compute normalized fairness\nimprovement over accuracy drop. Experiments on two dermatological datasets show\nthat our proposed method outperforms other methods on fairness criteria and\nFATE.\n", "date": "2023-03-15", "categories": []}
{"id": "2206.00667", "title": "How Biased are Your Features?: Computing Fairness Influence Functions\n  with Global Sensitivity Analysis", "abstract": "  Fairness in machine learning has attained significant focus due to the\nwidespread application in high-stake decision-making tasks. Unregulated machine\nlearning classifiers can exhibit bias towards certain demographic groups in\ndata, thus the quantification and mitigation of classifier bias is a central\nconcern in fairness in machine learning. In this paper, we aim to quantify the\ninfluence of different features in a dataset on the bias of a classifier. To do\nthis, we introduce the Fairness Influence Function (FIF). This function breaks\ndown bias into its components among individual features and the intersection of\nmultiple features. The key idea is to represent existing group fairness metrics\nas the difference of the scaled conditional variances in the classifier's\nprediction and apply a decomposition of variance according to global\nsensitivity analysis. To estimate FIFs, we instantiate an algorithm\nFairXplainer that applies variance decomposition of classifier's prediction\nfollowing local regression. Experiments demonstrate that FairXplainer captures\nFIFs of individual feature and intersectional features, provides a better\napproximation of bias based on FIFs, demonstrates higher correlation of FIFs\nwith fairness interventions, and detects changes in bias due to fairness\naffirmative/punitive actions in the classifier.\n  The code is available at https://github.com/ReAILe/bias-explainer.\n", "date": "2022-06-01", "categories": []}
{"id": "2307.00472", "title": "Equal Confusion Fairness: Measuring Group-Based Disparities in Automated\n  Decision Systems", "abstract": "  As artificial intelligence plays an increasingly substantial role in\ndecisions affecting humans and society, the accountability of automated\ndecision systems has been receiving increasing attention from researchers and\npractitioners. Fairness, which is concerned with eliminating unjust treatment\nand discrimination against individuals or sensitive groups, is a critical\naspect of accountability. Yet, for evaluating fairness, there is a plethora of\nfairness metrics in the literature that employ different perspectives and\nassumptions that are often incompatible. This work focuses on group fairness.\nMost group fairness metrics desire a parity between selected statistics\ncomputed from confusion matrices belonging to different sensitive groups.\nGeneralizing this intuition, this paper proposes a new equal confusion fairness\ntest to check an automated decision system for fairness and a new confusion\nparity error to quantify the extent of any unfairness. To further analyze the\nsource of potential unfairness, an appropriate post hoc analysis methodology is\nalso presented. The usefulness of the test, metric, and post hoc analysis is\ndemonstrated via a case study on the controversial case of COMPAS, an automated\ndecision system employed in the US to assist judges with assessing recidivism\nrisks. Overall, the methods and metrics provided here may assess automated\ndecision systems' fairness as part of a more extensive accountability\nassessment, such as those based on the system accountability benchmark.\n", "date": "2023-07-02", "categories": []}
{"id": "2306.17575", "title": "Augmenting Holistic Review in University Admission using Natural\n  Language Processing for Essays and Recommendation Letters", "abstract": "  University admission at many highly selective institutions uses a holistic\nreview process, where all aspects of the application, including protected\nattributes (e.g., race, gender), grades, essays, and recommendation letters are\nconsidered, to compose an excellent and diverse class. In this study, we\nempirically evaluate how influential protected attributes are for predicting\nadmission decisions using a machine learning (ML) model, and in how far textual\ninformation (e.g., personal essay, teacher recommendation) may substitute for\nthe loss of protected attributes in the model. Using data from 14,915\napplicants to an undergraduate admission office at a selective U.S. institution\nin the 2022-2023 cycle, we find that the exclusion of protected attributes from\nthe ML model leads to substantially reduced admission-prediction performance.\nThe inclusion of textual information via both a TF-IDF representation and a\nLatent Dirichlet allocation (LDA) model partially restores model performance,\nbut does not appear to provide a full substitute for admitting a similarly\ndiverse class. In particular, while the text helps with gender diversity, the\nproportion of URM applicants is severely impacted by the exclusion of protected\nattributes, and the inclusion of new attributes generated from the textual\ninformation does not recover this performance loss.\n", "date": "2023-06-30", "categories": []}
{"id": "2304.13933", "title": "Oversampling Higher-Performing Minorities During Machine Learning Model\n  Training Reduces Adverse Impact Slightly but Also Reduces Model Accuracy", "abstract": "  Organizations are increasingly adopting machine learning (ML) for personnel\nassessment. However, concerns exist about fairness in designing and\nimplementing ML assessments. Supervised ML models are trained to model patterns\nin data, meaning ML models tend to yield predictions that reflect subgroup\ndifferences in applicant attributes in the training data, regardless of the\nunderlying cause of subgroup differences. In this study, we systematically\nunder- and oversampled minority (Black and Hispanic) applicants to manipulate\nadverse impact ratios in training data and investigated how training data\nadverse impact ratios affect ML model adverse impact and accuracy. We used\nself-reports and interview transcripts from job applicants (N = 2,501) to train\n9,702 ML models to predict screening decisions. We found that training data\nadverse impact related linearly to ML model adverse impact. However, removing\nadverse impact from training data only slightly reduced ML model adverse impact\nand tended to negatively affect ML model accuracy. We observed consistent\neffects across self-reports and interview transcripts, whether oversampling\nreal (i.e., bootstrapping) or synthetic observations. As our study relied on\nlimited predictor sets from one organization, the observed effects on adverse\nimpact may be attenuated among more accurate ML models.\n", "date": "2023-04-27", "categories": []}
{"id": "2305.09035", "title": "Algorithmic Censoring in Dynamic Learning Systems", "abstract": "  Dynamic learning systems subject to selective labeling exhibit censoring,\ni.e. persistent negative predictions assigned to one or more subgroups of\npoints. In applications like consumer finance, this results in groups of\napplicants that are persistently denied and thus never enter into the training\ndata. In this work, we formalize censoring, demonstrate how it can arise, and\nhighlight difficulties in detection. We consider safeguards against censoring -\nrecourse and randomized-exploration - both of which ensure we collect labels\nfor points that would otherwise go unobserved. The resulting techniques allow\nexamples from censored groups to enter into the training data and correct the\nmodel. Our results highlight the otherwise unmeasured harms of censoring and\ndemonstrate the effectiveness of mitigation strategies across a range of data\ngenerating processes.\n", "date": "2023-05-15", "categories": []}
{"id": "2306.16552", "title": "Learning Fair Classifiers via Min-Max F-divergence Regularization", "abstract": "  As machine learning (ML) based systems are adopted in domains such as law\nenforcement, criminal justice, finance, hiring and admissions, ensuring the\nfairness of ML aided decision-making is becoming increasingly important. In\nthis paper, we focus on the problem of fair classification, and introduce a\nnovel min-max F-divergence regularization framework for learning fair\nclassification models while preserving high accuracy. Our framework consists of\ntwo trainable networks, namely, a classifier network and a bias/fairness\nestimator network, where the fairness is measured using the statistical notion\nof F-divergence. We show that F-divergence measures possess convexity and\ndifferentiability properties, and their variational representation make them\nwidely applicable in practical gradient based training methods. The proposed\nframework can be readily adapted to multiple sensitive attributes and for high\ndimensional datasets. We study the F-divergence based training paradigm for two\ntypes of group fairness constraints, namely, demographic parity and equalized\nodds. We present a comprehensive set of experiments for several real-world data\nsets arising in multiple domains (including COMPAS, Law Admissions, Adult\nIncome, and CelebA datasets). To quantify the fairness-accuracy tradeoff, we\nintroduce the notion of fairness-accuracy receiver operating characteristic\n(FA-ROC) and a corresponding \\textit{low-bias} FA-ROC, which we argue is an\nappropriate measure to evaluate different classifiers. In comparison to several\nexisting approaches for learning fair classifiers (including pre-processing,\npost-processing and other regularization methods), we show that the proposed\nF-divergence based framework achieves state-of-the-art performance with respect\nto the trade-off between accuracy and fairness.\n", "date": "2023-06-28", "categories": []}
{"id": "2306.15994", "title": "Systematic analysis of the impact of label noise correction on ML\n  Fairness", "abstract": "  Arbitrary, inconsistent, or faulty decision-making raises serious concerns,\nand preventing unfair models is an increasingly important challenge in Machine\nLearning. Data often reflect past discriminatory behavior, and models trained\non such data may reflect bias on sensitive attributes, such as gender, race, or\nage. One approach to developing fair models is to preprocess the training data\nto remove the underlying biases while preserving the relevant information, for\nexample, by correcting biased labels. While multiple label noise correction\nmethods are available, the information about their behavior in identifying\ndiscrimination is very limited. In this work, we develop an empirical\nmethodology to systematically evaluate the effectiveness of label noise\ncorrection techniques in ensuring the fairness of models trained on biased\ndatasets. Our methodology involves manipulating the amount of label noise and\ncan be used with fairness benchmarks but also with standard ML datasets. We\napply the methodology to analyze six label noise correction methods according\nto several fairness metrics on standard OpenML datasets. Our results suggest\nthat the Hybrid Label Noise Correction method achieves the best trade-off\nbetween predictive performance and fairness. Clustering-Based Correction can\nreduce discrimination the most, however, at the cost of lower predictive\nperformance.\n", "date": "2023-06-28", "categories": []}
{"id": "2306.16244", "title": "CBBQ: A Chinese Bias Benchmark Dataset Curated with Human-AI\n  Collaboration for Large Language Models", "abstract": "  Holistically measuring societal biases of large language models is crucial\nfor detecting and reducing ethical risks in highly capable AI models. In this\nwork, we present a Chinese Bias Benchmark dataset that consists of over 100K\nquestions jointly constructed by human experts and generative language models,\ncovering stereotypes and societal biases in 14 social dimensions related to\nChinese culture and values. The curation process contains 4 essential steps:\nbias identification via extensive literature review, ambiguous context\ngeneration, AI-assisted disambiguous context generation, snd manual review \\&\nrecomposition. The testing instances in the dataset are automatically derived\nfrom 3K+ high-quality templates manually authored with stringent quality\ncontrol. The dataset exhibits wide coverage and high diversity. Extensive\nexperiments demonstrate the effectiveness of the dataset in detecting model\nbias, with all 10 publicly available Chinese large language models exhibiting\nstrong bias in certain categories. Additionally, we observe from our\nexperiments that fine-tuned models could, to a certain extent, heed\ninstructions and avoid generating outputs that are morally harmful in some\ntypes, in the way of \"moral self-correction\". Our dataset and results are\npublicly available at\n\\href{https://github.com/YFHuangxxxx/CBBQ}{https://github.com/YFHuangxxxx/CBBQ},\noffering debiasing research opportunities to a widened community.\n", "date": "2023-06-28", "categories": []}
{"id": "2306.14978", "title": "Fairness Aware Counterfactuals for Subgroups", "abstract": "  In this work, we present Fairness Aware Counterfactuals for Subgroups\n(FACTS), a framework for auditing subgroup fairness through counterfactual\nexplanations. We start with revisiting (and generalizing) existing notions and\nintroducing new, more refined notions of subgroup fairness. We aim to (a)\nformulate different aspects of the difficulty of individuals in certain\nsubgroups to achieve recourse, i.e. receive the desired outcome, either at the\nmicro level, considering members of the subgroup individually, or at the macro\nlevel, considering the subgroup as a whole, and (b) introduce notions of\nsubgroup fairness that are robust, if not totally oblivious, to the cost of\nachieving recourse. We accompany these notions with an efficient,\nmodel-agnostic, highly parameterizable, and explainable framework for\nevaluating subgroup fairness. We demonstrate the advantages, the wide\napplicability, and the efficiency of our approach through a thorough\nexperimental evaluation of different benchmark datasets.\n", "date": "2023-06-26", "categories": []}
{"id": "2306.15299", "title": "FAIRER: Fairness as Decision Rationale Alignment", "abstract": "  Deep neural networks (DNNs) have made significant progress, but often suffer\nfrom fairness issues, as deep models typically show distinct accuracy\ndifferences among certain subgroups (e.g., males and females). Existing\nresearch addresses this critical issue by employing fairness-aware loss\nfunctions to constrain the last-layer outputs and directly regularize DNNs.\nAlthough the fairness of DNNs is improved, it is unclear how the trained\nnetwork makes a fair prediction, which limits future fairness improvements. In\nthis paper, we investigate fairness from the perspective of decision rationale\nand define the parameter parity score to characterize the fair decision process\nof networks by analyzing neuron influence in various subgroups. Extensive\nempirical studies show that the unfair issue could arise from the unaligned\ndecision rationales of subgroups. Existing fairness regularization terms fail\nto achieve decision rationale alignment because they only constrain last-layer\noutputs while ignoring intermediate neuron alignment. To address the issue, we\nformulate the fairness as a new task, i.e., decision rationale alignment that\nrequires DNNs' neurons to have consistent responses on subgroups at both\nintermediate processes and the final prediction. To make this idea practical\nduring optimization, we relax the naive objective function and propose\ngradient-guided parity alignment, which encourages gradient-weighted\nconsistency of neurons across subgroups. Extensive experiments on a variety of\ndatasets show that our method can significantly enhance fairness while\nsustaining a high level of accuracy and outperforming other approaches by a\nwide margin.\n", "date": "2023-06-27", "categories": []}
{"id": "2306.15567", "title": "A Three-Way Knot: Privacy, Fairness, and Predictive Performance Dynamics", "abstract": "  As the frontier of machine learning applications moves further into human\ninteraction, multiple concerns arise regarding automated decision-making. Two\nof the most critical issues are fairness and data privacy. On the one hand, one\nmust guarantee that automated decisions are not biased against certain groups,\nespecially those unprotected or marginalized. On the other hand, one must\nensure that the use of personal information fully abides by privacy regulations\nand that user identities are kept safe. The balance between privacy, fairness,\nand predictive performance is complex. However, despite their potential\nsocietal impact, we still demonstrate a poor understanding of the dynamics\nbetween these optimization vectors. In this paper, we study this three-way\ntension and how the optimization of each vector impacts others, aiming to\ninform the future development of safe applications. In light of claims that\npredictive performance and fairness can be jointly optimized, we find this is\nonly possible at the expense of data privacy. Overall, experimental results\nshow that one of the vectors will be penalized regardless of which of the three\nwe optimize. Nonetheless, we find promising avenues for future work in joint\noptimization solutions, where smaller trade-offs are observed between the three\nvectors.\n", "date": "2023-06-27", "categories": []}
{"id": "2206.05828", "title": "Bounding and Approximating Intersectional Fairness through Marginal\n  Fairness", "abstract": "  Discrimination in machine learning often arises along multiple dimensions\n(a.k.a. protected attributes); it is then desirable to ensure\n\\emph{intersectional fairness} -- i.e., that no subgroup is discriminated\nagainst. It is known that ensuring \\emph{marginal fairness} for every dimension\nindependently is not sufficient in general. Due to the exponential number of\nsubgroups, however, directly measuring intersectional fairness from data is\nimpossible. In this paper, our primary goal is to understand in detail the\nrelationship between marginal and intersectional fairness through statistical\nanalysis. We first identify a set of sufficient conditions under which an exact\nrelationship can be obtained. Then, we prove bounds (easily computable through\nmarginal fairness and other meaningful statistical quantities) in\nhigh-probability on intersectional fairness in the general case. Beyond their\ndescriptive value, we show that these theoretical bounds can be leveraged to\nderive a heuristic improving the approximation and bounds of intersectional\nfairness by choosing, in a relevant manner, protected attributes for which we\ndescribe intersectional subgroups. Finally, we test the performance of our\napproximations and bounds on real and synthetic data-sets.\n", "date": "2022-06-12", "categories": []}
{"id": "2306.13659", "title": "Toward A Logical Theory Of Fairness and Bias", "abstract": "  Fairness in machine learning is of considerable interest in recent years\nowing to the propensity of algorithms trained on historical data to amplify and\nperpetuate historical biases. In this paper, we argue for a formal\nreconstruction of fairness definitions, not so much to replace existing\ndefinitions but to ground their application in an epistemic setting and allow\nfor rich environmental modelling. Consequently we look into three notions:\nfairness through unawareness, demographic parity and counterfactual fairness,\nand formalise these in the epistemic situation calculus.\n", "date": "2023-06-08", "categories": []}
{"id": "2306.13738", "title": "Multi-Target Multiplicity: Flexibility and Fairness in Target\n  Specification under Resource Constraints", "abstract": "  Prediction models have been widely adopted as the basis for decision-making\nin domains as diverse as employment, education, lending, and health. Yet, few\nreal world problems readily present themselves as precisely formulated\nprediction tasks. In particular, there are often many reasonable target\nvariable options. Prior work has argued that this is an important and sometimes\nunderappreciated choice, and has also shown that target choice can have a\nsignificant impact on the fairness of the resulting model. However, the\nexisting literature does not offer a formal framework for characterizing the\nextent to which target choice matters in a particular task. Our work fills this\ngap by drawing connections between the problem of target choice and recent work\non predictive multiplicity. Specifically, we introduce a conceptual and\ncomputational framework for assessing how the choice of target affects\nindividuals' outcomes and selection rate disparities across groups. We call\nthis multi-target multiplicity. Along the way, we refine the study of\nsingle-target multiplicity by introducing notions of multiplicity that respect\nresource constraints -- a feature of many real-world tasks that is not captured\nby existing notions of predictive multiplicity. We apply our methods on a\nhealthcare dataset, and show that the level of multiplicity that stems from\ntarget variable choice can be greater than that stemming from nearly-optimal\nmodels of a single target.\n", "date": "2023-06-23", "categories": []}
{"id": "2306.14123", "title": "Privacy and Fairness in Federated Learning: on the Perspective of\n  Trade-off", "abstract": "  Federated learning (FL) has been a hot topic in recent years. Ever since it\nwas introduced, researchers have endeavored to devise FL systems that protect\nprivacy or ensure fair results, with most research focusing on one or the\nother. As two crucial ethical notions, the interactions between privacy and\nfairness are comparatively less studied. However, since privacy and fairness\ncompete, considering each in isolation will inevitably come at the cost of the\nother. To provide a broad view of these two critical topics, we presented a\ndetailed literature review of privacy and fairness issues, highlighting unique\nchallenges posed by FL and solutions in federated settings. We further\nsystematically surveyed different interactions between privacy and fairness,\ntrying to reveal how privacy and fairness could affect each other and point out\nnew research directions in fair and private FL.\n", "date": "2023-06-25", "categories": []}
{"id": "2209.07912", "title": "Algorithmic decision making methods for fair credit scoring", "abstract": "  The effectiveness of machine learning in evaluating the creditworthiness of\nloan applicants has been demonstrated for a long time. However, there is\nconcern that the use of automated decision-making processes may result in\nunequal treatment of groups or individuals, potentially leading to\ndiscriminatory outcomes. This paper seeks to address this issue by evaluating\nthe effectiveness of 12 leading bias mitigation methods across 5 different\nfairness metrics, as well as assessing their accuracy and potential\nprofitability for financial institutions. Through our analysis, we have\nidentified the challenges associated with achieving fairness while maintaining\naccuracy and profitabiliy, and have highlighted both the most successful and\nleast successful mitigation methods. Ultimately, our research serves to bridge\nthe gap between experimental machine learning and its practical applications in\nthe finance industry.\n", "date": "2022-09-16", "categories": []}
{"id": "2306.12912", "title": "Mitigating Discrimination in Insurance with Wasserstein Barycenters", "abstract": "  The insurance industry is heavily reliant on predictions of risks based on\ncharacteristics of potential customers. Although the use of said models is\ncommon, researchers have long pointed out that such practices perpetuate\ndiscrimination based on sensitive features such as gender or race. Given that\nsuch discrimination can often be attributed to historical data biases, an\nelimination or at least mitigation is desirable. With the shift from more\ntraditional models to machine-learning based predictions, calls for greater\nmitigation have grown anew, as simply excluding sensitive variables in the\npricing process can be shown to be ineffective. In this article, we first\ninvestigate why predictions are a necessity within the industry and why\ncorrecting biases is not as straightforward as simply identifying a sensitive\nvariable. We then propose to ease the biases through the use of Wasserstein\nbarycenters instead of simple scaling. To demonstrate the effects and\neffectiveness of the approach we employ it on real data and discuss its\nimplications.\n", "date": "2023-06-22", "categories": []}
{"id": "2306.13064", "title": "Auditing Predictive Models for Intersectional Biases", "abstract": "  Predictive models that satisfy group fairness criteria in aggregate for\nmembers of a protected class, but do not guarantee subgroup fairness, could\nproduce biased predictions for individuals at the intersection of two or more\nprotected classes. To address this risk, we propose Conditional Bias Scan\n(CBS), a flexible auditing framework for detecting intersectional biases in\nclassification models. CBS identifies the subgroup for which there is the most\nsignificant bias against the protected class, as compared to the equivalent\nsubgroup in the non-protected class, and can incorporate multiple commonly used\nfairness definitions for both probabilistic and binarized predictions. We show\nthat this methodology can detect previously unidentified intersectional and\ncontextual biases in the COMPAS pre-trial risk assessment tool and has higher\nbias detection power compared to similar methods that audit for subgroup\nfairness.\n", "date": "2023-06-22", "categories": []}
{"id": "2306.11964", "title": "Sampling Individually-Fair Rankings that are Always Group Fair", "abstract": "  Rankings on online platforms help their end-users find the relevant\ninformation -- people, news, media, and products -- quickly. Fair ranking\ntasks, which ask to rank a set of items to maximize utility subject to\nsatisfying group-fairness constraints, have gained significant interest in the\nAlgorithmic Fairness, Information Retrieval, and Machine Learning literature.\nRecent works, however, identify uncertainty in the utilities of items as a\nprimary cause of unfairness and propose introducing randomness in the output.\nThis randomness is carefully chosen to guarantee an adequate representation of\neach item (while accounting for the uncertainty). However, due to this\nrandomness, the output rankings may violate group fairness constraints. We give\nan efficient algorithm that samples rankings from an individually-fair\ndistribution while ensuring that every output ranking is group fair. The\nexpected utility of the output ranking is at least $\\alpha$ times the utility\nof the optimal fair solution. Here, $\\alpha$ depends on the utilities,\nposition-discounts, and constraints -- it approaches 1 as the range of\nutilities or the position-discounts shrinks, or when utilities satisfy\ndistributional assumptions. Empirically, we observe that our algorithm achieves\nindividual and group fairness and that Pareto dominates the state-of-the-art\nbaselines.\n", "date": "2023-06-21", "categories": []}
{"id": "2209.13012", "title": "Survey on Fairness Notions and Related Tensions", "abstract": "  Automated decision systems are increasingly used to take consequential\ndecisions in problems such as job hiring and loan granting with the hope of\nreplacing subjective human decisions with objective machine learning (ML)\nalgorithms. However, ML-based decision systems are prone to bias, which results\nin yet unfair decisions. Several notions of fairness have been defined in the\nliterature to capture the different subtleties of this ethical and social\nconcept (e.g., statistical parity, equal opportunity, etc.). Fairness\nrequirements to be satisfied while learning models created several types of\ntensions among the different notions of fairness and other desirable properties\nsuch as privacy and classification accuracy. This paper surveys the commonly\nused fairness notions and discusses the tensions among them with privacy and\naccuracy. Different methods to address the fairness-accuracy trade-off\n(classified into four approaches, namely, pre-processing, in-processing,\npost-processing, and hybrid) are reviewed. The survey is consolidated with\nexperimental analysis carried out on fairness benchmark datasets to illustrate\nthe relationship between fairness measures and accuracy in real-world\nscenarios.\n", "date": "2022-09-16", "categories": []}
{"id": "2210.10040", "title": "The Tail Wagging the Dog: Dataset Construction Biases of Social Bias\n  Benchmarks", "abstract": "  How reliably can we trust the scores obtained from social bias benchmarks as\nfaithful indicators of problematic social biases in a given language model? In\nthis work, we study this question by contrasting social biases with non-social\nbiases stemming from choices made during dataset construction that might not\neven be discernible to the human eye. To do so, we empirically simulate various\nalternative constructions for a given benchmark based on innocuous\nmodifications (such as paraphrasing or random-sampling) that maintain the\nessence of their social bias. On two well-known social bias benchmarks\n(Winogender and BiasNLI) we observe that these shallow modifications have a\nsurprising effect on the resulting degree of bias across various models. We\nhope these troubling observations motivate more robust measures of social\nbiases.\n", "date": "2022-10-18", "categories": []}
{"id": "2301.03758", "title": "Sequential Fair Resource Allocation under a Markov Decision Process\n  Framework", "abstract": "  We study the sequential decision-making problem of allocating a limited\nresource to agents that reveal their stochastic demands on arrival over a\nfinite horizon. Our goal is to design fair allocation algorithms that exhaust\nthe available resource budget. This is challenging in sequential settings where\ninformation on future demands is not available at the time of decision-making.\nWe formulate the problem as a discrete time Markov decision process (MDP). We\npropose a new algorithm, SAFFE, that makes fair allocations with respect to the\nentire demands revealed over the horizon by accounting for expected future\ndemands at each arrival time. The algorithm introduces regularization which\nenables the prioritization of current revealed demands over future potential\ndemands depending on the uncertainty in agents' future demands. Using the MDP\nformulation, we show that SAFFE optimizes allocations based on an upper bound\non the Nash Social Welfare fairness objective, and we bound its gap to\noptimality with the use of concentration bounds on total future demands. Using\nsynthetic and real data, we compare the performance of SAFFE against existing\napproaches and a reinforcement learning policy trained on the MDP. We show that\nSAFFE leads to more fair and efficient allocations and achieves\nclose-to-optimal performance in settings with dense arrivals.\n", "date": "2023-01-10", "categories": []}
{"id": "2302.06321", "title": "Parameter-efficient Modularised Bias Mitigation via AdapterFusion", "abstract": "  Large pre-trained language models contain societal biases and carry along\nthese biases to downstream tasks. Current in-processing bias mitigation\napproaches (like adversarial training) impose debiasing by updating a model's\nparameters, effectively transferring the model to a new, irreversible debiased\nstate. In this work, we propose a novel approach to develop stand-alone\ndebiasing functionalities separate from the model, which can be integrated into\nthe model on-demand, while keeping the core model untouched. Drawing from the\nconcept of AdapterFusion in multi-task learning, we introduce DAM (Debiasing\nwith Adapter Modules) - a debiasing approach to first encapsulate arbitrary\nbias mitigation functionalities into separate adapters, and then add them to\nthe model on-demand in order to deliver fairness qualities. We conduct a large\nset of experiments on three classification tasks with gender, race, and age as\nprotected attributes. Our results show that DAM improves or maintains the\neffectiveness of bias mitigation, avoids catastrophic forgetting in a\nmulti-attribute scenario, and maintains on-par task performance, while granting\nparameter-efficiency and easy switching between the original and debiased\nmodels.\n", "date": "2023-02-13", "categories": []}
{"id": "2306.10530", "title": "Gender Bias in Transformer Models: A comprehensive survey", "abstract": "  Gender bias in artificial intelligence (AI) has emerged as a pressing concern\nwith profound implications for individuals' lives. This paper presents a\ncomprehensive survey that explores gender bias in Transformer models from a\nlinguistic perspective. While the existence of gender bias in language models\nhas been acknowledged in previous studies, there remains a lack of consensus on\nhow to effectively measure and evaluate this bias. Our survey critically\nexamines the existing literature on gender bias in Transformers, shedding light\non the diverse methodologies and metrics employed to assess bias. Several\nlimitations in current approaches to measuring gender bias in Transformers are\nidentified, encompassing the utilization of incomplete or flawed metrics,\ninadequate dataset sizes, and a dearth of standardization in evaluation\nmethods. Furthermore, our survey delves into the potential ramifications of\ngender bias in Transformers for downstream applications, including dialogue\nsystems and machine translation. We underscore the importance of fostering\nequity and fairness in these systems by emphasizing the need for heightened\nawareness and accountability in developing and deploying language technologies.\nThis paper serves as a comprehensive overview of gender bias in Transformer\nmodels, providing novel insights and offering valuable directions for future\nresearch in this critical domain.\n", "date": "2023-06-18", "categories": []}
{"id": "2306.10919", "title": "Fairness Index Measures to Evaluate Bias in Biometric Recognition", "abstract": "  The demographic disparity of biometric systems has led to serious concerns\nregarding their societal impact as well as applicability of such systems in\nprivate and public domains. A quantitative evaluation of demographic fairness\nis an important step towards understanding, assessment, and mitigation of\ndemographic bias in biometric applications. While few, existing fairness\nmeasures are based on post-decision data (such as verification accuracy) of\nbiometric systems, we discuss how pre-decision data (score distributions)\nprovide useful insights towards demographic fairness. In this paper, we\nintroduce multiple measures, based on the statistical characteristics of score\ndistributions, for the evaluation of demographic fairness of a generic\nbiometric verification system. We also propose different variants for each\nfairness measure depending on how the contribution from constituent demographic\ngroups needs to be combined towards the final measure. In each case, the\nbehavior of the measure has been illustrated numerically and graphically on\nsynthetic data. The demographic imbalance in benchmarking datasets is often\noverlooked during fairness assessment. We provide a novel weighing strategy to\nreduce the effect of such imbalance through a non-linear function of sample\nsizes of demographic groups. The proposed measures are independent of the\nbiometric modality, and thus, applicable across commonly used biometric\nmodalities (e.g., face, fingerprint, etc.).\n", "date": "2023-06-19", "categories": []}
{"id": "2306.11132", "title": "Fairness-aware Message Passing for Graph Neural Networks", "abstract": "  Graph Neural Networks (GNNs) have shown great power in various domains.\nHowever, their predictions may inherit societal biases on sensitive attributes,\nlimiting their adoption in real-world applications. Although many efforts have\nbeen taken for fair GNNs, most existing works just adopt widely used fairness\ntechniques in machine learning to graph domains and ignore or don't have a\nthorough understanding of the message passing mechanism with fairness\nconstraints, which is a distinctive feature of GNNs. To fill the gap, we\npropose a novel fairness-aware message passing framework GMMD, which is derived\nfrom an optimization problem that considers both graph smoothness and\nrepresentation fairness. GMMD can be intuitively interpreted as encouraging a\nnode to aggregate representations of other nodes from different sensitive\ngroups while subtracting representations of other nodes from the same sensitive\ngroup, resulting in fair representations. We also provide a theoretical\nanalysis to justify that GMMD can guarantee fairness, which leads to a simpler\nand theory-guided variant GMMD-S. Extensive experiments on graph benchmarks\nshow that our proposed framework can significantly improve the fairness of\nvarious backbone GNN models while maintaining high accuracy.\n", "date": "2023-06-19", "categories": []}
{"id": "2306.11181", "title": "Insufficiently Justified Disparate Impact: A New Criterion for Subgroup\n  Fairness", "abstract": "  In this paper, we develop a new criterion, \"insufficiently justified\ndisparate impact\" (IJDI), for assessing whether recommendations (binarized\npredictions) made by an algorithmic decision support tool are fair. Our novel,\nutility-based IJDI criterion evaluates false positive and false negative error\nrate imbalances, identifying statistically significant disparities between\ngroups which are present even when adjusting for group-level differences in\nbase rates. We describe a novel IJDI-Scan approach which can efficiently\nidentify the intersectional subpopulations, defined across multiple observed\nattributes of the data, with the most significant IJDI. To evaluate IJDI-Scan's\nperformance, we conduct experiments on both simulated and real-world data,\nincluding recidivism risk assessment and credit scoring. Further, we implement\nand evaluate approaches to mitigating IJDI for the detected subpopulations in\nthese domains.\n", "date": "2023-06-19", "categories": []}
{"id": "2302.03810", "title": "Fairness in Matching under Uncertainty", "abstract": "  The prevalence and importance of algorithmic two-sided marketplaces has drawn\nattention to the issue of fairness in such settings. Algorithmic decisions are\nused in assigning students to schools, users to advertisers, and applicants to\njob interviews. These decisions should heed the preferences of individuals, and\nsimultaneously be fair with respect to their merits (synonymous with fit,\nfuture performance, or need). Merits conditioned on observable features are\nalways \\emph{uncertain}, a fact that is exacerbated by the widespread use of\nmachine learning algorithms to infer merit from the observables. As our key\ncontribution, we carefully axiomatize a notion of individual fairness in the\ntwo-sided marketplace setting which respects the uncertainty in the merits;\nindeed, it simultaneously recognizes uncertainty as the primary potential cause\nof unfairness and an approach to address it. We design a linear programming\nframework to find fair utility-maximizing distributions over allocations, and\nwe show that the linear program is robust to perturbations in the estimated\nparameters of the uncertain merit distributions, a key property in combining\nthe approach with machine learning techniques.\n", "date": "2023-02-08", "categories": []}
{"id": "2306.09425", "title": "Arbitrariness Lies Beyond the Fairness-Accuracy Frontier", "abstract": "  Machine learning tasks may admit multiple competing models that achieve\nsimilar performance yet produce conflicting outputs for individual samples -- a\nphenomenon known as predictive multiplicity. We demonstrate that fairness\ninterventions in machine learning optimized solely for group fairness and\naccuracy can exacerbate predictive multiplicity. Consequently, state-of-the-art\nfairness interventions can mask high predictive multiplicity behind favorable\ngroup fairness and accuracy metrics. We argue that a third axis of\n``arbitrariness'' should be considered when deploying models to aid\ndecision-making in applications of individual-level impact. To address this\nchallenge, we propose an ensemble algorithm applicable to any fairness\nintervention that provably ensures more consistent predictions.\n", "date": "2023-06-15", "categories": []}
{"id": "2306.09752", "title": "Politeness Stereotypes and Attack Vectors: Gender Stereotypes in\n  Japanese and Korean Language Models", "abstract": "  In efforts to keep up with the rapid progress and use of large language\nmodels, gender bias research is becoming more prevalent in NLP. Non-English\nbias research, however, is still in its infancy with most work focusing on\nEnglish. In our work, we study how grammatical gender bias relating to\npoliteness levels manifests in Japanese and Korean language models. Linguistic\nstudies in these languages have identified a connection between gender bias and\npoliteness levels, however it is not yet known if language models reproduce\nthese biases. We analyze relative prediction probabilities of the male and\nfemale grammatical genders using templates and find that informal polite speech\nis most indicative of the female grammatical gender, while rude and formal\nspeech is most indicative of the male grammatical gender. Further, we find\npoliteness levels to be an attack vector for allocational gender bias in\ncyberbullying detection models. Cyberbullies can evade detection through simple\ntechniques abusing politeness levels. We introduce an attack dataset to (i)\nidentify representational gender bias across politeness levels, (ii)\ndemonstrate how gender biases can be abused to bypass cyberbullying detection\nmodels and (iii) show that allocational biases can be mitigated via training on\nour proposed dataset. Through our findings we highlight the importance of bias\nresearch moving beyond its current English-centrism.\n", "date": "2023-06-16", "categories": []}
{"id": "2306.09835", "title": "Subset Selection Based On Multiple Rankings in the Presence of Bias:\n  Effectiveness of Fairness Constraints for Multiwinner Voting Score Functions", "abstract": "  We consider the problem of subset selection where one is given multiple\nrankings of items and the goal is to select the highest ``quality'' subset.\nScore functions from the multiwinner voting literature have been used to\naggregate rankings into quality scores for subsets. We study this setting of\nsubset selection problems when, in addition, rankings may contain systemic or\nunconscious biases toward a group of items. For a general model of input\nrankings and biases, we show that requiring the selected subset to satisfy\ngroup fairness constraints can improve the quality of the selection with\nrespect to unbiased rankings. Importantly, we show that for fairness\nconstraints to be effective, different multiwinner score functions may require\na drastically different number of rankings: While for some functions, fairness\nconstraints need an exponential number of rankings to recover a\nclose-to-optimal solution, for others, this dependency is only polynomial. This\nresult relies on a novel notion of ``smoothness'' of submodular functions in\nthis setting that quantifies how well a function can ``correctly'' assess the\nquality of items in the presence of bias. The results in this paper can be used\nto guide the choice of multiwinner score functions for the subset selection\nsetting considered here; we additionally provide a tool to empirically enable\nthis.\n", "date": "2023-06-16", "categories": []}
{"id": "2306.08394", "title": "Compatibility of Fairness Metrics with EU Non-Discrimination Laws:\n  Demographic Parity & Conditional Demographic Disparity", "abstract": "  Empirical evidence suggests that algorithmic decisions driven by Machine\nLearning (ML) techniques threaten to discriminate against legally protected\ngroups or create new sources of unfairness. This work supports the contextual\napproach to fairness in EU non-discrimination legal framework and aims at\nassessing up to what point we can assure legal fairness through fairness\nmetrics and under fairness constraints. For that, we analyze the legal notion\nof non-discrimination and differential treatment with the fairness definition\nDemographic Parity (DP) through Conditional Demographic Disparity (CDD). We\ntrain and compare different classifiers with fairness constraints to assess\nwhether it is possible to reduce bias in the prediction while enabling the\ncontextual approach to judicial interpretation practiced under EU\nnon-discrimination laws. Our experimental results on three scenarios show that\nthe in-processing bias mitigation algorithm leads to different performances in\neach of them. Our experiments and analysis suggest that AI-assisted\ndecision-making can be fair from a legal perspective depending on the case at\nhand and the legal justification. These preliminary results encourage future\nwork which will involve further case studies, metrics, and fairness notions.\n", "date": "2023-06-14", "categories": []}
{"id": "2306.07427", "title": "Towards Fair and Explainable AI using a Human-Centered AI Approach", "abstract": "  The rise of machine learning (ML) is accompanied by several high-profile\ncases that have stressed the need for fairness, accountability, explainability\nand trust in ML systems. The existing literature has largely focused on fully\nautomated ML approaches that try to optimize for some performance metric.\nHowever, human-centric measures like fairness, trust, explainability, etc. are\nsubjective in nature, context-dependent, and might not correlate with\nconventional performance metrics. To deal with these challenges, we explore a\nhuman-centered AI approach that empowers people by providing more transparency\nand human control.\n  In this dissertation, we present 5 research projects that aim to enhance\nexplainability and fairness in classification systems and word embeddings. The\nfirst project explores the utility/downsides of introducing local model\nexplanations as interfaces for machine teachers (crowd workers). Our study\nfound that adding explanations supports trust calibration for the resulting ML\nmodel and enables rich forms of teaching feedback. The second project presents\nD-BIAS, a causality-based human-in-the-loop visual tool for identifying and\nmitigating social biases in tabular datasets. Apart from fairness, we found\nthat our tool also enhances trust and accountability. The third project\npresents WordBias, a visual interactive tool that helps audit pre-trained\nstatic word embeddings for biases against groups, such as females, or\nsubgroups, such as Black Muslim females. The fourth project presents DramatVis\nPersonae, a visual analytics tool that helps identify social biases in creative\nwriting. Finally, the last project presents an empirical study aimed at\nunderstanding the cumulative impact of multiple fairness-enhancing\ninterventions at different stages of the ML pipeline on fairness, utility and\ndifferent population groups. We conclude by discussing some of the future\ndirections.\n", "date": "2023-06-12", "categories": []}
{"id": "2306.07500", "title": "Adding guardrails to advanced chatbots", "abstract": "  Generative AI models continue to become more powerful. The launch of ChatGPT\nin November 2022 has ushered in a new era of AI. ChatGPT and other similar\nchatbots have a range of capabilities, from answering student homework\nquestions to creating music and art. There are already concerns that humans may\nbe replaced by chatbots for a variety of jobs. Because of the wide spectrum of\ndata chatbots are built on, we know that they will have human errors and human\nbiases built into them. These biases may cause significant harm and/or inequity\ntoward different subpopulations. To understand the strengths and weakness of\nchatbot responses, we present a position paper that explores different use\ncases of ChatGPT to determine the types of questions that are answered fairly\nand the types that still need improvement. We find that ChatGPT is a fair\nsearch engine for the tasks we tested; however, it has biases on both text\ngeneration and code generation. We find that ChatGPT is very sensitive to\nchanges in the prompt, where small changes lead to different levels of\nfairness. This suggests that we need to immediately implement \"corrections\" or\nmitigation strategies in order to improve fairness of these systems. We suggest\ndifferent strategies to improve chatbots and also advocate for an impartial\nreview panel that has access to the model parameters to measure the levels of\ndifferent types of biases and then recommends safeguards that move toward\nresponses that are less discriminatory and more accurate.\n", "date": "2023-06-13", "categories": []}
{"id": "2211.07350", "title": "Does Debiasing Inevitably Degrade the Model Performance", "abstract": "  Gender bias in language models has attracted sufficient attention because it\nthreatens social justice. However, most of the current debiasing methods\ndegraded the model's performance on other tasks while the degradation mechanism\nis still mysterious. We propose a theoretical framework explaining the three\ncandidate mechanisms of the language model's gender bias. We use our\ntheoretical framework to explain why the current debiasing methods cause\nperformance degradation. We also discover a pathway through which debiasing\nwill not degrade the model performance. We further develop a\ncausality-detection fine-tuning approach to correct gender bias. The numerical\nexperiment demonstrates that our method is able to lead to double dividends:\npartially mitigating gender bias while avoiding performance degradation.\n", "date": "2022-11-14", "categories": []}
{"id": "2212.10563", "title": "BLIND: Bias Removal With No Demographics", "abstract": "  Models trained on real-world data tend to imitate and amplify social biases.\nCommon methods to mitigate biases require prior information on the types of\nbiases that should be mitigated (e.g., gender or racial bias) and the social\ngroups associated with each data sample. In this work, we introduce BLIND, a\nmethod for bias removal with no prior knowledge of the demographics in the\ndataset. While training a model on a downstream task, BLIND detects biased\nsamples using an auxiliary model that predicts the main model's success, and\ndown-weights those samples during the training process. Experiments with racial\nand gender biases in sentiment classification and occupation classification\ntasks demonstrate that BLIND mitigates social biases without relying on a\ncostly demographic annotation process. Our method is competitive with other\nmethods that require demographic information and sometimes even surpasses them.\n", "date": "2022-12-20", "categories": []}
{"id": "2301.13443", "title": "Retiring $\\Delta$DP: New Distribution-Level Metrics for Demographic\n  Parity", "abstract": "  Demographic parity is the most widely recognized measure of group fairness in\nmachine learning, which ensures equal treatment of different demographic\ngroups. Numerous works aim to achieve demographic parity by pursuing the\ncommonly used metric $\\Delta DP$. Unfortunately, in this paper, we reveal that\nthe fairness metric $\\Delta DP$ can not precisely measure the violation of\ndemographic parity, because it inherently has the following drawbacks: i)\nzero-value $\\Delta DP$ does not guarantee zero violation of demographic parity,\nii) $\\Delta DP$ values can vary with different classification thresholds. To\nthis end, we propose two new fairness metrics, Area Between Probability density\nfunction Curves (ABPC) and Area Between Cumulative density function Curves\n(ABCC), to precisely measure the violation of demographic parity at the\ndistribution level. The new fairness metrics directly measure the difference\nbetween the distributions of the prediction probability for different\ndemographic groups. Thus our proposed new metrics enjoy: i) zero-value\nABCC/ABPC guarantees zero violation of demographic parity; ii) ABCC/ABPC\nguarantees demographic parity while the classification thresholds are adjusted.\nWe further re-evaluate the existing fair models with our proposed fairness\nmetrics and observe different fairness behaviors of those models under the new\nmetrics. The code is available at\nhttps://github.com/ahxt/new_metric_for_demographic_parity\n", "date": "2023-01-31", "categories": []}
{"id": "2302.10908", "title": "Human-Centric Multimodal Machine Learning: Recent Advances and Testbed\n  on AI-based Recruitment", "abstract": "  The presence of decision-making algorithms in society is rapidly increasing\nnowadays, while concerns about their transparency and the possibility of these\nalgorithms becoming new sources of discrimination are arising. There is a\ncertain consensus about the need to develop AI applications with a\nHuman-Centric approach. Human-Centric Machine Learning needs to be developed\nbased on four main requirements: (i) utility and social good; (ii) privacy and\ndata ownership; (iii) transparency and accountability; and (iv) fairness in\nAI-driven decision-making processes. All these four Human-Centric requirements\nare closely related to each other. With the aim of studying how current\nmultimodal algorithms based on heterogeneous sources of information are\naffected by sensitive elements and inner biases in the data, we propose a\nfictitious case study focused on automated recruitment: FairCVtest. We train\nautomatic recruitment algorithms using a set of multimodal synthetic profiles\nincluding image, text, and structured data, which are consciously scored with\ngender and racial biases. FairCVtest shows the capacity of the Artificial\nIntelligence (AI) behind automatic recruitment tools built this way (a common\npractice in many other application scenarios beyond recruitment) to extract\nsensitive information from unstructured data and exploit it in combination to\ndata biases in undesirable (unfair) ways. We present an overview of recent\nworks developing techniques capable of removing sensitive information and\nbiases from the decision-making process of deep learning architectures, as well\nas commonly used databases for fairness research in AI. We demonstrate how\nlearning approaches developed to guarantee privacy in latent spaces can lead to\nunbiased and fair automatic decision-making process.\n", "date": "2023-02-13", "categories": []}
{"id": "2305.10201", "title": "Echoes of Biases: How Stigmatizing Language Affects AI Performance", "abstract": "  Electronic health records (EHRs) serve as an essential data source for the\nenvisioned artificial intelligence (AI)-driven transformation in healthcare.\nHowever, clinician biases reflected in EHR notes can lead to AI models\ninheriting and amplifying these biases, perpetuating health disparities. This\nstudy investigates the impact of stigmatizing language (SL) in EHR notes on\nmortality prediction using a Transformer-based deep learning model and\nexplainable AI (XAI) techniques. Our findings demonstrate that SL written by\nclinicians adversely affects AI performance, particularly so for black\npatients, highlighting SL as a source of racial disparity in AI model\ndevelopment. To explore an operationally efficient way to mitigate SL's impact,\nwe investigate patterns in the generation of SL through a clinicians'\ncollaborative network, identifying central clinicians as having a stronger\nimpact on racial disparity in the AI model. We find that removing SL written by\ncentral clinicians is a more efficient bias reduction strategy than eliminating\nall SL in the entire corpus of data. This study provides actionable insights\nfor responsible AI development and contributes to understanding clinician\nbehavior and EHR note writing in healthcare.\n", "date": "2023-05-17", "categories": []}
{"id": "2306.06135", "title": "Safety and Fairness for Content Moderation in Generative Models", "abstract": "  With significant advances in generative AI, new technologies are rapidly\nbeing deployed with generative components. Generative models are typically\ntrained on large datasets, resulting in model behaviors that can mimic the\nworst of the content in the training data. Responsible deployment of generative\ntechnologies requires content moderation strategies, such as safety input and\noutput filters. Here, we provide a theoretical framework for conceptualizing\nresponsible content moderation of text-to-image generative technologies,\nincluding a demonstration of how to empirically measure the constructs we\nenumerate. We define and distinguish the concepts of safety, fairness, and\nmetric equity, and enumerate example harms that can come in each domain. We\nthen provide a demonstration of how the defined harms can be quantified. We\nconclude with a summary of how the style of harms quantification we demonstrate\nenables data-driven content moderation decisions.\n", "date": "2023-06-09", "categories": []}
{"id": "2306.06542", "title": "Investigating Practices and Opportunities for Cross-functional\n  Collaboration around AI Fairness in Industry Practice", "abstract": "  An emerging body of research indicates that ineffective cross-functional\ncollaboration -- the interdisciplinary work done by industry practitioners\nacross roles -- represents a major barrier to addressing issues of fairness in\nAI design and development. In this research, we sought to better understand\npractitioners' current practices and tactics to enact cross-functional\ncollaboration for AI fairness, in order to identify opportunities to support\nmore effective collaboration. We conducted a series of interviews and design\nworkshops with 23 industry practitioners spanning various roles from 17\ncompanies. We found that practitioners engaged in bridging work to overcome\nfrictions in understanding, contextualization, and evaluation around AI\nfairness across roles. In addition, in organizational contexts with a lack of\nresources and incentives for fairness work, practitioners often piggybacked on\nexisting requirements (e.g., for privacy assessments) and AI development norms\n(e.g., the use of quantitative evaluation metrics), although they worry that\nthese tactics may be fundamentally compromised. Finally, we draw attention to\nthe invisible labor that practitioners take on as part of this bridging and\npiggybacking work to enact interdisciplinary collaboration for fairness. We\nclose by discussing opportunities for both FAccT researchers and AI\npractitioners to better support cross-functional collaboration for fairness in\nthe design and development of AI systems.\n", "date": "2023-06-10", "categories": []}
{"id": "2306.06696", "title": "Toward Fair Facial Expression Recognition with Improved Distribution\n  Alignment", "abstract": "  We present a novel approach to mitigate bias in facial expression recognition\n(FER) models. Our method aims to reduce sensitive attribute information such as\ngender, age, or race, in the embeddings produced by FER models. We employ a\nkernel mean shrinkage estimator to estimate the kernel mean of the\ndistributions of the embeddings associated with different sensitive attribute\ngroups, such as young and old, in the Hilbert space. Using this estimation, we\ncalculate the maximum mean discrepancy (MMD) distance between the distributions\nand incorporate it in the classifier loss along with an adversarial loss, which\nis then minimized through the learning process to improve the distribution\nalignment. Our method makes sensitive attributes less recognizable for the\nmodel, which in turn promotes fairness. Additionally, for the first time, we\nanalyze the notion of attractiveness as an important sensitive attribute in FER\nmodels and demonstrate that FER models can indeed exhibit biases towards more\nattractive faces. To prove the efficacy of our model in reducing bias regarding\ndifferent sensitive attributes (including the newly proposed attractiveness\nattribute), we perform several experiments on two widely used datasets, CelebA\nand RAF-DB. The results in terms of both accuracy and fairness measures\noutperform the state-of-the-art in most cases, demonstrating the effectiveness\nof the proposed method.\n", "date": "2023-06-11", "categories": []}
{"id": "2207.12497", "title": "Estimating and Controlling for Equalized Odds via Sensitive Attribute\n  Predictors", "abstract": "  As the use of machine learning models in real world high-stakes decision\nsettings continues to grow, it is highly important that we are able to audit\nand control for any potential fairness violations these models may exhibit\ntowards certain groups. To do so, one naturally requires access to sensitive\nattributes, such as demographics, gender, or other potentially sensitive\nfeatures that determine group membership. Unfortunately, in many settings, this\ninformation is often unavailable. In this work we study the well known\n\\emph{equalized odds} (EOD) definition of fairness. In a setting without\nsensitive attributes, we first provide tight and computable upper bounds for\nthe EOD violation of a predictor. These bounds precisely reflect the worst\npossible EOD violation. Second, we demonstrate how one can provably control the\nworst-case EOD by a new post-processing correction method. Our results\ncharacterize when directly controlling for EOD with respect to the predicted\nsensitive attributes is -- and when is not -- optimal when it comes to\ncontrolling worst-case EOD. Our results hold under assumptions that are milder\nthan previous works, and we illustrate these results with experiments on\nsynthetic and real datasets.\n", "date": "2022-07-25", "categories": []}
{"id": "2306.03293", "title": "Towards Fairness in Personalized Ads Using Impression Variance Aware\n  Reinforcement Learning", "abstract": "  Variances in ad impression outcomes across demographic groups are\nincreasingly considered to be potentially indicative of algorithmic bias in\npersonalized ads systems. While there are many definitions of fairness that\ncould be applicable in the context of personalized systems, we present a\nframework which we call the Variance Reduction System (VRS) for achieving more\nequitable outcomes in Meta's ads systems. VRS seeks to achieve a distribution\nof impressions with respect to selected protected class (PC) attributes that\nmore closely aligns the demographics of an ad's eligible audience (a function\nof advertiser targeting criteria) with the audience who sees that ad, in a\nprivacy-preserving manner. We first define metrics to quantify fairness gaps in\nterms of ad impression variances with respect to PC attributes including gender\nand estimated race. We then present the VRS for re-ranking ads in an impression\nvariance-aware manner. We evaluate VRS via extensive simulations over different\nparameter choices and study the effect of the VRS on the chosen fairness\nmetric. We finally present online A/B testing results from applying VRS to\nMeta's ads systems, concluding with a discussion of future work. We have\ndeployed the VRS to all users in the US for housing ads, resulting in\nsignificant improvement in our fairness metric. VRS is the first large-scale\ndeployed framework for pursuing fairness for multiple PC attributes in online\nadvertising.\n", "date": "2023-06-05", "categories": []}
{"id": "2306.06083", "title": "Improving Fairness and Robustness in End-to-End Speech Recognition\n  through unsupervised clustering", "abstract": "  The challenge of fairness arises when Automatic Speech Recognition (ASR)\nsystems do not perform equally well for all sub-groups of the population. In\nthe past few years there have been many improvements in overall speech\nrecognition quality, but without any particular focus on advancing Equality and\nEquity for all user groups for whom systems do not perform well. ASR fairness\nis therefore also a robustness issue. Meanwhile, data privacy also takes\npriority in production systems. In this paper, we present a privacy preserving\napproach to improve fairness and robustness of end-to-end ASR without using\nmetadata, zip codes, or even speaker or utterance embeddings directly in\ntraining. We extract utterance level embeddings using a speaker ID model\ntrained on a public dataset, which we then use in an unsupervised fashion to\ncreate acoustic clusters. We use cluster IDs instead of speaker utterance\nembeddings as extra features during model training, which shows improvements\nfor all demographic groups and in particular for different accents.\n", "date": "2023-06-06", "categories": []}
{"id": "2210.07213", "title": "FARE: Provably Fair Representation Learning with Practical Certificates", "abstract": "  Fair representation learning (FRL) is a popular class of methods aiming to\nproduce fair classifiers via data preprocessing. Recent regulatory directives\nstress the need for FRL methods that provide practical certificates, i.e.,\nprovable upper bounds on the unfairness of any downstream classifier trained on\npreprocessed data, which directly provides assurance in a practical scenario.\nCreating such FRL methods is an important challenge that remains unsolved. In\nthis work, we address that challenge and introduce FARE (Fairness with\nRestricted Encoders), the first FRL method with practical fairness\ncertificates. FARE is based on our key insight that restricting the\nrepresentation space of the encoder enables the derivation of practical\nguarantees, while still permitting favorable accuracy-fairness tradeoffs for\nsuitable instantiations, such as one we propose based on fair trees. To produce\na practical certificate, we develop and apply a statistical procedure that\ncomputes a finite sample high-confidence upper bound on the unfairness of any\ndownstream classifier trained on FARE embeddings. In our comprehensive\nexperimental evaluation, we demonstrate that FARE produces practical\ncertificates that are tight and often even comparable with purely empirical\nresults obtained by prior methods, which establishes the practical value of our\napproach.\n", "date": "2022-10-13", "categories": []}
{"id": "2212.06641", "title": "Simplicity Bias Leads to Amplified Performance Disparities", "abstract": "  Which parts of a dataset will a given model find difficult? Recent work has\nshown that SGD-trained models have a bias towards simplicity, leading them to\nprioritize learning a majority class, or to rely upon harmful spurious\ncorrelations. Here, we show that the preference for \"easy\" runs far deeper: A\nmodel may prioritize any class or group of the dataset that it finds simple-at\nthe expense of what it finds complex-as measured by performance difference on\nthe test set. When subsets with different levels of complexity align with\ndemographic groups, we term this difficulty disparity, a phenomenon that occurs\neven with balanced datasets that lack group/label associations. We show how\ndifficulty disparity is a model-dependent quantity, and is further amplified in\ncommonly-used models as selected by typical average performance scores. We\nquantify an amplification factor across a range of settings in order to compare\ndisparity of different models on a fixed dataset. Finally, we present two\nreal-world examples of difficulty amplification in action, resulting in\nworse-than-expected performance disparities between groups even when using a\nbalanced dataset. The existence of such disparities in balanced datasets\ndemonstrates that merely balancing sample sizes of groups is not sufficient to\nensure unbiased performance. We hope this work presents a step towards\nmeasurable understanding of the role of model bias as it interacts with the\nstructure of data, and call for additional model-dependent mitigation methods\nto be deployed alongside dataset audits.\n", "date": "2022-12-13", "categories": []}
{"id": "2305.03712", "title": "Statistical Inference for Fairness Auditing", "abstract": "  Before deploying a black-box model in high-stakes problems, it is important\nto evaluate the model's performance on sensitive subpopulations. For example,\nin a recidivism prediction task, we may wish to identify demographic groups for\nwhich our prediction model has unacceptably high false positive rates or\ncertify that no such groups exist. In this paper, we frame this task, often\nreferred to as \"fairness auditing,\" in terms of multiple hypothesis testing. We\nshow how the bootstrap can be used to simultaneously bound performance\ndisparities over a collection of groups with statistical guarantees. Our\nmethods can be used to flag subpopulations affected by model underperformance,\nand certify subpopulations for which the model performs adequately. Crucially,\nour audit is model-agnostic and applicable to nearly any performance metric or\ngroup fairness criterion. Our methods also accommodate extremely rich -- even\ninfinite -- collections of subpopulations. Further, we generalize beyond\nsubpopulations by showing how to assess performance over certain distribution\nshifts. We test the proposed methods on benchmark datasets in predictive\ninference and algorithmic fairness and find that our audits can provide\ninterpretable and trustworthy guarantees.\n", "date": "2023-05-05", "categories": []}
{"id": "2306.05066", "title": "Causal Fairness for Outcome Control", "abstract": "  As society transitions towards an AI-based decision-making infrastructure, an\never-increasing number of decisions once under control of humans are now\ndelegated to automated systems. Even though such developments make various\nparts of society more efficient, a large body of evidence suggests that a great\ndeal of care needs to be taken to make such automated decision-making systems\nfair and equitable, namely, taking into account sensitive attributes such as\ngender, race, and religion. In this paper, we study a specific decision-making\ntask called outcome control in which an automated system aims to optimize an\noutcome variable $Y$ while being fair and equitable. The interest in such a\nsetting ranges from interventions related to criminal justice and welfare, all\nthe way to clinical decision-making and public health. In this paper, we first\nanalyze through causal lenses the notion of benefit, which captures how much a\nspecific individual would benefit from a positive decision, counterfactually\nspeaking, when contrasted with an alternative, negative one. We introduce the\nnotion of benefit fairness, which can be seen as the minimal fairness\nrequirement in decision-making, and develop an algorithm for satisfying it. We\nthen note that the benefit itself may be influenced by the protected attribute,\nand propose causal tools which can be used to analyze this. Finally, if some of\nthe variations of the protected attribute in the benefit are considered as\ndiscriminatory, the notion of benefit fairness may need to be strengthened,\nwhich leads us to articulating a notion of causal benefit fairness. Using this\nnotion, we develop a new optimization procedure capable of maximizing $Y$ while\nascertaining causal fairness in the decision process.\n", "date": "2023-06-08", "categories": []}
{"id": "2306.05068", "title": "Shedding light on underrepresentation and Sampling Bias in machine\n  learning", "abstract": "  Accurately measuring discrimination is crucial to faithfully assessing\nfairness of trained machine learning (ML) models. Any bias in measuring\ndiscrimination leads to either amplification or underestimation of the existing\ndisparity. Several sources of bias exist and it is assumed that bias resulting\nfrom machine learning is born equally by different groups (e.g. females vs\nmales, whites vs blacks, etc.). If, however, bias is born differently by\ndifferent groups, it may exacerbate discrimination against specific\nsub-populations. Sampling bias, is inconsistently used in the literature to\ndescribe bias due to the sampling procedure. In this paper, we attempt to\ndisambiguate this term by introducing clearly defined variants of sampling\nbias, namely, sample size bias (SSB) and underrepresentation bias (URB). We\nshow also how discrimination can be decomposed into variance, bias, and noise.\nFinally, we challenge the commonly accepted mitigation approach that\ndiscrimination can be addressed by collecting more samples of the\nunderrepresented group.\n", "date": "2023-06-08", "categories": []}
{"id": "2306.05307", "title": "Are fairness metric scores enough to assess discrimination biases in\n  machine learning?", "abstract": "  This paper presents novel experiments shedding light on the shortcomings of\ncurrent metrics for assessing biases of gender discrimination made by machine\nlearning algorithms on textual data. We focus on the Bios dataset, and our\nlearning task is to predict the occupation of individuals, based on their\nbiography. Such prediction tasks are common in commercial Natural Language\nProcessing (NLP) applications such as automatic job recommendations. We address\nan important limitation of theoretical discussions dealing with group-wise\nfairness metrics: they focus on large datasets, although the norm in many\nindustrial NLP applications is to use small to reasonably large linguistic\ndatasets for which the main practical constraint is to get a good prediction\naccuracy. We then question how reliable are different popular measures of bias\nwhen the size of the training set is simply sufficient to learn reasonably\naccurate predictions. Our experiments sample the Bios dataset and learn more\nthan 200 models on different sample sizes. This allows us to statistically\nstudy our results and to confirm that common gender bias indices provide\ndiverging and sometimes unreliable results when applied to relatively small\ntraining and test samples. This highlights the crucial importance of variance\ncalculations for providing sound results in this field.\n", "date": "2023-06-08", "categories": []}
{"id": "2211.03759", "title": "Easily Accessible Text-to-Image Generation Amplifies Demographic\n  Stereotypes at Large Scale", "abstract": "  Machine learning models that convert user-written text descriptions into\nimages are now widely available online and used by millions of users to\ngenerate millions of images a day. We investigate the potential for these\nmodels to amplify dangerous and complex stereotypes. We find a broad range of\nordinary prompts produce stereotypes, including prompts simply mentioning\ntraits, descriptors, occupations, or objects. For example, we find cases of\nprompting for basic traits or social roles resulting in images reinforcing\nwhiteness as ideal, prompting for occupations resulting in amplification of\nracial and gender disparities, and prompting for objects resulting in\nreification of American norms. Stereotypes are present regardless of whether\nprompts explicitly mention identity and demographic language or avoid such\nlanguage. Moreover, stereotypes persist despite mitigation strategies; neither\nuser attempts to counter stereotypes by requesting images with specific\ncounter-stereotypes nor institutional attempts to add system ``guardrails''\nhave prevented the perpetuation of stereotypes. Our analysis justifies concerns\nregarding the impacts of today's models, presenting striking exemplars, and\nconnecting these findings with deep insights into harms drawn from social\nscientific and humanist disciplines. This work contributes to the effort to\nshed light on the uniquely complex biases in language-vision models and\ndemonstrates the ways that the mass deployment of text-to-image generation\nmodels results in mass dissemination of stereotypes and resulting harms.\n", "date": "2022-11-07", "categories": []}
{"id": "2302.08077", "title": "Group Fairness with Uncertainty in Sensitive Attributes", "abstract": "  Learning a fair predictive model is crucial to mitigate biased decisions\nagainst minority groups in high-stakes applications. A common approach to learn\nsuch a model involves solving an optimization problem that maximizes the\npredictive power of the model under an appropriate group fairness constraint.\nHowever, in practice, sensitive attributes are often missing or noisy resulting\nin uncertainty. We demonstrate that solely enforcing fairness constraints on\nuncertain sensitive attributes can fall significantly short in achieving the\nlevel of fairness of models trained without uncertainty. To overcome this\nlimitation, we propose a bootstrap-based algorithm that achieves the target\nlevel of fairness despite the uncertainty in sensitive attributes. The\nalgorithm is guided by a Gaussian analysis for the independence notion of\nfairness where we propose a robust quadratically constrained quadratic problem\nto ensure a strict fairness guarantee with uncertain sensitive attributes. Our\nalgorithm is applicable to both discrete and continuous sensitive attributes\nand is effective in real-world classification and regression tasks for various\ngroup fairness notions, e.g., independence and separation.\n", "date": "2023-02-16", "categories": []}
{"id": "2306.04118", "title": "M$^3$Fair: Mitigating Bias in Healthcare Data through Multi-Level and\n  Multi-Sensitive-Attribute Reweighting Method", "abstract": "  In the data-driven artificial intelligence paradigm, models heavily rely on\nlarge amounts of training data. However, factors like sampling distribution\nimbalance can lead to issues of bias and unfairness in healthcare data.\nSensitive attributes, such as race, gender, age, and medical condition, are\ncharacteristics of individuals that are commonly associated with discrimination\nor bias. In healthcare AI, these attributes can play a significant role in\ndetermining the quality of care that individuals receive. For example, minority\ngroups often receive fewer procedures and poorer-quality medical care than\nwhite individuals in US. Therefore, detecting and mitigating bias in data is\ncrucial to enhancing health equity. Bias mitigation methods include\npre-processing, in-processing, and post-processing. Among them, Reweighting\n(RW) is a widely used pre-processing method that performs well in balancing\nmachine learning performance and fairness performance. RW adjusts the weights\nfor samples within each (group, label) combination, where these weights are\nutilized in loss functions. However, RW is limited to considering only a single\nsensitive attribute when mitigating bias and assumes that each sensitive\nattribute is equally important. This may result in potential inaccuracies when\naddressing intersectional bias. To address these limitations, we propose\nM3Fair, a multi-level and multi-sensitive-attribute reweighting method by\nextending the RW method to multiple sensitive attributes at multiple levels.\nOur experiments on real-world datasets show that the approach is effective,\nstraightforward, and generalizable in addressing the healthcare fairness\nissues.\n", "date": "2023-06-07", "categories": []}
{"id": "2306.04400", "title": "A Fair Classifier Embracing Triplet Collapse", "abstract": "  In this paper, we study the behaviour of the triplet loss and show that it\ncan be exploited to limit the biases created and perpetuated by machine\nlearning models. Our fair classifier uses the collapse of the triplet loss when\nits margin is greater than the maximum distance between two points in the\nlatent space, in the case of stochastic triplet selection.\n", "date": "2023-06-07", "categories": []}
{"id": "2306.04482", "title": "ICON$^2$: Reliably Benchmarking Predictive Inequity in Object Detection", "abstract": "  As computer vision systems are being increasingly deployed at scale in\nhigh-stakes applications like autonomous driving, concerns about social bias in\nthese systems are rising. Analysis of fairness in real-world vision systems,\nsuch as object detection in driving scenes, has been limited to observing\npredictive inequity across attributes such as pedestrian skin tone, and lacks a\nconsistent methodology to disentangle the role of confounding variables e.g.\ndoes my model perform worse for a certain skin tone, or are such scenes in my\ndataset more challenging due to occlusion and crowds? In this work, we\nintroduce ICON$^2$, a framework for robustly answering this question. ICON$^2$\nleverages prior knowledge on the deficiencies of object detection systems to\nidentify performance discrepancies across sub-populations, compute correlations\nbetween these potential confounders and a given sensitive attribute, and\ncontrol for the most likely confounders to obtain a more reliable estimate of\nmodel bias. Using our approach, we conduct an in-depth study on the performance\nof object detection with respect to income from the BDD100K driving dataset,\nrevealing useful insights.\n", "date": "2023-06-07", "categories": []}
{"id": "2306.04597", "title": "Language Models Get a Gender Makeover: Mitigating Gender Bias with\n  Few-Shot Data Interventions", "abstract": "  Societal biases present in pre-trained large language models are a critical\nissue as these models have been shown to propagate biases in countless\ndownstream applications, rendering them unfair towards specific groups of\npeople. Since large-scale retraining of these models from scratch is both time\nand compute-expensive, a variety of approaches have been previously proposed\nthat de-bias a pre-trained model. While the majority of current\nstate-of-the-art debiasing methods focus on changes to the training regime, in\nthis paper, we propose data intervention strategies as a powerful yet simple\ntechnique to reduce gender bias in pre-trained models. Specifically, we\nempirically show that by fine-tuning a pre-trained model on only 10 de-biased\n(intervened) training examples, the tendency to favor any gender is\nsignificantly reduced. Since our proposed method only needs a few training\nexamples, our few-shot debiasing approach is highly feasible and practical.\nThrough extensive experimentation, we show that our debiasing technique\nperforms better than competitive state-of-the-art baselines with minimal loss\nin language modeling ability.\n", "date": "2023-06-07", "categories": []}
{"id": "2211.02139", "title": "Can Querying for Bias Leak Protected Attributes? Achieving Privacy With\n  Smooth Sensitivity", "abstract": "  Existing regulations prohibit model developers from accessing protected\nattributes (gender, race, etc.), often resulting in fairness assessments on\npopulations without knowing their protected groups. In such scenarios,\ninstitutions often adopt a separation between the model developers (who train\nmodels with no access to the protected attributes) and a compliance team (who\nmay have access to the entire dataset for auditing purposes). However, the\nmodel developers might be allowed to test their models for bias by querying the\ncompliance team for group fairness metrics. In this paper, we first demonstrate\nthat simply querying for fairness metrics, such as statistical parity and\nequalized odds can leak the protected attributes of individuals to the model\ndevelopers. We demonstrate that there always exist strategies by which the\nmodel developers can identify the protected attribute of a targeted individual\nin the test dataset from just a single query. In particular, we show that one\ncan reconstruct the protected attributes of all the individuals from O(Nk \\log(\nn /Nk)) queries when Nk<<n using techniques from compressed sensing (n: size of\nthe test dataset, Nk: size of smallest group). Our results pose an interesting\ndebate in algorithmic fairness: should querying for fairness metrics be viewed\nas a neutral-valued solution to ensure compliance with regulations? Or, does it\nconstitute a violation of regulations and privacy if the number of queries\nanswered is enough for the model developers to identify the protected\nattributes of specific individuals? To address this supposed violation, we also\npropose Attribute-Conceal, a novel technique that achieves differential privacy\nby calibrating noise to the smooth sensitivity of our bias query, outperforming\nnaive techniques such as the Laplace mechanism. We also include experimental\nresults on the Adult dataset and synthetic data.\n", "date": "2022-11-03", "categories": []}
{"id": "2301.08987", "title": "Tier Balancing: Towards Dynamic Fairness over Underlying Causal Factors", "abstract": "  The pursuit of long-term fairness involves the interplay between\ndecision-making and the underlying data generating process. In this paper,\nthrough causal modeling with a directed acyclic graph (DAG) on the\ndecision-distribution interplay, we investigate the possibility of achieving\nlong-term fairness from a dynamic perspective. We propose Tier Balancing, a\ntechnically more challenging but more natural notion to achieve in the context\nof long-term, dynamic fairness analysis. Different from previous fairness\nnotions that are defined purely on observed variables, our notion goes one step\nfurther, capturing behind-the-scenes situation changes on the unobserved latent\ncausal factors that directly carry out the influence from the current decision\nto the future data distribution. Under the specified dynamics, we prove that in\ngeneral one cannot achieve the long-term fairness goal only through one-step\ninterventions. Furthermore, in the effort of approaching long-term fairness, we\nconsider the mission of \"getting closer to\" the long-term fairness goal and\npresent possibility and impossibility results accordingly.\n", "date": "2023-01-21", "categories": []}
{"id": "2304.04874", "title": "ImageCaptioner$^2$: Image Captioner for Image Captioning Bias\n  Amplification Assessment", "abstract": "  Most pre-trained learning systems are known to suffer from bias, which\ntypically emerges from the data, the model, or both. Measuring and quantifying\nbias and its sources is a challenging task and has been extensively studied in\nimage captioning. Despite the significant effort in this direction, we observed\nthat existing metrics lack consistency in the inclusion of the visual signal.\nIn this paper, we introduce a new bias assessment metric, dubbed\n$ImageCaptioner^2$, for image captioning. Instead of measuring the absolute\nbias in the model or the data, $ImageCaptioner^2$ pay more attention to the\nbias introduced by the model w.r.t the data bias, termed bias amplification.\nUnlike the existing methods, which only evaluate the image captioning\nalgorithms based on the generated captions only, $ImageCaptioner^2$\nincorporates the image while measuring the bias. In addition, we design a\nformulation for measuring the bias of generated captions as prompt-based image\ncaptioning instead of using language classifiers. Finally, we apply our\n$ImageCaptioner^2$ metric across 11 different image captioning architectures on\nthree different datasets, i.e., MS-COCO caption dataset, Artemis V1, and\nArtemis V2, and on three different protected attributes, i.e., gender, race,\nand emotions. Consequently, we verify the effectiveness of our\n$ImageCaptioner^2$ metric by proposing AnonymousBench, which is a novel human\nevaluation paradigm for bias metrics. Our metric shows significant superiority\nover the recent bias metric; LIC, in terms of human alignment, where the\ncorrelation scores are 80% and 54% for our metric and LIC, respectively. The\ncode is available at https://eslambakr.github.io/imagecaptioner2.github.io/.\n", "date": "2023-04-10", "categories": []}
{"id": "2305.05608", "title": "The Role of Relevance in Fair Ranking", "abstract": "  Online platforms mediate access to opportunity: relevance-based rankings\ncreate and constrain options by allocating exposure to job openings and job\ncandidates in hiring platforms, or sellers in a marketplace. In order to do so\nresponsibly, these socially consequential systems employ various fairness\nmeasures and interventions, many of which seek to allocate exposure based on\nworthiness. Because these constructs are typically not directly observable,\nplatforms must instead resort to using proxy scores such as relevance and infer\nthem from behavioral signals such as searcher clicks. Yet, it remains an open\nquestion whether relevance fulfills its role as such a worthiness score in\nhigh-stakes fair rankings. In this paper, we combine perspectives and tools\nfrom the social sciences, information retrieval, and fairness in machine\nlearning to derive a set of desired criteria that relevance scores should\nsatisfy in order to meaningfully guide fairness interventions. We then\nempirically show that not all of these criteria are met in a case study of\nrelevance inferred from biased user click data. We assess the impact of these\nviolations on the estimated system fairness and analyze whether existing\nfairness interventions may mitigate the identified issues. Our analyses and\nresults surface the pressing need for new approaches to relevance collection\nand generation that are suitable for use in fair ranking.\n", "date": "2023-05-09", "categories": []}
{"id": "2306.03179", "title": "Fair Patient Model: Mitigating Bias in the Patient Representation\n  Learned from the Electronic Health Records", "abstract": "  Objective: To pre-train fair and unbiased patient representations from\nElectronic Health Records (EHRs) using a novel weighted loss function that\nreduces bias and improves fairness in deep representation learning models.\n  Methods: We defined a new loss function, called weighted loss function, in\nthe deep representation learning model to balance the importance of different\ngroups of patients and features. We applied the proposed model, called Fair\nPatient Model (FPM), to a sample of 34,739 patients from the MIMIC-III dataset\nand learned patient representations for four clinical outcome prediction tasks.\n  Results: FPM outperformed the baseline models in terms of three fairness\nmetrics: demographic parity, equality of opportunity difference, and equalized\nodds ratio. FPM also achieved comparable predictive performance with the\nbaselines, with an average accuracy of 0.7912. Feature analysis revealed that\nFPM captured more information from clinical features than the baselines.\n  Conclusion: FPM is a novel method to pre-train fair and unbiased patient\nrepresentations from EHR data using a weighted loss function. The learned\nrepresentations can be used for various downstream tasks in healthcare and can\nbe extended to other domains where bias and fairness are important.\n", "date": "2023-06-05", "categories": []}
{"id": "2306.03773", "title": "Some voices are too common: Building fair speech recognition systems\n  using the Common Voice dataset", "abstract": "  Automatic speech recognition (ASR) systems become increasingly efficient\nthanks to new advances in neural network training like self-supervised\nlearning. However, they are known to be unfair toward certain groups, for\ninstance, people speaking with an accent. In this work, we use the French\nCommon Voice dataset to quantify the biases of a pre-trained wav2vec~2.0 model\ntoward several demographic groups. By fine-tuning the pre-trained model on a\nvariety of fixed-size, carefully crafted training sets, we demonstrate the\nimportance of speaker diversity. We also run an in-depth analysis of the Common\nVoice corpus and identify important shortcomings that should be taken into\naccount by users of this dataset.\n", "date": "2023-06-01", "categories": []}
{"id": "2210.08781", "title": "Stochastic Differentially Private and Fair Learning", "abstract": "  Machine learning models are increasingly used in high-stakes decision-making\nsystems. In such applications, a major concern is that these models sometimes\ndiscriminate against certain demographic groups such as individuals with\ncertain race, gender, or age. Another major concern in these applications is\nthe violation of the privacy of users. While fair learning algorithms have been\ndeveloped to mitigate discrimination issues, these algorithms can still leak\nsensitive information, such as individuals' health or financial records.\nUtilizing the notion of differential privacy (DP), prior works aimed at\ndeveloping learning algorithms that are both private and fair. However,\nexisting algorithms for DP fair learning are either not guaranteed to converge\nor require full batch of data in each iteration of the algorithm to converge.\nIn this paper, we provide the first stochastic differentially private algorithm\nfor fair learning that is guaranteed to converge. Here, the term \"stochastic\"\nrefers to the fact that our proposed algorithm converges even when minibatches\nof data are used at each iteration (i.e. stochastic optimization). Our\nframework is flexible enough to permit different fairness notions, including\ndemographic parity and equalized odds. In addition, our algorithm can be\napplied to non-binary classification tasks with multiple (non-binary) sensitive\nattributes. As a byproduct of our convergence analysis, we provide the first\nutility guarantee for a DP algorithm for solving nonconvex-strongly concave\nmin-max problems. Our numerical experiments show that the proposed algorithm\nconsistently offers significant performance gains over the state-of-the-art\nbaselines, and can be applied to larger scale problems with non-binary\ntarget/sensitive attributes.\n", "date": "2022-10-17", "categories": []}
{"id": "2211.01528", "title": "Fair and Optimal Classification via Post-Processing", "abstract": "  To mitigate the bias exhibited by machine learning models, fairness criteria\ncan be integrated into the training process to ensure fair treatment across all\ndemographics, but it often comes at the expense of model performance.\nUnderstanding such tradeoffs, therefore, underlies the design of fair\nalgorithms. To this end, this paper provides a complete characterization of the\ninherent tradeoff of demographic parity on classification problems, under the\nmost general multi-group, multi-class, and noisy setting. Specifically, we show\nthat the minimum error rate achievable by randomized and attribute-aware fair\nclassifiers is given by the optimal value of a Wasserstein-barycenter problem.\nOn the practical side, our findings lead to a simple post-processing algorithm\nthat derives fair classifiers from score functions, which yields the optimal\nfair classifier when the score is Bayes optimal. We provide suboptimality\nanalysis and sample complexity for our algorithm, and demonstrate its\neffectiveness on benchmark datasets.\n", "date": "2022-11-03", "categories": []}
{"id": "2306.01943", "title": "NLPositionality: Characterizing Design Biases of Datasets and Models", "abstract": "  Design biases in NLP systems, such as performance differences for different\npopulations, often stem from their creator's positionality, i.e., views and\nlived experiences shaped by identity and background. Despite the prevalence and\nrisks of design biases, they are hard to quantify because researcher, system,\nand dataset positionality is often unobserved. We introduce NLPositionality, a\nframework for characterizing design biases and quantifying the positionality of\nNLP datasets and models. Our framework continuously collects annotations from a\ndiverse pool of volunteer participants on LabintheWild, and statistically\nquantifies alignment with dataset labels and model predictions. We apply\nNLPositionality to existing datasets and models for two tasks -- social\nacceptability and hate speech detection. To date, we have collected 16,299\nannotations in over a year for 600 instances from 1,096 annotators across 87\ncountries. We find that datasets and models align predominantly with Western,\nWhite, college-educated, and younger populations. Additionally, certain groups,\nsuch as non-binary people and non-native English speakers, are further\nmarginalized by datasets and models as they rank least in alignment across all\ntasks. Finally, we draw from prior literature to discuss how researchers can\nexamine their own positionality and that of their datasets and models, opening\nthe door for more inclusive NLP systems.\n", "date": "2023-06-02", "categories": []}
{"id": "2206.04101", "title": "What-is and How-to for Fairness in Machine Learning: A Survey,\n  Reflection, and Perspective", "abstract": "  Algorithmic fairness has attracted increasing attention in the machine\nlearning community. Various definitions are proposed in the literature, but the\ndifferences and connections among them are not clearly addressed. In this\npaper, we review and reflect on various fairness notions previously proposed in\nmachine learning literature, and make an attempt to draw connections to\narguments in moral and political philosophy, especially theories of justice. We\nalso consider fairness inquiries from a dynamic perspective, and further\nconsider the long-term impact that is induced by current prediction and\ndecision. In light of the differences in the characterized fairness, we present\na flowchart that encompasses implicit assumptions and expected outcomes of\ndifferent types of fairness inquiries on the data generating process, on the\npredicted outcome, and on the induced impact, respectively. This paper\ndemonstrates the importance of matching the mission (which kind of fairness one\nwould like to enforce) and the means (which spectrum of fairness analysis is of\ninterest, what is the appropriate analyzing scheme) to fulfill the intended\npurpose.\n", "date": "2022-06-08", "categories": []}
{"id": "2305.09941", "title": "\"I'm fully who I am\": Towards Centering Transgender and Non-Binary\n  Voices to Measure Biases in Open Language Generation", "abstract": "  Transgender and non-binary (TGNB) individuals disproportionately experience\ndiscrimination and exclusion from daily life. Given the recent popularity and\nadoption of language generation technologies, the potential to further\nmarginalize this population only grows. Although a multitude of NLP fairness\nliterature focuses on illuminating and addressing gender biases, assessing\ngender harms for TGNB identities requires understanding how such identities\nuniquely interact with societal gender norms and how they differ from gender\nbinary-centric perspectives. Such measurement frameworks inherently require\ncentering TGNB voices to help guide the alignment between gender-inclusive NLP\nand whom they are intended to serve. Towards this goal, we ground our work in\nthe TGNB community and existing interdisciplinary literature to assess how the\nsocial reality surrounding experienced marginalization of TGNB persons\ncontributes to and persists within Open Language Generation (OLG). This social\nknowledge serves as a guide for evaluating popular large language models (LLMs)\non two key aspects: (1) misgendering and (2) harmful responses to gender\ndisclosure. To do this, we introduce TANGO, a dataset of template-based\nreal-world text curated from a TGNB-oriented community. We discover a dominance\nof binary gender norms reflected by the models; LLMs least misgendered subjects\nin generated text when triggered by prompts whose subjects used binary\npronouns. Meanwhile, misgendering was most prevalent when triggering generation\nwith singular they and neopronouns. When prompted with gender disclosures, TGNB\ndisclosure generated the most stigmatizing language and scored most toxic, on\naverage. Our findings warrant further research on how TGNB harms manifest in\nLLMs and serve as a broader case study toward concretely grounding the design\nof gender-inclusive AI in community voices and interdisciplinary literature.\n", "date": "2023-05-17", "categories": []}
{"id": "2306.01117", "title": "Examining the Causal Effect of First Names on Language Models: The Case\n  of Social Commonsense Reasoning", "abstract": "  As language models continue to be integrated into applications of personal\nand societal relevance, ensuring these models' trustworthiness is crucial,\nparticularly with respect to producing consistent outputs regardless of\nsensitive attributes. Given that first names may serve as proxies for\n(intersectional) socio-demographic representations, it is imperative to examine\nthe impact of first names on commonsense reasoning capabilities. In this paper,\nwe study whether a model's reasoning given a specific input differs based on\nthe first names provided. Our underlying assumption is that the reasoning about\nAlice should not differ from the reasoning about James. We propose and\nimplement a controlled experimental framework to measure the causal effect of\nfirst names on commonsense reasoning, enabling us to distinguish between model\npredictions due to chance and caused by actual factors of interest. Our results\nindicate that the frequency of first names has a direct effect on model\nprediction, with less frequent names yielding divergent predictions compared to\nmore frequent names. To gain insights into the internal mechanisms of models\nthat are contributing to these behaviors, we also conduct an in-depth\nexplainable analysis. Overall, our findings suggest that to ensure model\nrobustness, it is essential to augment datasets with more diverse first names\nduring the configuration stage.\n", "date": "2023-06-01", "categories": []}
{"id": "2306.01333", "title": "Navigating Fairness in Radiology AI: Concepts, Consequences,and Crucial\n  Considerations", "abstract": "  Artificial Intelligence (AI) has significantly revolutionized radiology,\npromising improved patient outcomes and streamlined processes. However, it's\ncritical to ensure the fairness of AI models to prevent stealthy bias and\ndisparities from leading to unequal outcomes. This review discusses the concept\nof fairness in AI, focusing on bias auditing using the Aequitas toolkit, and\nits real-world implications in radiology, particularly in disease screening\nscenarios. Aequitas, an open-source bias audit toolkit, scrutinizes AI models'\ndecisions, identifying hidden biases that may result in disparities across\ndifferent demographic groups and imaging equipment brands. This toolkit\noperates on statistical theories, analyzing a large dataset to reveal a model's\nfairness. It excels in its versatility to handle various variables\nsimultaneously, especially in a field as diverse as radiology. The review\nexplicates essential fairness metrics: Equal and Proportional Parity, False\nPositive Rate Parity, False Discovery Rate Parity, False Negative Rate Parity,\nand False Omission Rate Parity. Each metric serves unique purposes and offers\ndifferent insights. We present hypothetical scenarios to demonstrate their\nrelevance in disease screening settings, and how disparities can lead to\nsignificant real-world impacts.\n", "date": "2023-06-02", "categories": []}
{"id": "2306.01417", "title": "The Flawed Foundations of Fair Machine Learning", "abstract": "  The definition and implementation of fairness in automated decisions has been\nextensively studied by the research community. Yet, there hides fallacious\nreasoning, misleading assertions, and questionable practices at the foundations\nof the current fair machine learning paradigm. Those flaws are the result of a\nfailure to understand that the trade-off between statistically accurate\noutcomes and group similar outcomes exists as independent, external constraint\nrather than as a subjective manifestation as has been commonly argued. First,\nwe explain that there is only one conception of fairness present in the fair\nmachine learning literature: group similarity of outcomes based on a sensitive\nattribute where the similarity benefits an underprivileged group. Second, we\nshow that there is, in fact, a trade-off between statistically accurate\noutcomes and group similar outcomes in any data setting where group disparities\nexist, and that the trade-off presents an existential threat to the equitable,\nfair machine learning approach. Third, we introduce a proof-of-concept\nevaluation to aid researchers and designers in understanding the relationship\nbetween statistically accurate outcomes and group similar outcomes. Finally,\nsuggestions for future work aimed at data scientists, legal scholars, and data\nethicists that utilize the conceptual and experimental framework described\nthroughout this article are provided.\n", "date": "2023-06-02", "categories": []}
{"id": "2306.01699", "title": "Affinity Clustering Framework for Data Debiasing Using Pairwise\n  Distribution Discrepancy", "abstract": "  Group imbalance, resulting from inadequate or unrepresentative data\ncollection methods, is a primary cause of representation bias in datasets.\nRepresentation bias can exist with respect to different groups of one or more\nprotected attributes and might lead to prejudicial and discriminatory outcomes\ntoward certain groups of individuals; in cases where a learning model is\ntrained on such biased data. This paper presents MASC, a data augmentation\napproach that leverages affinity clustering to balance the representation of\nnon-protected and protected groups of a target dataset by utilizing instances\nof the same protected attributes from similar datasets that are categorized in\nthe same cluster as the target dataset by sharing instances of the protected\nattribute. The proposed method involves constructing an affinity matrix by\nquantifying distribution discrepancies between dataset pairs and transforming\nthem into a symmetric pairwise similarity matrix. A non-parametric spectral\nclustering is then applied to this affinity matrix, automatically categorizing\nthe datasets into an optimal number of clusters. We perform a step-by-step\nexperiment as a demo of our method to show the procedure of the proposed data\naugmentation method and evaluate and discuss its performance. A comparison with\nother data augmentation methods, both pre- and post-augmentation, is conducted,\nalong with a model evaluation analysis of each method. Our method can handle\nnon-binary protected attributes so, in our experiments, bias is measured in a\nnon-binary protected attribute setup w.r.t. racial groups distribution for two\nseparate minority groups in comparison with the majority group before and after\ndebiasing. Empirical results imply that our method of augmenting dataset biases\nusing real (genuine) data from similar contexts can effectively debias the\ntarget datasets comparably to existing data augmentation strategies.\n", "date": "2023-06-02", "categories": []}
{"id": "2306.00025", "title": "Disentangling and Operationalizing AI Fairness at LinkedIn", "abstract": "  Operationalizing AI fairness at LinkedIn's scale is challenging not only\nbecause there are multiple mutually incompatible definitions of fairness but\nalso because determining what is fair depends on the specifics and context of\nthe product where AI is deployed. Moreover, AI practitioners need clarity on\nwhat fairness expectations need to be addressed at the AI level. In this paper,\nwe present the evolving AI fairness framework used at LinkedIn to address these\nthree challenges. The framework disentangles AI fairness by separating out\nequal treatment and equitable product expectations. Rather than imposing a\ntrade-off between these two commonly opposing interpretations of fairness, the\nframework provides clear guidelines for operationalizing equal AI treatment\ncomplemented with a product equity strategy. This paper focuses on the equal AI\ntreatment component of LinkedIn's AI fairness framework, shares the principles\nthat support it, and illustrates their application through a case study. We\nhope this paper will encourage other big tech companies to join us in sharing\ntheir approach to operationalizing AI fairness at scale, so that together we\ncan keep advancing this constantly evolving field.\n", "date": "2023-05-30", "categories": []}
{"id": "2306.00403", "title": "A Survey on Fairness-aware Recommender Systems", "abstract": "  As information filtering services, recommender systems have extremely\nenriched our daily life by providing personalized suggestions and facilitating\npeople in decision-making, which makes them vital and indispensable to human\nsociety in the information era. However, as people become more dependent on\nthem, recent studies show that recommender systems potentially own\nunintentional impacts on society and individuals because of their unfairness\n(e.g., gender discrimination in job recommendations). To develop trustworthy\nservices, it is crucial to devise fairness-aware recommender systems that can\nmitigate these bias issues. In this survey, we summarise existing methodologies\nand practices of fairness in recommender systems. Firstly, we present concepts\nof fairness in different recommendation scenarios, comprehensively categorize\ncurrent advances, and introduce typical methods to promote fairness in\ndifferent stages of recommender systems. Next, after introducing datasets and\nevaluation metrics applied to assess the fairness of recommender systems, we\nwill delve into the significant influence that fairness-aware recommender\nsystems exert on real-world industrial applications. Subsequently, we highlight\nthe connection between fairness and other principles of trustworthy recommender\nsystems, aiming to consider trustworthiness principles holistically while\nadvocating for fairness. Finally, we summarize this review, spotlighting\npromising opportunities in comprehending concepts, frameworks, the balance\nbetween accuracy and fairness, and the ties with trustworthiness, with the\nultimate goal of fostering the development of fairness-aware recommender\nsystems.\n", "date": "2023-06-01", "categories": []}
{"id": "2306.00905", "title": "T2IAT: Measuring Valence and Stereotypical Biases in Text-to-Image\n  Generation", "abstract": "  Warning: This paper contains several contents that may be toxic, harmful, or\noffensive.\n  In the last few years, text-to-image generative models have gained remarkable\nsuccess in generating images with unprecedented quality accompanied by a\nbreakthrough of inference speed. Despite their rapid progress, human biases\nthat manifest in the training examples, particularly with regard to common\nstereotypical biases, like gender and skin tone, still have been found in these\ngenerative models. In this work, we seek to measure more complex human biases\nexist in the task of text-to-image generations. Inspired by the well-known\nImplicit Association Test (IAT) from social psychology, we propose a novel\nText-to-Image Association Test (T2IAT) framework that quantifies the implicit\nstereotypes between concepts and valence, and those in the images. We replicate\nthe previously documented bias tests on generative models, including morally\nneutral tests on flowers and insects as well as demographic stereotypical tests\non diverse social attributes. The results of these experiments demonstrate the\npresence of complex stereotypical behaviors in image generations.\n", "date": "2023-06-01", "categories": []}
{"id": "2206.03656", "title": "Fair Classification via Domain Adaptation: A Dual Adversarial Learning\n  Approach", "abstract": "  Modern machine learning (ML) models are becoming increasingly popular and are\nwidely used in decision-making systems. However, studies have shown critical\nissues of ML discrimination and unfairness, which hinder their adoption on\nhigh-stake applications. Recent research on fair classifiers has drawn\nsignificant attention to developing effective algorithms to achieve fairness\nand good classification performance. Despite the great success of these\nfairness-aware machine learning models, most of the existing models require\nsensitive attributes to pre-process the data, regularize the model learning or\npost-process the prediction to have fair predictions. However, sensitive\nattributes are often incomplete or even unavailable due to privacy, legal or\nregulation restrictions. Though we lack the sensitive attribute for training a\nfair model in the target domain, there might exist a similar domain that has\nsensitive attributes. Thus, it is important to exploit auxiliary information\nfrom a similar domain to help improve fair classification in the target domain.\nTherefore, in this paper, we study a novel problem of exploring domain\nadaptation for fair classification. We propose a new framework that can learn\nto adapt the sensitive attributes from a source domain for fair classification\nin the target domain. Extensive experiments on real-world datasets illustrate\nthe effectiveness of the proposed model for fair classification, even when no\nsensitive attributes are available in the target domain.\n", "date": "2022-06-08", "categories": []}
{"id": "2305.02995", "title": "Accuracy on the Curve: On the Nonlinear Correlation of ML Performance\n  Between Data Subpopulations", "abstract": "  Understanding the performance of machine learning (ML) models across diverse\ndata distributions is critically important for reliable applications. Despite\nrecent empirical studies positing a near-perfect linear correlation between\nin-distribution (ID) and out-of-distribution (OOD) accuracies, we empirically\ndemonstrate that this correlation is more nuanced under subpopulation shifts.\nThrough rigorous experimentation and analysis across a variety of datasets,\nmodels, and training epochs, we demonstrate that OOD performance often has a\nnonlinear correlation with ID performance in subpopulation shifts. Our\nfindings, which contrast previous studies that have posited a linear\ncorrelation in model performance during distribution shifts, reveal a \"moon\nshape\" correlation (parabolic uptrend curve) between the test performance on\nthe majority subpopulation and the minority subpopulation. This non-trivial\nnonlinear correlation holds across model architectures, hyperparameters,\ntraining durations, and the imbalance between subpopulations. Furthermore, we\nfound that the nonlinearity of this \"moon shape\" is causally influenced by the\ndegree of spurious correlations in the training data. Our controlled\nexperiments show that stronger spurious correlation in the training data\ncreates more nonlinear performance correlation. We provide complementary\nexperimental and theoretical analyses for this phenomenon, and discuss its\nimplications for ML reliability and fairness. Our work highlights the\nimportance of understanding the nonlinear effects of model improvement on\nperformance in different subpopulations, and has the potential to inform the\ndevelopment of more equitable and responsible machine learning models.\n", "date": "2023-05-04", "categories": []}
{"id": "2305.19329", "title": "Mitigating Test-Time Bias for Fair Image Retrieval", "abstract": "  We address the challenge of generating fair and unbiased image retrieval\nresults given neutral textual queries (with no explicit gender or race\nconnotations), while maintaining the utility (performance) of the underlying\nvision-language (VL) model. Previous methods aim to disentangle learned\nrepresentations of images and text queries from gender and racial\ncharacteristics. However, we show these are inadequate at alleviating bias for\nthe desired equal representation result, as there usually exists test-time bias\nin the target retrieval set. So motivated, we introduce a straightforward\ntechnique, Post-hoc Bias Mitigation (PBM), that post-processes the outputs from\nthe pre-trained vision-language model. We evaluate our algorithm on real-world\nimage search datasets, Occupation 1 and 2, as well as two large-scale\nimage-text datasets, MS-COCO and Flickr30k. Our approach achieves the lowest\nbias, compared with various existing bias-mitigation methods, in text-based\nimage retrieval result while maintaining satisfactory retrieval performance.\nThe source code is publicly available at\n\\url{https://anonymous.4open.science/r/Fair_Text_based_Image_Retrieval-D8B2}.\n", "date": "2023-05-23", "categories": []}
{"id": "2305.19407", "title": "FRAMM: Fair Ranking with Missing Modalities for Clinical Trial Site\n  Selection", "abstract": "  Despite many efforts to address the disparities, the underrepresentation of\ngender, racial, and ethnic minorities in clinical trials remains a problem and\nundermines the efficacy of treatments on minorities. This paper focuses on the\ntrial site selection task and proposes FRAMM, a deep reinforcement learning\nframework for fair trial site selection. We focus on addressing two real-world\nchallenges that affect fair trial sites selection: the data modalities are\noften not complete for many potential trial sites, and the site selection needs\nto simultaneously optimize for both enrollment and diversity since the problem\nis necessarily a trade-off between the two with the only possible way to\nincrease diversity post-selection being through limiting enrollment via caps.\nTo address the missing data challenge, FRAMM has a modality encoder with a\nmasked cross-attention mechanism for handling missing data, bypassing data\nimputation and the need for complete data in training. To handle the need for\nmaking efficient trade-offs, FRAMM uses deep reinforcement learning with a\nspecifically designed reward function that simultaneously optimizes for both\nenrollment and fairness.\n  We evaluate FRAMM using 4,392 real-world clinical trials ranging from 2016 to\n2021 and show that FRAMM outperforms the leading baseline in enrollment-only\nsettings while also achieving large gains in diversity. Specifically, it is\nable to produce a 9% improvement in diversity with similar enrollment levels\nover the leading baselines. That improved diversity is further manifested in\nachieving up to a 14% increase in Hispanic enrollment, 27% increase in Black\nenrollment, and 60% increase in Asian enrollment compared to selecting sites\nwith an enrollment-only model.\n", "date": "2023-05-30", "categories": []}
{"id": "2305.19409", "title": "Examining risks of racial biases in NLP tools for child protective\n  services", "abstract": "  Although much literature has established the presence of demographic bias in\nnatural language processing (NLP) models, most work relies on curated bias\nmetrics that may not be reflective of real-world applications. At the same\ntime, practitioners are increasingly using algorithmic tools in high-stakes\nsettings, with particular recent interest in NLP. In this work, we focus on one\nsuch setting: child protective services (CPS). CPS workers often write copious\nfree-form text notes about families they are working with, and CPS agencies are\nactively seeking to deploy NLP models to leverage these data. Given\nwell-established racial bias in this setting, we investigate possible ways\ndeployed NLP is liable to increase racial disparities. We specifically examine\nword statistics within notes and algorithmic fairness in risk prediction,\ncoreference resolution, and named entity recognition (NER). We document\nconsistent algorithmic unfairness in NER models, possible algorithmic\nunfairness in coreference resolution models, and little evidence of exacerbated\nracial bias in risk prediction. While there is existing pronounced criticism of\nrisk prediction, our results expose previously undocumented risks of racial\nbias in realistic information extraction systems, highlighting potential\nconcerns in deploying them, even though they may appear more benign. Our work\nserves as a rare realistic examination of NLP algorithmic fairness in a\npotential deployed setting and a timely investigation of a specific risk\nassociated with deploying NLP in CPS settings.\n", "date": "2023-05-30", "categories": []}
{"id": "2305.19475", "title": "Doubly Constrained Fair Clustering", "abstract": "  The remarkable attention which fair clustering has received in the last few\nyears has resulted in a significant number of different notions of fairness.\nDespite the fact that these notions are well-justified, they are often\nmotivated and studied in a disjoint manner where one fairness desideratum is\nconsidered exclusively in isolation from the others. This leaves the\nunderstanding of the relations between different fairness notions as an\nimportant open problem in fair clustering. In this paper, we take the first\nstep in this direction. Specifically, we consider the two most prominent\ndemographic representation fairness notions in clustering: (1) Group Fairness\n(GF), where the different demographic groups are supposed to have close to\npopulation-level representation in each cluster and (2) Diversity in Center\nSelection (DS), where the selected centers are supposed to have close to\npopulation-level representation of each group. We show that given a constant\napproximation algorithm for one constraint (GF or DS only) we can obtain a\nconstant approximation solution that satisfies both constraints simultaneously.\nInterestingly, we prove that any given solution that satisfies the GF\nconstraint can always be post-processed at a bounded degradation to the\nclustering cost to additionally satisfy the DS constraint while the reverse is\nnot true. Furthermore, we show that both GF and DS are incompatible (having an\nempty feasibility set in the worst case) with a collection of other\ndistance-based fairness notions. Finally, we carry experiments to validate our\ntheoretical findings.\n", "date": "2023-05-31", "categories": []}
{"id": "2305.19671", "title": "Signal Is Harder To Learn Than Bias: Debiasing with Focal Loss", "abstract": "  Spurious correlations are everywhere. While humans often do not perceive\nthem, neural networks are notorious for learning unwanted associations, also\nknown as biases, instead of the underlying decision rule. As a result,\npractitioners are often unaware of the biased decision-making of their\nclassifiers. Such a biased model based on spurious correlations might not\ngeneralize to unobserved data, leading to unintended, adverse consequences. We\npropose Signal is Harder (SiH), a variational-autoencoder-based method that\nsimultaneously trains a biased and unbiased classifier using a novel,\ndisentangling reweighting scheme inspired by the focal loss. Using the unbiased\nclassifier, SiH matches or improves upon the performance of state-of-the-art\ndebiasing methods. To improve the interpretability of our technique, we propose\na perturbation scheme in the latent space for visualizing the bias that helps\npractitioners become aware of the sources of spurious correlations.\n", "date": "2023-05-31", "categories": []}
{"id": "2305.20020", "title": "Bias Mitigation Methods for Binary Classification Decision-Making\n  Systems: Survey and Recommendations", "abstract": "  Bias mitigation methods for binary classification decision-making systems\nhave been widely researched due to the ever-growing importance of designing\nfair machine learning processes that are impartial and do not discriminate\nagainst individuals or groups based on protected personal characteristics. In\nthis paper, we present a structured overview of the research landscape for bias\nmitigation methods, report on their benefits and limitations, and provide\nrecommendations for the development of future bias mitigation methods for\nbinary classification.\n", "date": "2023-05-31", "categories": []}
{"id": "2207.08336", "title": "When Fairness Meets Privacy: Fair Classification with Semi-Private\n  Sensitive Attributes", "abstract": "  Machine learning models have demonstrated promising performance in many\nareas. However, the concerns that they can be biased against specific\ndemographic groups hinder their adoption in high-stake applications. Thus, it\nis essential to ensure fairness in machine learning models. Most previous\nefforts require direct access to sensitive attributes for mitigating bias.\nNonetheless, it is often infeasible to obtain large-scale users' sensitive\nattributes considering users' concerns about privacy in the data collection\nprocess. Privacy mechanisms such as local differential privacy (LDP) are widely\nenforced on sensitive information in the data collection stage due to legal\ncompliance and people's increasing awareness of privacy. Therefore, a critical\nproblem is how to make fair predictions under privacy. We study a novel and\npractical problem of fair classification in a semi-private setting, where most\nof the sensitive attributes are private and only a small amount of clean ones\nare available. To this end, we propose a novel framework FairSP that can\nachieve Fair prediction under the Semi-Private setting. First, FairSP learns to\ncorrect the noise-protected sensitive attributes by exploiting the limited\nclean sensitive attributes. Then, it jointly models the corrected and clean\ndata in an adversarial way for debiasing and prediction. Theoretical analysis\nshows that the proposed model can ensure fairness under mild assumptions in the\nsemi-private setting. Extensive experimental results on real-world datasets\ndemonstrate the effectiveness of our method for making fair predictions under\nprivacy and maintaining high accuracy.\n", "date": "2022-07-18", "categories": []}
{"id": "2210.11924", "title": "Men Also Do Laundry: Multi-Attribute Bias Amplification", "abstract": "  As computer vision systems become more widely deployed, there is increasing\nconcern from both the research community and the public that these systems are\nnot only reproducing but amplifying harmful social biases. The phenomenon of\nbias amplification, which is the focus of this work, refers to models\namplifying inherent training set biases at test time. Existing metrics measure\nbias amplification with respect to single annotated attributes (e.g.,\n$\\texttt{computer}$). However, several visual datasets consist of images with\nmultiple attribute annotations. We show models can learn to exploit\ncorrelations with respect to multiple attributes (e.g., {$\\texttt{computer}$,\n$\\texttt{keyboard}$}), which are not accounted for by current metrics. In\naddition, we show current metrics can give the erroneous impression that\nminimal or no bias amplification has occurred as they involve aggregating over\npositive and negative values. Further, these metrics lack a clear desired\nvalue, making them difficult to interpret. To address these shortcomings, we\npropose a new metric: Multi-Attribute Bias Amplification. We validate our\nproposed metric through an analysis of gender bias amplification on the COCO\nand imSitu datasets. Finally, we benchmark bias mitigation methods using our\nproposed metric, suggesting possible avenues for future bias mitigation\n", "date": "2022-10-21", "categories": []}
{"id": "2305.17701", "title": "KoSBi: A Dataset for Mitigating Social Bias Risks Towards Safer Large\n  Language Model Application", "abstract": "  Large language models (LLMs) learn not only natural text generation abilities\nbut also social biases against different demographic groups from real-world\ndata. This poses a critical risk when deploying LLM-based applications.\nExisting research and resources are not readily applicable in South Korea due\nto the differences in language and culture, both of which significantly affect\nthe biases and targeted demographic groups. This limitation requires localized\nsocial bias datasets to ensure the safe and effective deployment of LLMs. To\nthis end, we present KO SB I, a new social bias dataset of 34k pairs of\ncontexts and sentences in Korean covering 72 demographic groups in 15\ncategories. We find that through filtering-based moderation, social biases in\ngenerated content can be reduced by 16.47%p on average for HyperCLOVA (30B and\n82B), and GPT-3.\n", "date": "2023-05-28", "categories": []}
{"id": "2305.18504", "title": "Generalized Disparate Impact for Configurable Fairness Solutions in ML", "abstract": "  We make two contributions in the field of AI fairness over continuous\nprotected attributes. First, we show that the Hirschfeld-Gebelein-Renyi (HGR)\nindicator (the only one currently available for such a case) is valuable but\nsubject to a few crucial limitations regarding semantics, interpretability, and\nrobustness. Second, we introduce a family of indicators that are: 1)\ncomplementary to HGR in terms of semantics; 2) fully interpretable and\ntransparent; 3) robust over finite samples; 4) configurable to suit specific\napplications. Our approach also allows us to define fine-grained constraints to\npermit certain types of dependence and forbid others selectively. By expanding\nthe available options for continuous protected attributes, our approach\nrepresents a significant contribution to the area of fair artificial\nintelligence.\n", "date": "2023-05-29", "categories": []}
{"id": "2305.17557", "title": "Fair Clustering via Hierarchical Fair-Dirichlet Process", "abstract": "  The advent of ML-driven decision-making and policy formation has led to an\nincreasing focus on algorithmic fairness. As clustering is one of the most\ncommonly used unsupervised machine learning approaches, there has naturally\nbeen a proliferation of literature on {\\em fair clustering}. A popular notion\nof fairness in clustering mandates the clusters to be {\\em balanced}, i.e.,\neach level of a protected attribute must be approximately equally represented\nin each cluster. Building upon the original framework, this literature has\nrapidly expanded in various aspects. In this article, we offer a novel\nmodel-based formulation of fair clustering, complementing the existing\nliterature which is almost exclusively based on optimizing appropriate\nobjective functions.\n", "date": "2023-05-27", "categories": []}
{"id": "2305.16471", "title": "Bias, Consistency, and Partisanship in U.S. Asylum Cases: A Machine\n  Learning Analysis of Extraneous Factors in Immigration Court Decisions", "abstract": "  In this study, we introduce a novel two-pronged scoring system to measure\nindividual and systemic bias in immigration courts under the U.S. Executive\nOffice of Immigration Review (EOIR). We analyze nearly 6 million immigration\ncourt proceedings and 228 case features to build on prior research showing that\nU.S. asylum decisions vary dramatically based on factors that are extraneous to\nthe merits of a case. We close a critical gap in the literature of variability\nmetrics that can span space and time. Using predictive modeling, we explain\n58.54% of the total decision variability using two metrics: partisanship and\ninter-judge cohort consistency. Thus, whether the EOIR grants asylum to an\napplicant or not depends in majority on the combined effects of the political\nclimate and the individual variability of the presiding judge - not the\nindividual merits of the case. Using time series analysis, we also demonstrate\nthat partisanship increased in the early 1990s but plateaued following the turn\nof the century. These conclusions are striking to the extent that they diverge\nfrom the U.S. immigration system's commitments to independence and due process.\nOur contributions expose systemic inequities in the U.S. asylum decision-making\nprocess, and we recommend improved and standardized variability metrics to\nbetter diagnose and monitor these issues.\n", "date": "2023-05-25", "categories": []}
{"id": "2305.17072", "title": "Stereotypes and Smut: The (Mis)representation of Non-cisgender\n  Identities by Text-to-Image Models", "abstract": "  Cutting-edge image generation has been praised for producing high-quality\nimages, suggesting a ubiquitous future in a variety of applications. However,\ninitial studies have pointed to the potential for harm due to predictive bias,\nreflecting and potentially reinforcing cultural stereotypes. In this work, we\nare the first to investigate how multimodal models handle diverse gender\nidentities. Concretely, we conduct a thorough analysis in which we compare the\noutput of three image generation models for prompts containing cisgender vs.\nnon-cisgender identity terms. Our findings demonstrate that certain\nnon-cisgender identities are consistently (mis)represented as less human, more\nstereotyped and more sexualised. We complement our experimental analysis with\n(a)~a survey among non-cisgender individuals and (b) a series of interviews, to\nestablish which harms affected individuals anticipate, and how they would like\nto be represented. We find respondents are particularly concerned about\nmisrepresentation, and the potential to drive harmful behaviours and beliefs.\nSimple heuristics to limit offensive content are widely rejected, and instead\nrespondents call for community involvement, curated training data and the\nability to customise. These improvements could pave the way for a future where\nchange is led by the affected community, and technology is used to positively\n``[portray] queerness in ways that we haven't even thought of'' rather than\nreproducing stale, offensive stereotypes.\n", "date": "2023-05-26", "categories": []}
{"id": "2305.13938", "title": "Algorithmic Unfairness through the Lens of EU Non-Discrimination Law: Or\n  Why the Law is not a Decision Tree", "abstract": "  Concerns regarding unfairness and discrimination in the context of artificial\nintelligence (AI) systems have recently received increased attention from both\nlegal and computer science scholars. Yet, the degree of overlap between notions\nof algorithmic bias and fairness on the one hand, and legal notions of\ndiscrimination and equality on the other, is often unclear, leading to\nmisunderstandings between computer science and law. What types of bias and\nunfairness does the law address when it prohibits discrimination? What role can\nfairness metrics play in establishing legal compliance? In this paper, we aim\nto illustrate to what extent European Union (EU) non-discrimination law\ncoincides with notions of algorithmic fairness proposed in computer science\nliterature and where they differ. The contributions of this paper are as\nfollows. First, we analyse seminal examples of algorithmic unfairness through\nthe lens of EU non-discrimination law, drawing parallels with EU case law.\nSecond, we set out the normative underpinnings of fairness metrics and\ntechnical interventions and compare these to the legal reasoning of the Court\nof Justice of the EU. Specifically, we show how normative assumptions often\nremain implicit in both disciplinary approaches and explain the ensuing\nlimitations of current AI practice and non-discrimination law. We conclude with\nimplications for AI practitioners and regulators.\n", "date": "2023-05-05", "categories": []}
{"id": "2305.15979", "title": "Monitoring Algorithmic Fairness", "abstract": "  Machine-learned systems are in widespread use for making decisions about\nhumans, and it is important that they are fair, i.e., not biased against\nindividuals based on sensitive attributes. We present runtime verification of\nalgorithmic fairness for systems whose models are unknown, but are assumed to\nhave a Markov chain structure. We introduce a specification language that can\nmodel many common algorithmic fairness properties, such as demographic parity,\nequal opportunity, and social burden. We build monitors that observe a long\nsequence of events as generated by a given system, and output, after each\nobservation, a quantitative estimate of how fair or biased the system was on\nthat run until that point in time. The estimate is proven to be correct modulo\na variable error bound and a given confidence level, where the error bound gets\ntighter as the observed sequence gets longer. Our monitors are of two types,\nand use, respectively, frequentist and Bayesian statistical inference\ntechniques. While the frequentist monitors compute estimates that are\nobjectively correct with respect to the ground truth, the Bayesian monitors\ncompute estimates that are correct subject to a given prior belief about the\nsystem's model. Using a prototype implementation, we show how we can monitor if\na bank is fair in giving loans to applicants from different social backgrounds,\nand if a college is fair in admitting students while maintaining a reasonable\nfinancial burden on the society. Although they exhibit different theoretical\ncomplexities in certain cases, in our experiments, both frequentist and\nBayesian monitors took less than a millisecond to update their verdicts after\neach observation.\n", "date": "2023-05-25", "categories": []}
{"id": "2211.15897", "title": "Learning Antidote Data to Individual Unfairness", "abstract": "  Fairness is essential for machine learning systems deployed in high-stake\napplications. Among all fairness notions, individual fairness, deriving from a\nconsensus that `similar individuals should be treated similarly,' is a vital\nnotion to describe fair treatment for individual cases. Previous studies\ntypically characterize individual fairness as a prediction-invariant problem\nwhen perturbing sensitive attributes on samples, and solve it by\nDistributionally Robust Optimization (DRO) paradigm. However, such adversarial\nperturbations along a direction covering sensitive information used in DRO do\nnot consider the inherent feature correlations or innate data constraints,\ntherefore could mislead the model to optimize at off-manifold and unrealistic\nsamples. In light of this drawback, in this paper, we propose to learn and\ngenerate antidote data that approximately follows the data distribution to\nremedy individual unfairness. These generated on-manifold antidote data can be\nused through a generic optimization procedure along with original training\ndata, resulting in a pure pre-processing approach to individual unfairness, or\ncan also fit well with the in-processing DRO paradigm. Through extensive\nexperiments on multiple tabular datasets, we demonstrate our method resists\nindividual unfairness at a minimal or zero cost to predictive utility compared\nto baselines.\n", "date": "2022-11-29", "categories": []}
{"id": "2304.13680", "title": "Measuring Bias in AI Models: An Statistical Approach Introducing N-Sigma", "abstract": "  The new regulatory framework proposal on Artificial Intelligence (AI)\npublished by the European Commission establishes a new risk-based legal\napproach. The proposal highlights the need to develop adequate risk assessments\nfor the different uses of AI. This risk assessment should address, among\nothers, the detection and mitigation of bias in AI. In this work we analyze\nstatistical approaches to measure biases in automatic decision-making systems.\nWe focus our experiments in face recognition technologies. We propose a novel\nway to measure the biases in machine learning models using a statistical\napproach based on the N-Sigma method. N-Sigma is a popular statistical approach\nused to validate hypotheses in general science such as physics and social areas\nand its application to machine learning is yet unexplored. In this work we\nstudy how to apply this methodology to develop new risk assessment frameworks\nbased on bias analysis and we discuss the main advantages and drawbacks with\nrespect to other popular statistical tests.\n", "date": "2023-04-26", "categories": []}
{"id": "2305.14396", "title": "FITNESS: A Causal De-correlation Approach for Mitigating Bias in Machine\n  Learning Software", "abstract": "  Software built on top of machine learning algorithms is becoming increasingly\nprevalent in a variety of fields, including college admissions, healthcare,\ninsurance, and justice. The effectiveness and efficiency of these systems\nheavily depend on the quality of the training datasets. Biased datasets can\nlead to unfair and potentially harmful outcomes, particularly in such critical\ndecision-making systems where the allocation of resources may be affected. This\ncan exacerbate discrimination against certain groups and cause significant\nsocial disruption. To mitigate such unfairness, a series of bias-mitigating\nmethods are proposed. Generally, these studies improve the fairness of the\ntrained models to a certain degree but with the expense of sacrificing the\nmodel performance. In this paper, we propose FITNESS, a bias mitigation\napproach via de-correlating the causal effects between sensitive features\n(e.g., the sex) and the label. Our key idea is that by de-correlating such\neffects from a causality perspective, the model would avoid making predictions\nbased on sensitive features and thus fairness could be improved. Furthermore,\nFITNESS leverages multi-objective optimization to achieve a better\nperformance-fairness trade-off. To evaluate the effectiveness, we compare\nFITNESS with 7 state-of-the-art methods in 8 benchmark tasks by multiple\nmetrics. Results show that FITNESS can outperform the state-of-the-art methods\non bias mitigation while preserve the model's performance: it improved the\nmodel's fairness under all the scenarios while decreased the model's\nperformance under only 26.67% of the scenarios. Additionally, FITNESS surpasses\nthe Fairea Baseline in 96.72% cases, outperforming all methods we compared.\n", "date": "2023-05-23", "categories": []}
{"id": "2305.14468", "title": "Run Like a Girl! Sports-Related Gender Bias in Language and Vision", "abstract": "  Gender bias in Language and Vision datasets and models has the potential to\nperpetuate harmful stereotypes and discrimination. We analyze gender bias in\ntwo Language and Vision datasets. Consistent with prior work, we find that both\ndatasets underrepresent women, which promotes their invisibilization. Moreover,\nwe hypothesize and find that a bias affects human naming choices for people\nplaying sports: speakers produce names indicating the sport (e.g. 'tennis\nplayer' or 'surfer') more often when it is a man or a boy participating in the\nsport than when it is a woman or a girl, with an average of 46% vs. 35% of\nsports-related names for each gender. A computational model trained on these\nnaming data reproduces the bias. We argue that both the data and the model\nresult in representational harm against women.\n", "date": "2023-05-23", "categories": []}
{"id": "2305.13875", "title": "Fair Oversampling Technique using Heterogeneous Clusters", "abstract": "  Class imbalance and group (e.g., race, gender, and age) imbalance are\nacknowledged as two reasons in data that hinder the trade-off between fairness\nand utility of machine learning classifiers. Existing techniques have jointly\naddressed issues regarding class imbalance and group imbalance by proposing\nfair over-sampling techniques. Unlike the common oversampling techniques, which\nonly address class imbalance, fair oversampling techniques significantly\nimprove the abovementioned trade-off, as they can also address group imbalance.\nHowever, if the size of the original clusters is too small, these techniques\nmay cause classifier overfitting. To address this problem, we herein develop a\nfair oversampling technique using data from heterogeneous clusters. The\nproposed technique generates synthetic data that have class-mix features or\ngroup-mix features to make classifiers robust to overfitting. Moreover, we\ndevelop an interpolation method that can enhance the validity of generated\nsynthetic data by considering the original cluster distribution and data noise.\nFinally, we conduct experiments on five realistic datasets and three\nclassifiers, and the experimental results demonstrate the effectiveness of the\nproposed technique in terms of fairness and utility.\n", "date": "2023-05-23", "categories": []}
{"id": "2305.12434", "title": "BiasAsker: Measuring the Bias in Conversational AI System", "abstract": "  Powered by advanced Artificial Intelligence (AI) techniques, conversational\nAI systems, such as ChatGPT and digital assistants like Siri, have been widely\ndeployed in daily life. However, such systems may still produce content\ncontaining biases and stereotypes, causing potential social problems. Due to\nthe data-driven, black-box nature of modern AI techniques, comprehensively\nidentifying and measuring biases in conversational systems remains a\nchallenging task. Particularly, it is hard to generate inputs that can\ncomprehensively trigger potential bias due to the lack of data containing both\nsocial groups as well as biased properties. In addition, modern conversational\nsystems can produce diverse responses (e.g., chatting and explanation), which\nmakes existing bias detection methods simply based on the sentiment and the\ntoxicity hardly being adopted. In this paper, we propose BiasAsker, an\nautomated framework to identify and measure social bias in conversational AI\nsystems. To obtain social groups and biased properties, we construct a\ncomprehensive social bias dataset, containing a total of 841 groups and 8,110\nbiased properties. Given the dataset, BiasAsker automatically generates\nquestions and adopts a novel method based on existence measurement to identify\ntwo types of biases (i.e., absolute bias and related bias) in conversational\nsystems. Extensive experiments on 8 commercial systems and 2 famous research\nmodels, such as ChatGPT and GPT-3, show that 32.83% of the questions generated\nby BiasAsker can trigger biased behaviors in these widely deployed\nconversational systems. All the code, data, and experimental results have been\nreleased to facilitate future research.\n", "date": "2023-05-21", "categories": []}
{"id": "2305.12709", "title": "Cross-lingual Transfer Can Worsen Bias in Sentiment Analysis", "abstract": "  Sentiment analysis (SA) systems are widely deployed in many of the world's\nlanguages, and there is well-documented evidence of demographic bias in these\nsystems. In languages beyond English, scarcer training data is often\nsupplemented with transfer learning using pre-trained models, including\nmultilingual models trained on other languages. In some cases, even supervision\ndata comes from other languages. Does cross-lingual transfer also import new\nbiases? To answer this question, we use counterfactual evaluation to test\nwhether gender or racial biases are imported when using cross-lingual transfer,\ncompared to a monolingual transfer setting. Across five languages, we find that\nsystems using cross-lingual transfer usually become more biased than their\nmonolingual counterparts. We also find racial biases to be much more prevalent\nthan gender biases. To spur further research on this topic, we release the\nsentiment models we used for this study, and the intermediate checkpoints\nthroughout training, yielding 1,525 distinct models; we also release our\nevaluation code.\n", "date": "2023-05-22", "categories": []}
{"id": "2301.03421", "title": "A review of clustering models in educational data science towards\n  fairness-aware learning", "abstract": "  Ensuring fairness is essential for every education system. Machine learning\nis increasingly supporting the education system and educational data science\n(EDS) domain, from decision support to educational activities and learning\nanalytics. However, the machine learning-based decisions can be biased because\nthe algorithms may generate the results based on students' protected attributes\nsuch as race or gender. Clustering is an important machine learning technique\nto explore student data in order to support the decision-maker, as well as\nsupport educational activities, such as group assignments. Therefore, ensuring\nhigh-quality clustering models along with satisfying fairness constraints are\nimportant requirements. This chapter comprehensively surveys clustering models\nand their fairness in EDS. We especially focus on investigating the fair\nclustering models applied in educational activities. These models are believed\nto be practical tools for analyzing students' data and ensuring fairness in\nEDS.\n", "date": "2023-01-09", "categories": []}
{"id": "2305.11242", "title": "Comparing Biases and the Impact of Multilingual Training across Multiple\n  Languages", "abstract": "  Studies in bias and fairness in natural language processing have primarily\nexamined social biases within a single language and/or across few attributes\n(e.g. gender, race). However, biases can manifest differently across various\nlanguages for individual attributes. As a result, it is critical to examine\nbiases within each language and attribute. Of equal importance is to study how\nthese biases compare across languages and how the biases are affected when\ntraining a model on multilingual data versus monolingual data. We present a\nbias analysis across Italian, Chinese, English, Hebrew, and Spanish on the\ndownstream sentiment analysis task to observe whether specific demographics are\nviewed more positively. We study bias similarities and differences across these\nlanguages and investigate the impact of multilingual vs. monolingual training\ndata. We adapt existing sentiment bias templates in English to Italian,\nChinese, Hebrew, and Spanish for four attributes: race, religion, nationality,\nand gender. Our results reveal similarities in bias expression such as\nfavoritism of groups that are dominant in each language's culture (e.g.\nmajority religions and nationalities). Additionally, we find an increased\nvariation in predictions across protected groups, indicating bias\namplification, after multilingual finetuning in comparison to multilingual\npretraining.\n", "date": "2023-05-18", "categories": []}
{"id": "2305.11262", "title": "CHBias: Bias Evaluation and Mitigation of Chinese Conversational\n  Language Models", "abstract": "  \\textit{\\textbf{\\textcolor{red}{Warning}:} This paper contains content that\nmay be offensive or upsetting.} Pretrained conversational agents have been\nexposed to safety issues, exhibiting a range of stereotypical human biases such\nas gender bias. However, there are still limited bias categories in current\nresearch, and most of them only focus on English. In this paper, we introduce a\nnew Chinese dataset, CHBias, for bias evaluation and mitigation of Chinese\nconversational language models. Apart from those previous well-explored bias\ncategories, CHBias includes under-explored bias categories, such as ageism and\nappearance biases, which received less attention. We evaluate two popular\npretrained Chinese conversational models, CDial-GPT and EVA2.0, using CHBias.\nFurthermore, to mitigate different biases, we apply several debiasing methods\nto the Chinese pretrained models. Experimental results show that these Chinese\npretrained models are potentially risky for generating texts that contain\nsocial biases, and debiasing methods using the proposed dataset can make\nresponse generation less biased while preserving the models' conversational\ncapabilities.\n", "date": "2023-05-18", "categories": []}
{"id": "2305.11386", "title": "Improving Fairness in AI Models on Electronic Health Records: The Case\n  for Federated Learning Methods", "abstract": "  Developing AI tools that preserve fairness is of critical importance,\nspecifically in high-stakes applications such as those in healthcare. However,\nhealth AI models' overall prediction performance is often prioritized over the\npossible biases such models could have. In this study, we show one possible\napproach to mitigate bias concerns by having healthcare institutions\ncollaborate through a federated learning paradigm (FL; which is a popular\nchoice in healthcare settings). While FL methods with an emphasis on fairness\nhave been previously proposed, their underlying model and local implementation\ntechniques, as well as their possible applications to the healthcare domain\nremain widely underinvestigated. Therefore, we propose a comprehensive FL\napproach with adversarial debiasing and a fair aggregation method, suitable to\nvarious fairness metrics, in the healthcare domain where electronic health\nrecords are used. Not only our approach explicitly mitigates bias as part of\nthe optimization process, but an FL-based paradigm would also implicitly help\nwith addressing data imbalance and increasing the data size, offering a\npractical solution for healthcare applications. We empirically demonstrate our\nmethod's superior performance on multiple experiments simulating large-scale\nreal-world scenarios and compare it to several baselines. Our method has\nachieved promising fairness performance with the lowest impact on overall\ndiscrimination performance (accuracy).\n", "date": "2023-05-19", "categories": []}
{"id": "2305.11673", "title": "Bias Beyond English: Counterfactual Tests for Bias in Sentiment Analysis\n  in Four Languages", "abstract": "  Sentiment analysis (SA) systems are used in many products and hundreds of\nlanguages. Gender and racial biases are well-studied in English SA systems, but\nunderstudied in other languages, with few resources for such studies. To remedy\nthis, we build a counterfactual evaluation corpus for gender and racial/migrant\nbias in four languages. We demonstrate its usefulness by answering a simple but\nimportant question that an engineer might need to answer when deploying a\nsystem: What biases do systems import from pre-trained models when compared to\na baseline with no pre-training? Our evaluation corpus, by virtue of being\ncounterfactual, not only reveals which models have less bias, but also\npinpoints changes in model bias behaviour, which enables more targeted\nmitigation strategies. We release our code and evaluation corpora to facilitate\nfuture research.\n", "date": "2023-05-19", "categories": []}
{"id": "2305.11807", "title": "On the Fairness Impacts of Private Ensembles Models", "abstract": "  The Private Aggregation of Teacher Ensembles (PATE) is a machine learning\nframework that enables the creation of private models through the combination\nof multiple \"teacher\" models and a \"student\" model. The student model learns to\npredict an output based on the voting of the teachers, and the resulting model\nsatisfies differential privacy. PATE has been shown to be effective in creating\nprivate models in semi-supervised settings or when protecting data labels is a\npriority. This paper explores whether the use of PATE can result in unfairness,\nand demonstrates that it can lead to accuracy disparities among groups of\nindividuals. The paper also analyzes the algorithmic and data properties that\ncontribute to these disproportionate impacts, why these aspects are affecting\ndifferent groups disproportionately, and offers recommendations for mitigating\nthese effects\n", "date": "2023-05-19", "categories": []}
{"id": "2305.09399", "title": "Measuring Implicit Bias Using SHAP Feature Importance and Fuzzy\n  Cognitive Maps", "abstract": "  In this paper, we integrate the concepts of feature importance with implicit\nbias in the context of pattern classification. This is done by means of a\nthree-step methodology that involves (i) building a classifier and tuning its\nhyperparameters, (ii) building a Fuzzy Cognitive Map model able to quantify\nimplicit bias, and (iii) using the SHAP feature importance to active the neural\nconcepts when performing simulations. The results using a real case study\nconcerning fairness research support our two-fold hypothesis. On the one hand,\nit is illustrated the risks of using a feature importance method as an absolute\ntool to measure implicit bias. On the other hand, it is concluded that the\namount of bias towards protected features might differ depending on whether the\nfeatures are numerically or categorically encoded.\n", "date": "2023-05-16", "categories": []}
{"id": "2305.09686", "title": "Data Bias Management", "abstract": "  Due to the widespread use of data-powered systems in our everyday lives,\nconcepts like bias and fairness gained significant attention among researchers\nand practitioners, in both industry and academia. Such issues typically emerge\nfrom the data, which comes with varying levels of quality, used to train\nsupervised machine learning systems. With the commercialization and deployment\nof such systems that are sometimes delegated to make life-changing decisions,\nsignificant efforts are being made towards the identification and removal of\npossible sources of data bias that may resurface to the final end user or in\nthe decisions being made. In this paper, we present research results that show\nhow bias in data affects end users, where bias is originated, and provide a\nviewpoint about what we should do about it. We argue that data bias is not\nsomething that should necessarily be removed in all cases, and that research\nattention should instead shift from bias removal towards the identification,\nmeasurement, indexing, surfacing, and adapting for bias, which we name bias\nmanagement.\n", "date": "2023-05-15", "categories": []}
{"id": "2305.09931", "title": "Mitigating Group Bias in Federated Learning: Beyond Local Fairness", "abstract": "  The issue of group fairness in machine learning models, where certain\nsub-populations or groups are favored over others, has been recognized for some\ntime. While many mitigation strategies have been proposed in centralized\nlearning, many of these methods are not directly applicable in federated\nlearning, where data is privately stored on multiple clients. To address this,\nmany proposals try to mitigate bias at the level of clients before aggregation,\nwhich we call locally fair training. However, the effectiveness of these\napproaches is not well understood. In this work, we investigate the theoretical\nfoundation of locally fair training by studying the relationship between global\nmodel fairness and local model fairness. Additionally, we prove that for a\nbroad class of fairness metrics, the global model's fairness can be obtained\nusing only summary statistics from local clients. Based on that, we propose a\nglobally fair training algorithm that directly minimizes the penalized\nempirical loss. Real-data experiments demonstrate the promising performance of\nour proposed approach for enhancing fairness while retaining high accuracy\ncompared to locally fair training methods.\n", "date": "2023-05-17", "categories": []}
{"id": "2305.10204", "title": "Shielded Representations: Protecting Sensitive Attributes Through\n  Iterative Gradient-Based Projection", "abstract": "  Natural language processing models tend to learn and encode social biases\npresent in the data. One popular approach for addressing such biases is to\neliminate encoded information from the model's representations. However,\ncurrent methods are restricted to removing only linearly encoded information.\nIn this work, we propose Iterative Gradient-Based Projection (IGBP), a novel\nmethod for removing non-linear encoded concepts from neural representations.\nOur method consists of iteratively training neural classifiers to predict a\nparticular attribute we seek to eliminate, followed by a projection of the\nrepresentation on a hypersurface, such that the classifiers become oblivious to\nthe target attribute. We evaluate the effectiveness of our method on the task\nof removing gender and race information as sensitive attributes. Our results\ndemonstrate that IGBP is effective in mitigating bias through intrinsic and\nextrinsic evaluations, with minimal impact on downstream task accuracy.\n", "date": "2023-05-17", "categories": []}
{"id": "2209.05274", "title": "Fairness in Forecasting of Observations of Linear Dynamical Systems", "abstract": "  In machine learning, training data often capture the behaviour of multiple\nsubgroups of some underlying human population. This behaviour can often be\nmodelled as observations of an unknown dynamical system with an unobserved\nstate. When the training data for the subgroups are not controlled carefully,\nhowever, under-representation bias arises. To counter under-representation\nbias, we introduce two natural notions of fairness in time-series forecasting\nproblems: subgroup fairness and instantaneous fairness. These notions extend\npredictive parity to the learning of dynamical systems. We also show globally\nconvergent methods for the fairness-constrained learning problems using\nhierarchies of convexifications of non-commutative polynomial optimisation\nproblems. We also show that by exploiting sparsity in the convexifications, we\ncan reduce the run time of our methods considerably. Our empirical results on a\nbiased data set motivated by insurance applications and the well-known COMPAS\ndata set demonstrate the efficacy of our methods.\n", "date": "2022-09-12", "categories": []}
{"id": "2212.11261", "title": "Contrastive Language-Vision AI Models Pretrained on Web-Scraped\n  Multimodal Data Exhibit Sexual Objectification Bias", "abstract": "  Nine language-vision AI models trained on web scrapes with the Contrastive\nLanguage-Image Pretraining (CLIP) objective are evaluated for evidence of a\nbias studied by psychologists: the sexual objectification of girls and women,\nwhich occurs when a person's human characteristics, such as emotions, are\ndisregarded and the person is treated as a body. We replicate three experiments\nin psychology quantifying sexual objectification and show that the phenomena\npersist in AI. A first experiment uses standardized images of women from the\nSexual OBjectification and EMotion Database, and finds that human\ncharacteristics are disassociated from images of objectified women: the model's\nrecognition of emotional state is mediated by whether the subject is fully or\npartially clothed. Embedding association tests (EATs) return significant effect\nsizes for both anger (d >0.80) and sadness (d >0.50), associating images of\nfully clothed subjects with emotions. GRAD-CAM saliency maps highlight that\nCLIP gets distracted from emotional expressions in objectified images. A second\nexperiment measures the effect in a representative application: an automatic\nimage captioner (Antarctic Captions) includes words denoting emotion less than\n50% as often for images of partially clothed women than for images of fully\nclothed women. A third experiment finds that images of female professionals\n(scientists, doctors, executives) are likely to be associated with sexual\ndescriptions relative to images of male professionals. A fourth experiment\nshows that a prompt of \"a [age] year old girl\" generates sexualized images (as\ndetermined by an NSFW classifier) up to 73% of the time for VQGAN-CLIP and\nStable Diffusion; the corresponding rate for boys never surpasses 9%. The\nevidence indicates that language-vision AI models trained on web scrapes learn\nbiases of sexual objectification, which propagate to downstream applications.\n", "date": "2022-12-21", "categories": []}
{"id": "2305.09072", "title": "Skin Deep: Investigating Subjectivity in Skin Tone Annotations for\n  Computer Vision Benchmark Datasets", "abstract": "  To investigate the well-observed racial disparities in computer vision\nsystems that analyze images of humans, researchers have turned to skin tone as\nmore objective annotation than race metadata for fairness performance\nevaluations. However, the current state of skin tone annotation procedures is\nhighly varied. For instance, researchers use a range of untested scales and\nskin tone categories, have unclear annotation procedures, and provide\ninadequate analyses of uncertainty. In addition, little attention is paid to\nthe positionality of the humans involved in the annotation process--both\ndesigners and annotators alike--and the historical and sociological context of\nskin tone in the United States. Our work is the first to investigate the skin\ntone annotation process as a sociotechnical project. We surveyed recent skin\ntone annotation procedures and conducted annotation experiments to examine how\nsubjective understandings of skin tone are embedded in skin tone annotation\nprocedures. Our systematic literature review revealed the uninterrogated\nassociation between skin tone and race and the limited effort to analyze\nannotator uncertainty in current procedures for skin tone annotation in\ncomputer vision evaluation. Our experiments demonstrated that design decisions\nin the annotation procedure such as the order in which the skin tone scale is\npresented or additional context in the image (i.e., presence of a face)\nsignificantly affected the resulting inter-annotator agreement and individual\nuncertainty of skin tone annotations. We call for greater reflexivity in the\ndesign, analysis, and documentation of procedures for evaluation using skin\ntone.\n", "date": "2023-05-15", "categories": []}
{"id": "2302.04118", "title": "On the Richness of Calibration", "abstract": "  Probabilistic predictions can be evaluated through comparisons with observed\nlabel frequencies, that is, through the lens of calibration. Recent scholarship\non algorithmic fairness has started to look at a growing variety of\ncalibration-based objectives under the name of multi-calibration but has still\nremained fairly restricted. In this paper, we explore and analyse forms of\nevaluation through calibration by making explicit the choices involved in\ndesigning calibration scores. We organise these into three grouping choices and\na choice concerning the agglomeration of group errors. This provides a\nframework for comparing previously proposed calibration scores and helps to\nformulate novel ones with desirable mathematical properties. In particular, we\nexplore the possibility of grouping datapoints based on their input features\nrather than on predictions and formally demonstrate advantages of such\napproaches. We also characterise the space of suitable agglomeration functions\nfor group errors, generalising previously proposed calibration scores.\nComplementary to such population-level scores, we explore calibration scores at\nthe individual level and analyse their relationship to choices of grouping. We\ndraw on these insights to introduce and axiomatise fairness deviation measures\nfor population-level scores. We demonstrate that with appropriate choices of\ngrouping, these novel global fairness scores can provide notions of (sub-)group\nor individual fairness.\n", "date": "2023-02-08", "categories": []}
{"id": "2207.13319", "title": "Should Bank Stress Tests Be Fair?", "abstract": "  Regulatory stress tests have become one of the main tools for setting capital\nrequirements at the largest U.S. banks. The Federal Reserve uses confidential\nmodels to evaluate bank-specific outcomes for bank-specific portfolios in\nshared stress scenarios. As a matter of policy, the same models are used for\nall banks, despite considerable heterogeneity across institutions; individual\nbanks have contended that some models are not suited to their businesses.\nMotivated by this debate, we ask, what is a fair aggregation of individually\ntailored models into a common model? We argue that simply pooling data across\nbanks treats banks equally but is subject to two deficiencies: it may distort\nthe impact of legitimate portfolio features, and it is vulnerable to implicit\nmisdirection of legitimate information to infer bank identity. We compare\nvarious notions of regression fairness to address these deficiencies,\nconsidering both forecast accuracy and equal treatment. In the setting of\nlinear models, we argue for estimating and then discarding centered bank fixed\neffects as preferable to simply ignoring differences across banks. We present\nevidence that the overall impact can be material. We also discuss extensions to\nnonlinear models.\n", "date": "2022-07-27", "categories": []}
{"id": "2305.06402", "title": "Analyzing Bias in Diffusion-based Face Generation Models", "abstract": "  Diffusion models are becoming increasingly popular in synthetic data\ngeneration and image editing applications. However, these models can amplify\nexisting biases and propagate them to downstream applications. Therefore, it is\ncrucial to understand the sources of bias in their outputs. In this paper, we\ninvestigate the presence of bias in diffusion-based face generation models with\nrespect to attributes such as gender, race, and age. Moreover, we examine how\ndataset size affects the attribute composition and perceptual quality of both\ndiffusion and Generative Adversarial Network (GAN) based face generation models\nacross various attribute classes. Our findings suggest that diffusion models\ntend to worsen distribution bias in the training data for various attributes,\nwhich is heavily influenced by the size of the dataset. Conversely, GAN models\ntrained on balanced datasets with a larger number of samples show less bias\nacross different attributes.\n", "date": "2023-05-10", "categories": []}
{"id": "2305.06994", "title": "A statistical approach to detect sensitive features in a group fairness\n  setting", "abstract": "  The use of machine learning models in decision support systems with high\nsocietal impact raised concerns about unfair (disparate) results for different\ngroups of people. When evaluating such unfair decisions, one generally relies\non predefined groups that are determined by a set of features that are\nconsidered sensitive. However, such an approach is subjective and does not\nguarantee that these features are the only ones to be considered as sensitive\nnor that they entail unfair (disparate) outcomes.\n  In this paper, we propose a preprocessing step to address the task of\nautomatically recognizing sensitive features that does not require a trained\nmodel to verify unfair results. Our proposal is based on the Hilber-Schmidt\nindependence criterion, which measures the statistical dependence of variable\ndistributions. We hypothesize that if the dependence between the label vector\nand a candidate is high for a sensitive feature, then the information provided\nby this feature will entail disparate performance measures between groups. Our\nempirical results attest our hypothesis and show that several features\nconsidered as sensitive in the literature do not necessarily entail disparate\n(unfair) results.\n", "date": "2023-05-11", "categories": []}
{"id": "2305.05827", "title": "Inclusive FinTech Lending via Contrastive Learning and Domain Adaptation", "abstract": "  FinTech lending (e.g., micro-lending) has played a significant role in\nfacilitating financial inclusion. It has reduced processing times and costs,\nenhanced the user experience, and made it possible for people to obtain loans\nwho may not have qualified for credit from traditional lenders. However, there\nare concerns about the potentially biased algorithmic decision-making during\nloan screening. Machine learning algorithms used to evaluate credit quality can\nbe influenced by representation bias in the training data, as we only have\naccess to the default outcome labels of approved loan applications, for which\nthe borrowers' socioeconomic characteristics are better than those of rejected\nones. In this case, the model trained on the labeled data performs well on the\nhistorically approved population, but does not generalize well to borrowers of\nlow socioeconomic background. In this paper, we investigate the problem of\nrepresentation bias in loan screening for a real-world FinTech lending\nplatform. We propose a new Transformer-based sequential loan screening model\nwith self-supervised contrastive learning and domain adaptation to tackle this\nchallenging issue. We use contrastive learning to train our feature extractor\non unapproved (unlabeled) loan applications and use domain adaptation to\ngeneralize the performance of our label predictor. We demonstrate the\neffectiveness of our model through extensive experimentation in the real-world\nmicro-lending setting. Our results show that our model significantly promotes\nthe inclusiveness of funding decisions, while also improving loan screening\naccuracy and profit by 7.10% and 8.95%, respectively. We also show that\nincorporating the test data into contrastive learning and domain adaptation and\nlabeling a small ratio of test data can further boost model performance.\n", "date": "2023-05-10", "categories": []}
{"id": "2305.05101", "title": "Towards unraveling calibration biases in medical image analysis", "abstract": "  In recent years the development of artificial intelligence (AI) systems for\nautomated medical image analysis has gained enormous momentum. At the same\ntime, a large body of work has shown that AI systems can systematically and\nunfairly discriminate against certain populations in various application\nscenarios. These two facts have motivated the emergence of algorithmic fairness\nstudies in this field. Most research on healthcare algorithmic fairness to date\nhas focused on the assessment of biases in terms of classical discrimination\nmetrics such as AUC and accuracy. Potential biases in terms of model\ncalibration, however, have only recently begun to be evaluated. This is\nespecially important when working with clinical decision support systems, as\npredictive uncertainty is key for health professionals to optimally evaluate\nand combine multiple sources of information. In this work we study\ndiscrimination and calibration biases in models trained for automatic detection\nof malignant dermatological conditions from skin lesions images. Importantly,\nwe show how several typically employed calibration metrics are systematically\nbiased with respect to sample sizes, and how this can lead to erroneous\nfairness analysis if not taken into consideration. This is of particular\nrelevance to fairness studies, where data imbalance results in drastic sample\nsize differences between demographic sub-groups, which, if not taken into\naccount, can act as confounders.\n", "date": "2023-05-09", "categories": []}
{"id": "2305.04699", "title": "Runtime Monitoring of Dynamic Fairness Properties", "abstract": "  A machine-learned system that is fair in static decision-making tasks may\nhave biased societal impacts in the long-run. This may happen when the system\ninteracts with humans and feedback patterns emerge, reinforcing old biases in\nthe system and creating new biases. While existing works try to identify and\nmitigate long-run biases through smart system design, we introduce techniques\nfor monitoring fairness in real time. Our goal is to build and deploy a monitor\nthat will continuously observe a long sequence of events generated by the\nsystem in the wild, and will output, with each event, a verdict on how fair the\nsystem is at the current point in time. The advantages of monitoring are\ntwo-fold. Firstly, fairness is evaluated at run-time, which is important\nbecause unfair behaviors may not be eliminated a priori, at design-time, due to\npartial knowledge about the system and the environment, as well as\nuncertainties and dynamic changes in the system and the environment, such as\nthe unpredictability of human behavior. Secondly, monitors are by design\noblivious to how the monitored system is constructed, which makes them suitable\nto be used as trusted third-party fairness watchdogs. They function as\ncomputationally lightweight statistical estimators, and their correctness\nproofs rely on the rigorous analysis of the stochastic process that models the\nassumptions about the underlying dynamics of the system. We show, both in\ntheory and experiments, how monitors can warn us (1) if a bank's credit policy\nover time has created an unfair distribution of credit scores among the\npopulation, and (2) if a resource allocator's allocation policy over time has\nmade unfair allocations. Our experiments demonstrate that the monitors\nintroduce very low overhead. We believe that runtime monitoring is an important\nand mathematically rigorous new addition to the fairness toolbox.\n", "date": "2023-05-08", "categories": []}
{"id": "2305.03691", "title": "Mining bias-target Alignment from Voronoi Cells", "abstract": "  Despite significant research efforts, deep neural networks are still\nvulnerable to biases: this raises concerns about their fairness and limits\ntheir generalization. In this paper, we propose a bias-agnostic approach to\nmitigate the impact of bias in deep neural networks. Unlike traditional\ndebiasing approaches, we rely on a metric to quantify ``bias\nalignment/misalignment'' on target classes, and use this information to\ndiscourage the propagation of bias-target alignment information through the\nnetwork. We conduct experiments on several commonly used datasets for debiasing\nand compare our method to supervised and bias-specific approaches. Our results\nindicate that the proposed method achieves comparable performance to\nstate-of-the-art supervised approaches, although it is bias-agnostic, even in\npresence of multiple biases in the same sample.\n", "date": "2023-05-05", "categories": []}
{"id": "2305.02806", "title": "Maximizing Submodular Functions for Recommendation in the Presence of\n  Biases", "abstract": "  Subset selection tasks, arise in recommendation systems and search engines\nand ask to select a subset of items that maximize the value for the user. The\nvalues of subsets often display diminishing returns, and hence, submodular\nfunctions have been used to model them. If the inputs defining the submodular\nfunction are known, then existing algorithms can be used. In many applications,\nhowever, inputs have been observed to have social biases that reduce the\nutility of the output subset. Hence, interventions to improve the utility are\ndesired. Prior works focus on maximizing linear functions -- a special case of\nsubmodular functions -- and show that fairness constraint-based interventions\ncan not only ensure proportional representation but also achieve near-optimal\nutility in the presence of biases. We study the maximization of a family of\nsubmodular functions that capture functions arising in the aforementioned\napplications. Our first result is that, unlike linear functions,\nconstraint-based interventions cannot guarantee any constant fraction of the\noptimal utility for this family of submodular functions. Our second result is\nan algorithm for submodular maximization. The algorithm provably outputs\nsubsets that have near-optimal utility for this family under mild assumptions\nand that proportionally represent items from each group. In empirical\nevaluation, with both synthetic and real-world data, we observe that this\nalgorithm improves the utility of the output subset for this family of\nsubmodular functions over baselines.\n", "date": "2023-05-03", "categories": []}
{"id": "2305.01783", "title": "Fairness and representation in satellite-based poverty maps: Evidence of\n  urban-rural disparities and their impacts on downstream policy", "abstract": "  Poverty maps derived from satellite imagery are increasingly used to inform\nhigh-stakes policy decisions, such as the allocation of humanitarian aid and\nthe distribution of government resources. Such poverty maps are typically\nconstructed by training machine learning algorithms on a relatively modest\namount of ``ground truth\" data from surveys, and then predicting poverty levels\nin areas where imagery exists but surveys do not. Using survey and satellite\ndata from ten countries, this paper investigates disparities in representation,\nsystematic biases in prediction errors, and fairness concerns in\nsatellite-based poverty mapping across urban and rural lines, and shows how\nthese phenomena affect the validity of policies based on predicted maps. Our\nfindings highlight the importance of careful error and bias analysis before\nusing satellite-based poverty maps in real-world policy decisions.\n", "date": "2023-05-02", "categories": []}
{"id": "2305.01888", "title": "Fairness in AI Systems: Mitigating gender bias from language-vision\n  models", "abstract": "  Our society is plagued by several biases, including racial biases, caste\nbiases, and gender bias. As a matter of fact, several years ago, most of these\nnotions were unheard of. These biases passed through generations along with\namplification have lead to scenarios where these have taken the role of\nexpected norms by certain groups in the society. One notable example is of\ngender bias. Whether we talk about the political world, lifestyle or corporate\nworld, some generic differences are observed regarding the involvement of both\nthe groups. This differential distribution, being a part of the society at\nlarge, exhibits its presence in the recorded data as well. Machine learning is\nalmost entirely dependent on the availability of data; and the idea of learning\nfrom data and making predictions assumes that data defines the expected\nbehavior at large. Hence, with biased data the resulting models are corrupted\nwith those inherent biases too; and with the current popularity of ML in\nproducts, this can result in a huge obstacle in the path of equality and\njustice. This work studies and attempts to alleviate gender bias issues from\nlanguage vision models particularly the task of image captioning. We study the\nextent of the impact of gender bias in existing datasets and propose a\nmethodology to mitigate its impact in caption based language vision models.\n", "date": "2023-05-03", "categories": []}
{"id": "2305.02009", "title": "fairml: A Statistician's Take on Fair Machine Learning Modelling", "abstract": "  The adoption of machine learning in applications where it is crucial to\nensure fairness and accountability has led to a large number of model proposals\nin the literature, largely formulated as optimisation problems with constraints\nreducing or eliminating the effect of sensitive attributes on the response.\nWhile this approach is very flexible from a theoretical perspective, the\nresulting models are somewhat black-box in nature: very little can be said\nabout their statistical properties, what are the best practices in their\napplied use, and how they can be extended to problems other than those they\nwere originally designed for. Furthermore, the estimation of each model\nrequires a bespoke implementation involving an appropriate solver which is less\nthan desirable from a software engineering perspective.\n  In this paper, we describe the fairml R package which implements our previous\nwork (Scutari, Panero, and Proissl 2022) and related models in the literature.\nfairml is designed around classical statistical models (generalised linear\nmodels) and penalised regression results (ridge regression) to produce fair\nmodels that are interpretable and whose properties are well-known. The\nconstraint used to enforce fairness is orthogonal to model estimation, making\nit possible to mix-and-match the desired model family and fairness definition\nfor each application. Furthermore, fairml provides facilities for model\nestimation, model selection and validation including diagnostic plots.\n", "date": "2023-05-03", "categories": []}
{"id": "2206.02891", "title": "A Justice-Based Framework for the Analysis of Algorithmic\n  Fairness-Utility Trade-Offs", "abstract": "  In prediction-based decision-making systems, different perspectives can be at\nodds: The short-term business goals of the decision makers are often in\nconflict with the decision subjects' wish to be treated fairly. Balancing these\ntwo perspectives is a question of values. However, these values are often\nhidden in the technicalities of the implementation of the decision-making\nsystem. In this paper, we propose a framework to make these value-laden choices\nclearly visible. We focus on a setting in which we want to find decision rules\nthat balance the perspective of the decision maker and of the decision\nsubjects. We provide an approach to formalize both perspectives, i.e., to\nassess the utility of the decision maker and the fairness towards the decision\nsubjects. In both cases, the idea is to elicit values from decision makers and\ndecision subjects that are then turned into something measurable. For the\nfairness evaluation, we build on well-known theories of distributive justice\nand on the algorithmic literature to ask what a fair distribution of utility\n(or welfare) looks like. This allows us to derive a fairness score that we then\ncompare to the decision maker's utility. As we focus on a setting in which we\nare given a trained model and have to choose a decision rule, we use the\nconcept of Pareto efficiency to compare decision rules. Our proposed framework\ncan both guide the implementation of a decision-making system and help with\naudits, as it allows us to resurface the values implemented in a\ndecision-making system.\n", "date": "2022-06-06", "categories": []}
{"id": "2206.02897", "title": "Distributive Justice as the Foundational Premise of Fair ML:\n  Unification, Extension, and Interpretation of Group Fairness Metrics", "abstract": "  Group fairness metrics are an established way of assessing the fairness of\nprediction-based decision-making systems. However, these metrics are still\ninsufficiently linked to philosophical theories, and their moral meaning is\noften unclear. In this paper, we propose a comprehensive framework for group\nfairness metrics, which links them to more theories of distributive justice.\nThe different group fairness metrics differ in their choices about how to\nmeasure the benefit or harm of a decision for the affected individuals, and\nwhat moral claims to benefits are assumed. Our unifying framework reveals the\nnormative choices associated with standard group fairness metrics and allows an\ninterpretation of their moral substance. In addition, this broader view\nprovides a structure for the expansion of standard fairness metrics that we\nfind in the literature. This expansion allows addressing several criticisms of\nstandard group fairness metrics, specifically: (1) they are parity-based, i.e.,\nthey demand some form of equality between groups, which may sometimes be\ndetrimental to marginalized groups; (2) they only compare decisions across\ngroups but not the resulting consequences for these groups; and (3) the full\nbreadth of the distributive justice literature is not sufficiently represented.\n", "date": "2022-06-06", "categories": []}
{"id": "2305.01595", "title": "On the Impact of Data Quality on Image Classification Fairness", "abstract": "  With the proliferation of algorithmic decision-making, increased scrutiny has\nbeen placed on these systems. This paper explores the relationship between the\nquality of the training data and the overall fairness of the models trained\nwith such data in the context of supervised classification. We measure key\nfairness metrics across a range of algorithms over multiple image\nclassification datasets that have a varying level of noise in both the labels\nand the training data itself. We describe noise in the labels as inaccuracies\nin the labelling of the data in the training set and noise in the data as\ndistortions in the data, also in the training set. By adding noise to the\noriginal datasets, we can explore the relationship between the quality of the\ntraining data and the fairness of the output of the models trained on that\ndata.\n", "date": "2023-05-02", "categories": []}
{"id": "2206.10843", "title": "Learning Debiased Classifier with Biased Committee", "abstract": "  Neural networks are prone to be biased towards spurious correlations between\nclasses and latent attributes exhibited in a major portion of training data,\nwhich ruins their generalization capability. We propose a new method for\ntraining debiased classifiers with no spurious attribute label. The key idea is\nto employ a committee of classifiers as an auxiliary module that identifies\nbias-conflicting data, i.e., data without spurious correlation, and assigns\nlarge weights to them when training the main classifier. The committee is\nlearned as a bootstrapped ensemble so that a majority of its classifiers are\nbiased as well as being diverse, and intentionally fail to predict classes of\nbias-conflicting data accordingly. The consensus within the committee on\nprediction difficulty thus provides a reliable cue for identifying and\nweighting bias-conflicting data. Moreover, the committee is also trained with\nknowledge transferred from the main classifier so that it gradually becomes\ndebiased along with the main classifier and emphasizes more difficult data as\ntraining progresses. On five real-world datasets, our method outperforms prior\narts using no spurious attribute label like ours and even surpasses those\nrelying on bias labels occasionally.\n", "date": "2022-06-22", "categories": []}
{"id": "2305.00817", "title": "Racial Bias within Face Recognition: A Survey", "abstract": "  Facial recognition is one of the most academically studied and industrially\ndeveloped areas within computer vision where we readily find associated\napplications deployed globally. This widespread adoption has uncovered\nsignificant performance variation across subjects of different racial profiles\nleading to focused research attention on racial bias within face recognition\nspanning both current causation and future potential solutions. In support,\nthis study provides an extensive taxonomic review of research on racial bias\nwithin face recognition exploring every aspect and stage of the face\nrecognition processing pipeline. Firstly, we discuss the problem definition of\nracial bias, starting with race definition, grouping strategies, and the\nsocietal implications of using race or race-related groupings. Secondly, we\ndivide the common face recognition processing pipeline into four stages: image\nacquisition, face localisation, face representation, face verification and\nidentification, and review the relevant corresponding literature associated\nwith each stage. The overall aim is to provide comprehensive coverage of the\nracial bias problem with respect to each and every stage of the face\nrecognition processing pipeline whilst also highlighting the potential pitfalls\nand limitations of contemporary mitigation strategies that need to be\nconsidered within future research endeavours or commercial applications alike.\n", "date": "2023-05-01", "categories": []}
{"id": "2305.00927", "title": "Cross-Institutional Transfer Learning for Educational Models:\n  Implications for Model Performance, Fairness, and Equity", "abstract": "  Modern machine learning increasingly supports paradigms that are\nmulti-institutional (using data from multiple institutions during training) or\ncross-institutional (using models from multiple institutions for inference),\nbut the empirical effects of these paradigms are not well understood. This\nstudy investigates cross-institutional learning via an empirical case study in\nhigher education. We propose a framework and metrics for assessing the utility\nand fairness of student dropout prediction models that are transferred across\ninstitutions. We examine the feasibility of cross-institutional transfer under\nreal-world data- and model-sharing constraints, quantifying model biases for\nintersectional student identities, characterizing potential disparate impact\ndue to these biases, and investigating the impact of various\ncross-institutional ensembling approaches on fairness and overall model\nperformance. We perform this analysis on data representing over 200,000\nenrolled students annually from four universities without sharing training data\nbetween institutions.\n  We find that a simple zero-shot cross-institutional transfer procedure can\nachieve similar performance to locally-trained models for all institutions in\nour study, without sacrificing model fairness. We also find that stacked\nensembling provides no additional benefits to overall performance or fairness\ncompared to either a local model or the zero-shot transfer procedure we tested.\nWe find no evidence of a fairness-accuracy tradeoff across dozens of models and\ntransfer schemes evaluated. Our auditing procedure also highlights the\nimportance of intersectional fairness analysis, revealing performance\ndisparities at the intersection of sensitive identity groups that are concealed\nunder one-dimensional analysis.\n", "date": "2023-05-01", "categories": []}
{"id": "2209.15605", "title": "Bias Mimicking: A Simple Sampling Approach for Bias Mitigation", "abstract": "  Prior work has shown that Visual Recognition datasets frequently\nunderrepresent bias groups $B$ (\\eg Female) within class labels $Y$ (\\eg\nProgrammers). This dataset bias can lead to models that learn spurious\ncorrelations between class labels and bias groups such as age, gender, or race.\nMost recent methods that address this problem require significant architectural\nchanges or additional loss functions requiring more hyper-parameter tuning.\nAlternatively, data sampling baselines from the class imbalance literature (\\eg\nUndersampling, Upweighting), which can often be implemented in a single line of\ncode and often have no hyperparameters, offer a cheaper and more efficient\nsolution. However, these methods suffer from significant shortcomings. For\nexample, Undersampling drops a significant part of the input distribution per\nepoch while Oversampling repeats samples, causing overfitting. To address these\nshortcomings, we introduce a new class-conditioned sampling method: Bias\nMimicking. The method is based on the observation that if a class $c$ bias\ndistribution, \\ie $P_D(B|Y=c)$ is mimicked across every $c^{\\prime}\\neq c$,\nthen $Y$ and $B$ are statistically independent. Using this notion, BM, through\na novel training procedure, ensures that the model is exposed to the entire\ndistribution per epoch without repeating samples. Consequently, Bias Mimicking\nimproves underrepresented groups' accuracy of sampling methods by 3\\% over four\nbenchmarks while maintaining and sometimes improving performance over\nnonsampling methods. Code: \\url{https://github.com/mqraitem/Bias-Mimicking}\n", "date": "2022-09-30", "categories": []}
{"id": "2304.13950", "title": "Fairness Uncertainty Quantification: How certain are you that the model\n  is fair?", "abstract": "  Fairness-aware machine learning has garnered significant attention in recent\nyears because of extensive use of machine learning in sensitive applications\nlike judiciary systems. Various heuristics, and optimization frameworks have\nbeen proposed to enforce fairness in classification \\cite{del2020review} where\nthe later approaches either provides empirical results or provides fairness\nguarantee for the exact minimizer of the objective function\n\\cite{celis2019classification}. In modern machine learning, Stochastic Gradient\nDescent (SGD) type algorithms are almost always used as training algorithms\nimplying that the learned model, and consequently, its fairness properties are\nrandom. Hence, especially for crucial applications, it is imperative to\nconstruct Confidence Interval (CI) for the fairness of the learned model. In\nthis work we provide CI for test unfairness when a group-fairness-aware,\nspecifically, Disparate Impact (DI), and Disparate Mistreatment (DM) aware\nlinear binary classifier is trained using online SGD-type algorithms. We show\nthat asymptotically a Central Limit Theorem holds for the estimated model\nparameter of both DI and DM-aware models. We provide online multiplier\nbootstrap method to estimate the asymptotic covariance to construct online CI.\nTo do so, we extend the known theoretical guarantees shown on the consistency\nof the online bootstrap method for unconstrained SGD to constrained\noptimization which could be of independent interest. We illustrate our results\non synthetic and real datasets.\n", "date": "2023-04-27", "categories": []}
{"id": "2301.02053", "title": "Max-Min Diversification with Fairness Constraints: Exact and\n  Approximation Algorithms", "abstract": "  Diversity maximization aims to select a diverse and representative subset of\nitems from a large dataset. It is a fundamental optimization task that finds\napplications in data summarization, feature selection, web search, recommender\nsystems, and elsewhere. However, in a setting where data items are associated\nwith different groups according to sensitive attributes like sex or race, it is\npossible that algorithmic solutions for this task, if left unchecked, will\nunder- or over-represent some of the groups. Therefore, we are motivated to\naddress the problem of \\emph{max-min diversification with fairness\nconstraints}, aiming to select $k$ items to maximize the minimum distance\nbetween any pair of selected items while ensuring that the number of items\nselected from each group falls within predefined lower and upper bounds. In\nthis work, we propose an exact algorithm based on integer linear programming\nthat is suitable for small datasets as well as a\n$\\frac{1-\\varepsilon}{5}$-approximation algorithm for any $\\varepsilon \\in (0,\n1)$ that scales to large datasets. Extensive experiments on real-world datasets\ndemonstrate the superior performance of our proposed algorithms over existing\nones.\n", "date": "2023-01-05", "categories": []}
{"id": "2304.13493", "title": "Towards clinical AI fairness: A translational perspective", "abstract": "  Artificial intelligence (AI) has demonstrated the ability to extract insights\nfrom data, but the issue of fairness remains a concern in high-stakes fields\nsuch as healthcare. Despite extensive discussion and efforts in algorithm\ndevelopment, AI fairness and clinical concerns have not been adequately\naddressed. In this paper, we discuss the misalignment between technical and\nclinical perspectives of AI fairness, highlight the barriers to AI fairness'\ntranslation to healthcare, advocate multidisciplinary collaboration to bridge\nthe knowledge gap, and provide possible solutions to address the clinical\nconcerns pertaining to AI fairness.\n", "date": "2023-04-26", "categories": []}
{"id": "2211.07907", "title": "MMD-B-Fair: Learning Fair Representations with Statistical Testing", "abstract": "  We introduce a method, MMD-B-Fair, to learn fair representations of data via\nkernel two-sample testing. We find neural features of our data where a maximum\nmean discrepancy (MMD) test cannot distinguish between representations of\ndifferent sensitive groups, while preserving information about the target\nattributes. Minimizing the power of an MMD test is more difficult than\nmaximizing it (as done in previous work), because the test threshold's complex\nbehavior cannot be simply ignored. Our method exploits the simple asymptotics\nof block testing schemes to efficiently find fair representations without\nrequiring complex adversarial optimization or generative modelling schemes\nwidely used by existing work on fair representation learning. We evaluate our\napproach on various datasets, showing its ability to ``hide'' information about\nsensitive attributes, and its effectiveness in downstream transfer tasks.\n", "date": "2022-11-15", "categories": []}
{"id": "2304.12573", "title": "Fairness and Bias in Truth Discovery Algorithms: An Experimental\n  Analysis", "abstract": "  Machine learning (ML) based approaches are increasingly being used in a\nnumber of applications with societal impact. Training ML models often require\nvast amounts of labeled data, and crowdsourcing is a dominant paradigm for\nobtaining labels from multiple workers. Crowd workers may sometimes provide\nunreliable labels, and to address this, truth discovery (TD) algorithms such as\nmajority voting are applied to determine the consensus labels from conflicting\nworker responses. However, it is important to note that these consensus labels\nmay still be biased based on sensitive attributes such as gender, race, or\npolitical affiliation. Even when sensitive attributes are not involved, the\nlabels can be biased due to different perspectives of subjective aspects such\nas toxicity. In this paper, we conduct a systematic study of the bias and\nfairness of TD algorithms. Our findings using two existing crowd-labeled\ndatasets, reveal that a non-trivial proportion of workers provide biased\nresults, and using simple approaches for TD is sub-optimal. Our study also\ndemonstrates that popular TD algorithms are not a panacea. Additionally, we\nquantify the impact of these unfair workers on downstream ML tasks and show\nthat conventional methods for achieving fairness and correcting label biases\nare ineffective in this setting. We end the paper with a plea for the design of\nnovel bias-aware truth discovery algorithms that can ameliorate these issues.\n", "date": "2023-04-25", "categories": []}
{"id": "2304.11217", "title": "ACROCPoLis: A Descriptive Framework for Making Sense of Fairness", "abstract": "  Fairness is central to the ethical and responsible development and use of AI\nsystems, with a large number of frameworks and formal notions of algorithmic\nfairness being available. However, many of the fairness solutions proposed\nrevolve around technical considerations and not the needs of and consequences\nfor the most impacted communities. We therefore want to take the focus away\nfrom definitions and allow for the inclusion of societal and relational aspects\nto represent how the effects of AI systems impact and are experienced by\nindividuals and social groups. In this paper, we do this by means of proposing\nthe ACROCPoLis framework to represent allocation processes with a modeling\nemphasis on fairness aspects. The framework provides a shared vocabulary in\nwhich the factors relevant to fairness assessments for different situations and\nprocedures are made explicit, as well as their interrelationships. This enables\nus to compare analogous situations, to highlight the differences in dissimilar\nsituations, and to capture differing interpretations of the same situation by\ndifferent stakeholders.\n", "date": "2023-04-19", "categories": []}
{"id": "2304.12190", "title": "Optimizing fairness tradeoffs in machine learning with multiobjective\n  meta-models", "abstract": "  Improving the fairness of machine learning models is a nuanced task that\nrequires decision makers to reason about multiple, conflicting criteria. The\nmajority of fair machine learning methods transform the error-fairness\ntrade-off into a single objective problem with a parameter controlling the\nrelative importance of error versus fairness. We propose instead to directly\noptimize the error-fairness tradeoff by using multi-objective optimization. We\npresent a flexible framework for defining the fair machine learning task as a\nweighted classification problem with multiple cost functions. This framework is\nagnostic to the underlying prediction model as well as the metrics. We use\nmultiobjective optimization to define the sample weights used in model training\nfor a given machine learner, and adapt the weights to optimize multiple metrics\nof fairness and accuracy across a set of tasks. To reduce the number of\noptimized parameters, and to constrain their complexity with respect to\npopulation subgroups, we propose a novel meta-model approach that learns to map\nprotected attributes to sample weights, rather than optimizing those weights\ndirectly. On a set of real-world problems, this approach outperforms current\nstate-of-the-art methods by finding solution sets with preferable\nerror/fairness trade-offs.\n", "date": "2023-04-21", "categories": []}
{"id": "2209.12396", "title": "Deep Fair Clustering via Maximizing and Minimizing Mutual Information:\n  Theory, Algorithm and Metric", "abstract": "  Fair clustering aims to divide data into distinct clusters while preventing\nsensitive attributes (\\textit{e.g.}, gender, race, RNA sequencing technique)\nfrom dominating the clustering. Although a number of works have been conducted\nand achieved huge success recently, most of them are heuristical, and there\nlacks a unified theory for algorithm design. In this work, we fill this blank\nby developing a mutual information theory for deep fair clustering and\naccordingly designing a novel algorithm, dubbed FCMI. In brief, through\nmaximizing and minimizing mutual information, FCMI is designed to achieve four\ncharacteristics highly expected by deep fair clustering, \\textit{i.e.},\ncompact, balanced, and fair clusters, as well as informative features. Besides\nthe contributions to theory and algorithm, another contribution of this work is\nproposing a novel fair clustering metric built upon information theory as well.\nUnlike existing evaluation metrics, our metric measures the clustering quality\nand fairness as a whole instead of separate manner. To verify the effectiveness\nof the proposed FCMI, we conduct experiments on six benchmarks including a\nsingle-cell RNA-seq atlas compared with 11 state-of-the-art methods in terms of\nfive metrics. The code could be accessed from \\url{ https://pengxi.me}.\n", "date": "2022-09-26", "categories": []}
{"id": "2304.10655", "title": "The Dataset Multiplicity Problem: How Unreliable Data Impacts\n  Predictions", "abstract": "  We introduce dataset multiplicity, a way to study how inaccuracies,\nuncertainty, and social bias in training datasets impact test-time predictions.\nThe dataset multiplicity framework asks a counterfactual question of what the\nset of resultant models (and associated test-time predictions) would be if we\ncould somehow access all hypothetical, unbiased versions of the dataset. We\ndiscuss how to use this framework to encapsulate various sources of uncertainty\nin datasets' factualness, including systemic social bias, data collection\npractices, and noisy labels or features. We show how to exactly analyze the\nimpacts of dataset multiplicity for a specific model architecture and type of\nuncertainty: linear models with label errors. Our empirical analysis shows that\nreal-world datasets, under reasonable assumptions, contain many test samples\nwhose predictions are affected by dataset multiplicity. Furthermore, the choice\nof domain-specific dataset multiplicity definition determines what samples are\naffected, and whether different demographic groups are disparately impacted.\nFinally, we discuss implications of dataset multiplicity for machine learning\npractice and research, including considerations for when model outcomes should\nnot be trusted.\n", "date": "2023-04-20", "categories": []}
{"id": "2304.10828", "title": "Individual Fairness in Bayesian Neural Networks", "abstract": "  We study Individual Fairness (IF) for Bayesian neural networks (BNNs).\nSpecifically, we consider the $\\epsilon$-$\\delta$-individual fairness notion,\nwhich requires that, for any pair of input points that are $\\epsilon$-similar\naccording to a given similarity metrics, the output of the BNN is within a\ngiven tolerance $\\delta>0.$ We leverage bounds on statistical sampling over the\ninput space and the relationship between adversarial robustness and individual\nfairness to derive a framework for the systematic estimation of\n$\\epsilon$-$\\delta$-IF, designing Fair-FGSM and Fair-PGD as\nglobal,fairness-aware extensions to gradient-based attacks for BNNs. We\nempirically study IF of a variety of approximately inferred BNNs with different\narchitectures on fairness benchmarks, and compare against deterministic models\nlearnt using frequentist techniques. Interestingly, we find that BNNs trained\nby means of approximate Bayesian inference consistently tend to be markedly\nmore individually fair than their deterministic counterparts.\n", "date": "2023-04-21", "categories": []}
{"id": "2304.11094", "title": "Effectiveness of Debiasing Techniques: An Indigenous Qualitative\n  Analysis", "abstract": "  An indigenous perspective on the effectiveness of debiasing techniques for\npre-trained language models (PLMs) is presented in this paper. The current\ntechniques used to measure and debias PLMs are skewed towards the US racial\nbiases and rely on pre-defined bias attributes (e.g. \"black\" vs \"white\"). Some\nrequire large datasets and further pre-training. Such techniques are not\ndesigned to capture the underrepresented indigenous populations in other\ncountries, such as M\\=aori in New Zealand. Local knowledge and understanding\nmust be incorporated to ensure unbiased algorithms, especially when addressing\na resource-restricted society.\n", "date": "2023-04-17", "categories": []}
{"id": "2304.09867", "title": "Introducing Construct Theory as a Standard Methodology for Inclusive AI\n  Models", "abstract": "  Construct theory in social psychology, developed by George Kelly are mental\nconstructs to predict and anticipate events. Constructs are how humans\ninterpret, curate, predict and validate data; information. AI today is biased\nbecause it is trained with a narrow construct as defined by the training data\nlabels. Machine Learning algorithms for facial recognition discriminate against\ndarker skin colors and in the ground breaking research papers (Buolamwini, Joy\nand Timnit Gebru. Gender Shades: Intersectional Accuracy Disparities in\nCommercial Gender Classification. FAT (2018), the inclusion of phenotypic\nlabeling is proposed as a viable solution. In Construct theory, phenotype is\njust one of the many subelements that make up the construct of a face. In this\npaper, we present 15 main elements of the construct of face, with 50\nsubelements and tested Google Cloud Vision API and Microsoft Cognitive Services\nAPI using FairFace dataset that currently has data for 7 races, genders and\nages, and we retested against FairFace Plus dataset curated by us. Our results\nshow exactly where they have gaps for inclusivity. Based on our experiment\nresults, we propose that validated, inclusive constructs become industry\nstandards for AI ML models going forward.\n", "date": "2023-04-19", "categories": []}
{"id": "2304.10153", "title": "On the Independence of Association Bias and Empirical Fairness in\n  Language Models", "abstract": "  The societal impact of pre-trained language models has prompted researchers\nto probe them for strong associations between protected attributes and\nvalue-loaded terms, from slur to prestigious job titles. Such work is said to\nprobe models for bias or fairness-or such probes 'into representational biases'\nare said to be 'motivated by fairness'-suggesting an intimate connection\nbetween bias and fairness. We provide conceptual clarity by distinguishing\nbetween association biases (Caliskan et al., 2022) and empirical fairness (Shen\net al., 2022) and show the two can be independent. Our main contribution,\nhowever, is showing why this should not come as a surprise. To this end, we\nfirst provide a thought experiment, showing how association bias and empirical\nfairness can be completely orthogonal. Next, we provide empirical evidence that\nthere is no correlation between bias metrics and fairness metrics across the\nmost widely used language models. Finally, we survey the sociological and\npsychological literature and show how this literature provides ample support\nfor expecting these metrics to be uncorrelated.\n", "date": "2023-04-20", "categories": []}
{"id": "2304.09218", "title": "Generative models improve fairness of medical classifiers under\n  distribution shifts", "abstract": "  A ubiquitous challenge in machine learning is the problem of domain\ngeneralisation. This can exacerbate bias against groups or labels that are\nunderrepresented in the datasets used for model development. Model bias can\nlead to unintended harms, especially in safety-critical applications like\nhealthcare. Furthermore, the challenge is compounded by the difficulty of\nobtaining labelled data due to high cost or lack of readily available domain\nexpertise. In our work, we show that learning realistic augmentations\nautomatically from data is possible in a label-efficient manner using\ngenerative models. In particular, we leverage the higher abundance of\nunlabelled data to capture the underlying data distribution of different\nconditions and subgroups for an imaging modality. By conditioning generative\nmodels on appropriate labels, we can steer the distribution of synthetic\nexamples according to specific requirements. We demonstrate that these learned\naugmentations can surpass heuristic ones by making models more robust and\nstatistically fair in- and out-of-distribution. To evaluate the generality of\nour approach, we study 3 distinct medical imaging contexts of varying\ndifficulty: (i) histopathology images from a publicly available generalisation\nbenchmark, (ii) chest X-rays from publicly available clinical datasets, and\n(iii) dermatology images characterised by complex shifts and imaging\nconditions. Complementing real training samples with synthetic ones improves\nthe robustness of models in all three medical tasks and increases fairness by\nimproving the accuracy of diagnosis within underrepresented groups. This\napproach leads to stark improvements OOD across modalities: 7.7% prediction\naccuracy improvement in histopathology, 5.2% in chest radiology with 44.6%\nlower fairness gap and a striking 63.5% improvement in high-risk sensitivity\nfor dermatology with a 7.5x reduction in fairness gap.\n", "date": "2023-04-18", "categories": []}
{"id": "2210.16435", "title": "Scalable Spectral Clustering with Group Fairness Constraints", "abstract": "  There are synergies of research interests and industrial efforts in modeling\nfairness and correcting algorithmic bias in machine learning. In this paper, we\npresent a scalable algorithm for spectral clustering (SC) with group fairness\nconstraints. Group fairness is also known as statistical parity where in each\ncluster, each protected group is represented with the same proportion as in the\nentirety. While FairSC algorithm (Kleindessner et al., 2019) is able to find\nthe fairer clustering, it is compromised by high costs due to the kernels of\ncomputing nullspaces and the square roots of dense matrices explicitly. We\npresent a new formulation of underlying spectral computation by incorporating\nnullspace projection and Hotelling's deflation such that the resulting\nalgorithm, called s-FairSC, only involves the sparse matrix-vector products and\nis able to fully exploit the sparsity of the fair SC model. The experimental\nresults on the modified stochastic block model demonstrate that s-FairSC is\ncomparable with FairSC in recovering fair clustering. Meanwhile, it is sped up\nby a factor of 12 for moderate model sizes. s-FairSC is further demonstrated to\nbe scalable in the sense that the computational costs of s-FairSC only increase\nmarginally compared to the SC without fairness constraints.\n", "date": "2022-10-28", "categories": []}
{"id": "2211.12703", "title": "Subgroup Robustness Grows On Trees: An Empirical Baseline Investigation", "abstract": "  Researchers have proposed many methods for fair and robust machine learning,\nbut comprehensive empirical evaluation of their subgroup robustness is lacking.\nIn this work, we address this gap in the context of tabular data, where\nsensitive subgroups are clearly-defined, real-world fairness problems abound,\nand prior works often do not compare to state-of-the-art tree-based models as\nbaselines. We conduct an empirical comparison of several previously-proposed\nmethods for fair and robust learning alongside state-of-the-art tree-based\nmethods and other baselines. Via experiments with more than $340{,}000$ model\nconfigurations on eight datasets, we show that tree-based methods have strong\nsubgroup robustness, even when compared to robustness- and fairness-enhancing\nmethods. Moreover, the best tree-based models tend to show good performance\nover a range of metrics, while robust or group-fair models can show\nbrittleness, with significant performance differences across different metrics\nfor a fixed model. We also demonstrate that tree-based models show less\nsensitivity to hyperparameter configurations, and are less costly to train. Our\nwork suggests that tree-based ensemble models make an effective baseline for\ntabular data, and are a sensible default when subgroup robustness is desired.\nFor associated code and detailed results, see\nhttps://github.com/jpgard/subgroup-robustness-grows-on-trees .\n", "date": "2022-11-23", "categories": []}
{"id": "2206.03680", "title": "Improving Evaluation of Debiasing in Image Classification", "abstract": "  Image classifiers often rely overly on peripheral attributes that have a\nstrong correlation with the target class (i.e., dataset bias) when making\npredictions. Due to the dataset bias, the model correctly classifies data\nsamples including bias attributes (i.e., bias-aligned samples) while failing to\ncorrectly predict those without bias attributes (i.e., bias-conflicting\nsamples). Recently, a myriad of studies focus on mitigating such dataset bias,\nthe task of which is referred to as debiasing. However, our comprehensive study\nindicates several issues need to be improved when conducting evaluation of\ndebiasing in image classification. First, most of the previous studies do not\nspecify how they select their hyper-parameters and model checkpoints (i.e.,\ntuning criterion). Second, the debiasing studies until now evaluated their\nproposed methods on datasets with excessively high bias-severities, showing\ndegraded performance on datasets with low bias severity. Third, the debiasing\nstudies do not share consistent experimental settings (e.g., datasets and\nneural networks) which need to be standardized for fair comparisons. Based on\nsuch issues, this paper 1) proposes an evaluation metric `Align-Conflict (AC)\nscore' for the tuning criterion, 2) includes experimental settings with low\nbias severity and shows that they are yet to be explored, and 3) unifies the\nstandardized experimental settings to promote fair comparisons between\ndebiasing methods. We believe that our findings and lessons inspire future\nresearchers in debiasing to further push state-of-the-art performances with\nfair comparisons.\n", "date": "2022-06-08", "categories": []}
{"id": "2304.06901", "title": "Systemic Fairness", "abstract": "  Machine learning algorithms are increasingly used to make or support\ndecisions in a wide range of settings. With such expansive use there is also\ngrowing concern about the fairness of such methods. Prior literature on\nalgorithmic fairness has extensively addressed risks and in many cases\npresented approaches to manage some of them. However, most studies have focused\non fairness issues that arise from actions taken by a (single) focal\ndecision-maker or agent. In contrast, most real-world systems have many agents\nthat work collectively as part of a larger ecosystem. For example, in a lending\nscenario, there are multiple lenders who evaluate loans for applicants, along\nwith policymakers and other institutions whose decisions also affect outcomes.\nThus, the broader impact of any lending decision of a single decision maker\nwill likely depend on the actions of multiple different agents in the\necosystem. This paper develops formalisms for firm versus systemic fairness,\nand calls for a greater focus in the algorithmic fairness literature on\necosystem-wide fairness - or more simply systemic fairness - in real-world\ncontexts.\n", "date": "2023-04-14", "categories": []}
{"id": "2304.07030", "title": "FairRec: Fairness Testing for Deep Recommender Systems", "abstract": "  Deep learning-based recommender systems (DRSs) are increasingly and widely\ndeployed in the industry, which brings significant convenience to people's\ndaily life in different ways. However, recommender systems are also shown to\nsuffer from multiple issues,e.g., the echo chamber and the Matthew effect, of\nwhich the notation of \"fairness\" plays a core role.While many fairness\nnotations and corresponding fairness testing approaches have been developed for\ntraditional deep classification models, they are essentially hardly applicable\nto DRSs. One major difficulty is that there still lacks a systematic\nunderstanding and mapping between the existing fairness notations and the\ndiverse testing requirements for deep recommender systems, not to mention\nfurther testing or debugging activities. To address the gap, we propose\nFairRec, a unified framework that supports fairness testing of DRSs from\nmultiple customized perspectives, e.g., model utility, item diversity, item\npopularity, etc. We also propose a novel, efficient search-based testing\napproach to tackle the new challenge, i.e., double-ended discrete particle\nswarm optimization (DPSO) algorithm, to effectively search for hidden fairness\nissues in the form of certain disadvantaged groups from a vast number of\ncandidate groups. Given the testing report, by adopting a simple re-ranking\nmitigation strategy on these identified disadvantaged groups, we show that the\nfairness of DRSs can be significantly improved. We conducted extensive\nexperiments on multiple industry-level DRSs adopted by leading companies. The\nresults confirm that FairRec is effective and efficient in identifying the\ndeeply hidden fairness issues, e.g., achieving 95% testing accuracy with half\nto 1/8 time.\n", "date": "2023-04-14", "categories": []}
{"id": "2304.07175", "title": "Exploring Causes of Demographic Variations In Face Recognition Accuracy", "abstract": "  In recent years, media reports have called out bias and racism in face\nrecognition technology. We review experimental results exploring several\nspeculated causes for asymmetric cross-demographic performance. We consider\naccuracy differences as represented by variations in non-mated (impostor) and /\nor mated (genuine) distributions for 1-to-1 face matching. Possible causes\nexplored include differences in skin tone, face size and shape, imbalance in\nnumber of identities and images in the training data, and amount of face\nvisible in the test data (\"face pixels\"). We find that demographic differences\nin face pixel information of the test images appear to most directly impact the\nresultant differences in face recognition accuracy.\n", "date": "2023-04-14", "categories": []}
{"id": "2304.06031", "title": "Fairness: from the ethical principle to the practice of Machine Learning\n  development as an ongoing agreement with stakeholders", "abstract": "  This paper clarifies why bias cannot be completely mitigated in Machine\nLearning (ML) and proposes an end-to-end methodology to translate the ethical\nprinciple of justice and fairness into the practice of ML development as an\nongoing agreement with stakeholders. The pro-ethical iterative process\npresented in the paper aims to challenge asymmetric power dynamics in the\nfairness decision making within ML design and support ML development teams to\nidentify, mitigate and monitor bias at each step of ML systems development. The\nprocess also provides guidance on how to explain the always imperfect\ntrade-offs in terms of bias to users.\n", "date": "2023-03-22", "categories": []}
{"id": "2304.06057", "title": "Maximal Fairness", "abstract": "  Fairness in AI has garnered quite some attention in research, and\nincreasingly also in society. The so-called \"Impossibility Theorem\" has been\none of the more striking research results with both theoretical and practical\nconsequences, as it states that satisfying a certain combination of fairness\nmeasures is impossible. To date, this negative result has not yet been\ncomplemented with a positive one: a characterization of which combinations of\nfairness notions are possible. This work aims to fill this gap by identifying\nmaximal sets of commonly used fairness measures that can be simultaneously\nsatisfied. The fairness measures used are demographic parity, equal\nopportunity, false positive parity, predictive parity, predictive equality,\noverall accuracy equality and treatment equality. We conclude that in total 12\nmaximal sets of these fairness measures are possible, among which seven\ncombinations of two measures, and five combinations of three measures. Our work\nraises interest questions regarding the practical relevance of each of these 12\nmaximal fairness notions in various scenarios.\n", "date": "2023-04-12", "categories": []}
{"id": "2304.06484", "title": "Exploring Gender and Race Biases in the NFT Market", "abstract": "  Non-Fungible Tokens (NFTs) are non-interchangeable assets, usually digital\nart, which are stored on the blockchain. Preliminary studies find that female\nand darker-skinned NFTs are valued less than their male and lighter-skinned\ncounterparts. However, these studies analyze only the CryptoPunks collection.\nWe test the statistical significance of race and gender biases in the prices of\nCryptoPunks and present the first study of gender bias in the broader NFT\nmarket. We find evidence of racial bias but not gender bias. Our work also\nintroduces a dataset of gender-labeled NFT collections to advance the broader\nstudy of social equity in this emerging market.\n", "date": "2023-03-29", "categories": []}
{"id": "2304.06596", "title": "Beyond Submodularity: A Unified Framework of Randomized Set Selection\n  with Group Fairness Constraints", "abstract": "  Machine learning algorithms play an important role in a variety of important\ndecision-making processes, including targeted advertisement displays, home loan\napprovals, and criminal behavior predictions. Given the far-reaching impact of\nthese algorithms, it is crucial that they operate fairly, free from bias or\nprejudice towards certain groups in the population. Ensuring impartiality in\nthese algorithms is essential for promoting equality and avoiding\ndiscrimination. To this end we introduce a unified framework for randomized\nsubset selection that incorporates group fairness constraints. Our problem\ninvolves a global utility function and a set of group utility functions for\neach group, here a group refers to a group of individuals (e.g., people)\nsharing the same attributes (e.g., gender). Our aim is to generate a\ndistribution across feasible subsets, specifying the selection probability of\neach feasible set, to maximize the global utility function while meeting a\npredetermined quota for each group utility function in expectation. Note that\nthere may not necessarily be any direct connections between the global utility\nfunction and each group utility function. We demonstrate that this framework\nunifies and generalizes many significant applications in machine learning and\noperations research. Our algorithmic results either improves the best known\nresult or provide the first approximation algorithms for new applications.\n", "date": "2023-04-13", "categories": []}
{"id": "2304.05764", "title": "Measuring Normative and Descriptive Biases in Language Models Using\n  Census Data", "abstract": "  We investigate in this paper how distributions of occupations with respect to\ngender is reflected in pre-trained language models. Such distributions are not\nalways aligned to normative ideals, nor do they necessarily reflect a\ndescriptive assessment of reality. In this paper, we introduce an approach for\nmeasuring to what degree pre-trained language models are aligned to normative\nand descriptive occupational distributions. To this end, we use official\ndemographic information about gender--occupation distributions provided by the\nnational statistics agencies of France, Norway, United Kingdom, and the United\nStates. We manually generate template-based sentences combining gendered\npronouns and nouns with occupations, and subsequently probe a selection of ten\nlanguage models covering the English, French, and Norwegian languages. The\nscoring system we introduce in this work is language independent, and can be\nused on any combination of template-based sentences, occupations, and\nlanguages. The approach could also be extended to other dimensions of national\ncensus data and other demographic variables.\n", "date": "2023-04-12", "categories": []}
{"id": "2304.04761", "title": "Connecting Fairness in Machine Learning with Public Health Equity", "abstract": "  Machine learning (ML) has become a critical tool in public health, offering\nthe potential to improve population health, diagnosis, treatment selection, and\nhealth system efficiency. However, biases in data and model design can result\nin disparities for certain protected groups and amplify existing inequalities\nin healthcare. To address this challenge, this study summarizes seminal\nliterature on ML fairness and presents a framework for identifying and\nmitigating biases in the data and model. The framework provides guidance on\nincorporating fairness into different stages of the typical ML pipeline, such\nas data processing, model design, deployment, and evaluation. To illustrate the\nimpact of biases in data on ML models, we present examples that demonstrate how\nsystematic biases can be amplified through model predictions. These case\nstudies suggest how the framework can be used to prevent these biases and\nhighlight the need for fair and equitable ML models in public health. This work\naims to inform and guide the use of ML in public health towards a more ethical\nand equitable outcome for all populations.\n", "date": "2023-04-08", "categories": []}
{"id": "2304.05023", "title": "Learning Optimal Fair Scoring Systems for Multi-Class Classification", "abstract": "  Machine Learning models are increasingly used for decision making, in\nparticular in high-stakes applications such as credit scoring, medicine or\nrecidivism prediction. However, there are growing concerns about these models\nwith respect to their lack of interpretability and the undesirable biases they\ncan generate or reproduce. While the concepts of interpretability and fairness\nhave been extensively studied by the scientific community in recent years, few\nworks have tackled the general multi-class classification problem under\nfairness constraints, and none of them proposes to generate fair and\ninterpretable models for multi-class classification. In this paper, we use\nMixed-Integer Linear Programming (MILP) techniques to produce inherently\ninterpretable scoring systems under sparsity and fairness constraints, for the\ngeneral multi-class classification setup. Our work generalizes the SLIM\n(Supersparse Linear Integer Models) framework that was proposed by Rudin and\nUstun to learn optimal scoring systems for binary classification. The use of\nMILP techniques allows for an easy integration of diverse operational\nconstraints (such as, but not restricted to, fairness or sparsity), but also\nfor the building of certifiably optimal models (or sub-optimal models with\nbounded optimality gap).\n", "date": "2023-04-11", "categories": []}
{"id": "2210.06732", "title": "Equal Improvability: A New Fairness Notion Considering the Long-term\n  Impact", "abstract": "  Devising a fair classifier that does not discriminate against different\ngroups is an important problem in machine learning. Although researchers have\nproposed various ways of defining group fairness, most of them only focused on\nthe immediate fairness, ignoring the long-term impact of a fair classifier\nunder the dynamic scenario where each individual can improve its feature over\ntime. Such dynamic scenarios happen in real world, e.g., college admission and\ncredit loaning, where each rejected sample makes effort to change its features\nto get accepted afterwards. In this dynamic setting, the long-term fairness\nshould equalize the samples' feature distribution across different groups after\nthe rejected samples make some effort to improve. In order to promote long-term\nfairness, we propose a new fairness notion called Equal Improvability (EI),\nwhich equalizes the potential acceptance rate of the rejected samples across\ndifferent groups assuming a bounded level of effort will be spent by each\nrejected sample. We analyze the properties of EI and its connections with\nexisting fairness notions. To find a classifier that satisfies the EI\nrequirement, we propose and study three different approaches that solve\nEI-regularized optimization problems. Through experiments on both synthetic and\nreal datasets, we demonstrate that the proposed EI-regularized algorithms\nencourage us to find a fair classifier in terms of EI. Finally, we provide\nexperimental results on dynamic scenarios which highlight the advantages of our\nEI metric in achieving the long-term fairness. Codes are available in a GitHub\nrepository, see https://github.com/guldoganozgur/ei_fairness.\n", "date": "2022-10-13", "categories": []}
{"id": "2304.04091", "title": "Best Arm Identification with Fairness Constraints on Subpopulations", "abstract": "  We formulate, analyze and solve the problem of best arm identification with\nfairness constraints on subpopulations (BAICS). Standard best arm\nidentification problems aim at selecting an arm that has the largest expected\nreward where the expectation is taken over the entire population. The BAICS\nproblem requires that an selected arm must be fair to all subpopulations (e.g.,\ndifferent ethnic groups, age groups, or customer types) by satisfying\nconstraints that the expected reward conditional on every subpopulation needs\nto be larger than some thresholds. The BAICS problem aims at correctly\nidentify, with high confidence, the arm with the largest expected reward from\nall arms that satisfy subpopulation constraints. We analyze the complexity of\nthe BAICS problem by proving a best achievable lower bound on the sample\ncomplexity with closed-form representation. We then design an algorithm and\nprove that the algorithm's sample complexity matches with the lower bound in\nterms of order. A brief account of numerical experiments are conducted to\nillustrate the theoretical findings.\n", "date": "2023-04-08", "categories": []}
{"id": "2304.04199", "title": "Information-Theoretic Testing and Debugging of Fairness Defects in Deep\n  Neural Networks", "abstract": "  The deep feedforward neural networks (DNNs) are increasingly deployed in\nsocioeconomic critical decision support software systems. DNNs are\nexceptionally good at finding minimal, sufficient statistical patterns within\ntheir training data. Consequently, DNNs may learn to encode decisions --\namplifying existing biases or introducing new ones -- that may disadvantage\nprotected individuals/groups and may stand to violate legal protections. While\nthe existing search based software testing approaches have been effective in\ndiscovering fairness defects, they do not supplement these defects with\ndebugging aids -- such as severity and causal explanations -- crucial to help\ndevelopers triage and decide on the next course of action. Can we measure the\nseverity of fairness defects in DNNs? Are these defects symptomatic of improper\ntraining or they merely reflect biases present in the training data? To answer\nsuch questions, we present DICE: an information-theoretic testing and debugging\nframework to discover and localize fairness defects in DNNs.\n  The key goal of DICE is to assist software developers in triaging fairness\ndefects by ordering them by their severity. Towards this goal, we quantify\nfairness in terms of protected information (in bits) used in decision making. A\nquantitative view of fairness defects not only helps in ordering these defects,\nour empirical evaluation shows that it improves the search efficiency due to\nresulting smoothness of the search space. Guided by the quantitative fairness,\nwe present a causal debugging framework to localize inadequately trained layers\nand neurons responsible for fairness defects. Our experiments over ten DNNs,\ndeveloped for socially critical tasks, show that DICE efficiently characterizes\nthe amounts of discrimination, effectively generates discriminatory instances,\nand localizes layers/neurons with significant biases.\n", "date": "2023-04-09", "categories": []}
{"id": "2304.04222", "title": "CILIATE: Towards Fairer Class-based Incremental Learning by Dataset and\n  Training Refinement", "abstract": "  Due to the model aging problem, Deep Neural Networks (DNNs) need updates to\nadjust them to new data distributions. The common practice leverages\nincremental learning (IL), e.g., Class-based Incremental Learning (CIL) that\nupdates output labels, to update the model with new data and a limited number\nof old data. This avoids heavyweight training (from scratch) using conventional\nmethods and saves storage space by reducing the number of old data to store.\nBut it also leads to poor performance in fairness. In this paper, we show that\nCIL suffers both dataset and algorithm bias problems, and existing solutions\ncan only partially solve the problem. We propose a novel framework, CILIATE,\nthat fixes both dataset and algorithm bias in CIL. It features a novel\ndifferential analysis guided dataset and training refinement process that\nidentifies unique and important samples overlooked by existing CIL and enforces\nthe model to learn from them. Through this process, CILIATE improves the\nfairness of CIL by 17.03%, 22.46%, and 31.79% compared to state-of-the-art\nmethods, iCaRL, BiC, and WA, respectively, based on our evaluation on three\npopular datasets and widely used ResNet models.\n", "date": "2023-04-09", "categories": []}
{"id": "2304.04679", "title": "FairPilot: An Explorative System for Hyperparameter Tuning through the\n  Lens of Fairness", "abstract": "  Despite the potential benefits of machine learning (ML) in high-risk\ndecision-making domains, the deployment of ML is not accessible to\npractitioners, and there is a risk of discrimination. To establish trust and\nacceptance of ML in such domains, democratizing ML tools and fairness\nconsideration are crucial. In this paper, we introduce FairPilot, an\ninteractive system designed to promote the responsible development of ML models\nby exploring a combination of various models, different hyperparameters, and a\nwide range of fairness definitions. We emphasize the challenge of selecting the\n``best\" ML model and demonstrate how FairPilot allows users to select a set of\nevaluation criteria and then displays the Pareto frontier of models and\nhyperparameters as an interactive map. FairPilot is the first system to combine\nthese features, offering a unique opportunity for users to responsibly choose\ntheir model.\n", "date": "2023-04-10", "categories": []}
{"id": "2304.04700", "title": "Achieving Long-term Fairness in Submodular Maximization through\n  Randomization", "abstract": "  Submodular function optimization has numerous applications in machine\nlearning and data analysis, including data summarization which aims to identify\na concise and diverse set of data points from a large dataset. It is important\nto implement fairness-aware algorithms when dealing with data items that may\ncontain sensitive attributes like race or gender, to prevent biases that could\nlead to unequal representation of different groups. With this in mind, we\ninvestigate the problem of maximizing a monotone submodular function while\nmeeting group fairness constraints. Unlike previous studies in this area, we\nallow for randomized solutions, with the objective being to calculate a\ndistribution over feasible sets such that the expected number of items selected\nfrom each group is subject to constraints in the form of upper and lower\nthresholds, ensuring that the representation of each group remains balanced in\nthe long term. Here a set is considered feasible if its size does not exceed a\nconstant value of $b$. Our research includes the development of a series of\napproximation algorithms for this problem.\n", "date": "2023-04-10", "categories": []}
{"id": "2304.02828", "title": "Uncurated Image-Text Datasets: Shedding Light on Demographic Bias", "abstract": "  The increasing tendency to collect large and uncurated datasets to train\nvision-and-language models has raised concerns about fair representations. It\nis known that even small but manually annotated datasets, such as MSCOCO, are\naffected by societal bias. This problem, far from being solved, may be getting\nworse with data crawled from the Internet without much control. In addition,\nthe lack of tools to analyze societal bias in big collections of images makes\naddressing the problem extremely challenging. Our first contribution is to\nannotate part of the Google Conceptual Captions dataset, widely used for\ntraining vision-and-language models, with four demographic and two contextual\nattributes. Our second contribution is to conduct a comprehensive analysis of\nthe annotations, focusing on how different demographic groups are represented.\nOur last contribution lies in evaluating three prevailing vision-and-language\ntasks: image captioning, text-image CLIP embeddings, and text-to-image\ngeneration, showing that societal bias is a persistent problem in all of them.\n", "date": "2023-04-06", "categories": []}
{"id": "2304.00061", "title": "To be Robust and to be Fair: Aligning Fairness with Robustness", "abstract": "  Adversarial training has been shown to be reliable in improving robustness\nagainst adversarial samples. However, the problem of adversarial training in\nterms of fairness has not yet been properly studied, and the relationship\nbetween fairness and accuracy attack still remains unclear. Can we\nsimultaneously improve robustness w.r.t. both fairness and accuracy? To tackle\nthis topic, in this paper, we study the problem of adversarial training and\nadversarial attack w.r.t. both metrics. We propose a unified structure for\nfairness attack which brings together common notions in group fairness, and we\ntheoretically prove the equivalence of fairness attack against different\nnotions. Moreover, we show the alignment of fairness and accuracy attack, and\ntheoretically demonstrate that robustness w.r.t. one metric benefits from\nrobustness w.r.t. the other metric. Our study suggests a novel way to unify\nadversarial training and attack w.r.t. fairness and accuracy, and experimental\nresults show that our proposed method achieves better performance in terms of\nrobustness w.r.t. both metrics.\n", "date": "2023-03-31", "categories": []}
{"id": "2304.00295", "title": "Fair-CDA: Continuous and Directional Augmentation for Group Fairness", "abstract": "  In this work, we propose {\\it Fair-CDA}, a fine-grained data augmentation\nstrategy for imposing fairness constraints. We use a feature disentanglement\nmethod to extract the features highly related to the sensitive attributes. Then\nwe show that group fairness can be achieved by regularizing the models on\ntransition paths of sensitive features between groups. By adjusting the\nperturbation strength in the direction of the paths, our proposed augmentation\nis controllable and auditable. To alleviate the accuracy degradation caused by\nfairness constraints, we further introduce a calibrated model to impute labels\nfor the augmented data. Our proposed method does not assume any data generative\nmodel and ensures good generalization for both accuracy and fairness.\nExperimental results show that Fair-CDA consistently outperforms\nstate-of-the-art methods on widely-used benchmarks, e.g., Adult, CelebA and\nMovieLens. Especially, Fair-CDA obtains an 86.3\\% relative improvement for\nfairness while maintaining the accuracy on the Adult dataset. Moreover, we\nevaluate Fair-CDA in an online recommendation system to demonstrate the\neffectiveness of our method in terms of accuracy and fairness.\n", "date": "2023-04-01", "categories": []}
{"id": "2304.00305", "title": "Predictive Heterogeneity: Measures and Applications", "abstract": "  As an intrinsic and fundamental property of big data, data heterogeneity\nexists in a variety of real-world applications, such as precision medicine,\nautonomous driving, financial applications, etc. For machine learning\nalgorithms, the ignorance of data heterogeneity will greatly hurt the\ngeneralization performance and the algorithmic fairness, since the prediction\nmechanisms among different sub-populations are likely to differ from each\nother. In this work, we focus on the data heterogeneity that affects the\nprediction of machine learning models, and firstly propose the \\emph{usable\npredictive heterogeneity}, which takes into account the model capacity and\ncomputational constraints. We prove that it can be reliably estimated from\nfinite data with probably approximately correct (PAC) bounds. Additionally, we\ndesign a bi-level optimization algorithm to explore the usable predictive\nheterogeneity from data. Empirically, the explored heterogeneity provides\ninsights for sub-population divisions in income prediction, crop yield\nprediction and image classification tasks, and leveraging such heterogeneity\nbenefits the out-of-distribution generalization performance.\n", "date": "2023-04-01", "categories": []}
{"id": "2304.00486", "title": "Learning by Grouping: A Multilevel Optimization Framework for Improving\n  Fairness in Classification without Losing Accuracy", "abstract": "  The integration of machine learning models in various real-world applications\nis becoming more prevalent to assist humans in their daily decision-making\ntasks as a result of recent advancements in this field. However, it has been\ndiscovered that there is a tradeoff between the accuracy and fairness of these\ndecision-making tasks. In some cases, these AI systems can be unfair by\nexhibiting bias or discrimination against certain social groups, which can have\nsevere consequences in real life. Inspired by one of the most well-known human\nlearning skills called grouping, we address this issue by proposing a novel\nmachine learning framework where the ML model learns to group a diverse set of\nproblems into distinct subgroups to solve each subgroup using its specific\nsub-model. Our proposed framework involves three stages of learning, which are\nformulated as a three-level optimization problem: (i) learning to group\nproblems into different subgroups; (ii) learning group-specific sub-models for\nproblem-solving; and (iii) updating group assignments of training examples by\nminimizing the validation loss. These three learning stages are performed\nend-to-end in a joint manner using gradient descent. To improve fairness and\naccuracy, we develop an efficient optimization algorithm to solve this\nthree-level optimization problem. To further reduce the risk of overfitting in\nsmall datasets, we incorporate domain adaptation techniques in the second stage\nof training. We further apply our method to neural architecture search.\nExtensive experiments on various datasets demonstrate our method's\neffectiveness and performance improvements in both fairness and accuracy. Our\nproposed Learning by Grouping can reduce overfitting and achieve\nstate-of-the-art performances with fixed human-designed network architectures\nand searchable network architectures on various datasets.\n", "date": "2023-04-02", "categories": []}
{"id": "2303.16963", "title": "Fairness-Aware Data Valuation for Supervised Learning", "abstract": "  Data valuation is a ML field that studies the value of training instances\ntowards a given predictive task. Although data bias is one of the main sources\nof downstream model unfairness, previous work in data valuation does not\nconsider how training instances may influence both performance and fairness of\nML models. Thus, we propose Fairness-Aware Data vauatiOn (FADO), a data\nvaluation framework that can be used to incorporate fairness concerns into a\nseries of ML-related tasks (e.g., data pre-processing, exploratory data\nanalysis, active learning). We propose an entropy-based data valuation metric\nsuited to address our two-pronged goal of maximizing both performance and\nfairness, which is more computationally efficient than existing metrics. We\nthen show how FADO can be applied as the basis for unfairness mitigation\npre-processing techniques. Our methods achieve promising results -- up to a 40\np.p. improvement in fairness at a less than 1 p.p. loss in performance compared\nto a baseline -- and promote fairness in a data-centric way, where a deeper\nunderstanding of data quality takes center stage.\n", "date": "2023-03-29", "categories": []}
{"id": "2207.06591", "title": "A methodology to characterize bias and harmful stereotypes in natural\n  language processing in Latin America", "abstract": "  Automated decision-making systems, especially those based on natural language\nprocessing, are pervasive in our lives. They are not only behind the internet\nsearch engines we use daily, but also take more critical roles: selecting\ncandidates for a job, determining suspects of a crime, diagnosing autism and\nmore. Such automated systems make errors, which may be harmful in many ways, be\nit because of the severity of the consequences (as in health issues) or because\nof the sheer number of people they affect. When errors made by an automated\nsystem affect a population more than others, we call the system\n\\textit{biased}.\n  Most modern natural language technologies are based on artifacts obtained\nfrom enormous volumes of text using machine learning, namely language models\nand word embeddings. Since they are created by applying subsymbolic machine\nlearning, mostly artificial neural networks, they are opaque and practically\nuninterpretable by direct inspection, thus making it very difficult to audit\nthem.\n  In this paper, we present a methodology that spells out how social\nscientists, domain experts, and machine learning experts can collaboratively\nexplore biases and harmful stereotypes in word embeddings and large language\nmodels. Our methodology is based on the following principles:\n  * focus on the linguistic manifestations of discrimination on word embeddings\nand language models, not on the mathematical properties of the models * reduce\nthe technical barrier for discrimination experts%, be it social scientists,\ndomain experts or other * characterize through a qualitative exploratory\nprocess in addition to a metric-based approach * address mitigation as part of\nthe training process, not as an afterthought\n", "date": "2022-07-14", "categories": []}
{"id": "2303.16626", "title": "Fairlearn: Assessing and Improving Fairness of AI Systems", "abstract": "  Fairlearn is an open source project to help practitioners assess and improve\nfairness of artificial intelligence (AI) systems. The associated Python\nlibrary, also named fairlearn, supports evaluation of a model's output across\naffected populations and includes several algorithms for mitigating fairness\nissues. Grounded in the understanding that fairness is a sociotechnical\nchallenge, the project integrates learning resources that aid practitioners in\nconsidering a system's broader societal context.\n", "date": "2023-03-29", "categories": []}
{"id": "2303.15697", "title": "Model and Evaluation: Towards Fairness in Multilingual Text\n  Classification", "abstract": "  Recently, more and more research has focused on addressing bias in text\nclassification models. However, existing research mainly focuses on the\nfairness of monolingual text classification models, and research on fairness\nfor multilingual text classification is still very limited. In this paper, we\nfocus on the task of multilingual text classification and propose a debiasing\nframework for multilingual text classification based on contrastive learning.\nOur proposed method does not rely on any external language resources and can be\nextended to any other languages. The model contains four modules: multilingual\ntext representation module, language fusion module, text debiasing module, and\ntext classification module. The multilingual text representation module uses a\nmultilingual pre-trained language model to represent the text, the language\nfusion module makes the semantic spaces of different languages tend to be\nconsistent through contrastive learning, and the text debiasing module uses\ncontrastive learning to make the model unable to identify sensitive attributes'\ninformation. The text classification module completes the basic tasks of\nmultilingual text classification. In addition, the existing research on the\nfairness of multilingual text classification is relatively simple in the\nevaluation mode. The evaluation method of fairness is the same as the\nmonolingual equality difference evaluation method, that is, the evaluation is\nperformed on a single language. We propose a multi-dimensional fairness\nevaluation framework for multilingual text classification, which evaluates the\nmodel's monolingual equality difference, multilingual equality difference,\nmultilingual equality performance difference, and destructiveness of the\nfairness strategy. We hope that our work can provide a more general debiasing\nmethod and a more comprehensive evaluation framework for multilingual text\nfairness tasks.\n", "date": "2023-03-28", "categories": []}
{"id": "2303.14332", "title": "Using Simple Incentives to Improve Two-Sided Fairness in Ridesharing\n  Systems", "abstract": "  State-of-the-art order dispatching algorithms for ridesharing batch passenger\nrequests and allocate them to a fleet of vehicles in a centralized manner,\noptimizing over the estimated values of each passenger-vehicle matching using\ninteger linear programming (ILP). Using good estimates of future values, such\nILP-based approaches are able to significantly increase the service rates\n(percentage of requests served) for a fixed fleet of vehicles. However, such\napproaches that focus solely on maximizing efficiency can lead to disparities\nfor both drivers (e.g., income inequality) and passengers (e.g., inequality of\nservice for different groups). Existing approaches that consider fairness only\ndo it for naive assignment policies, require extensive training, or look at\nonly single-sided fairness. We propose a simple incentive-based fairness scheme\nthat can be implemented online as a part of this ILP formulation that allows us\nto improve fairness over a variety of fairness metrics. Deriving from a lens of\nvariance minimization, we describe how these fairness incentives can be\nformulated for two distinct use cases for passenger groups and driver fairness.\nWe show that under mild conditions, our approach can guarantee an improvement\nin the chosen metric for the worst-off individual. We also show empirically\nthat our Simple Incentives approach significantly outperforms prior art,\ndespite requiring no retraining; indeed, it often leads to a large improvement\nover the state-of-the-art fairness-aware approach in both overall service rate\nand fairness.\n", "date": "2023-03-25", "categories": []}
{"id": "2303.14411", "title": "Fairness meets Cross-Domain Learning: a new perspective on Models and\n  Metrics", "abstract": "  Deep learning-based recognition systems are deployed at scale for several\nreal-world applications that inevitably involve our social life. Although being\nof great support when making complex decisions, they might capture spurious\ndata correlations and leverage sensitive attributes (e.g. age, gender,\nethnicity). How to factor out this information while keeping a high prediction\nperformance is a task with still several open questions, many of which are\nshared with those of the domain adaptation and generalization literature which\nfocuses on avoiding visual domain biases. In this work, we propose an in-depth\nstudy of the relationship between cross-domain learning (CD) and model fairness\nby introducing a benchmark on face and medical images spanning several\ndemographic groups as well as classification and localization tasks. After\nhaving highlighted the limits of the current evaluation metrics, we introduce a\nnew Harmonic Fairness (HF) score to assess jointly how fair and accurate every\nmodel is with respect to a reference baseline. Our study covers 14 CD\napproaches alongside three state-of-the-art fairness algorithms and shows how\nthe former can outperform the latter. Overall, our work paves the way for a\nmore systematic analysis of fairness problems in computer vision. Code\navailable at: https://github.com/iurada/fairness_crossdomain\n", "date": "2023-03-25", "categories": []}
{"id": "2303.14591", "title": "FairGAT: Fairness-aware Graph Attention Networks", "abstract": "  Graphs can facilitate modeling various complex systems such as gene networks\nand power grids, as well as analyzing the underlying relations within them.\nLearning over graphs has recently attracted increasing attention, particularly\ngraph neural network-based (GNN) solutions, among which graph attention\nnetworks (GATs) have become one of the most widely utilized neural network\nstructures for graph-based tasks. Although it is shown that the use of graph\nstructures in learning results in the amplification of algorithmic bias, the\ninfluence of the attention design in GATs on algorithmic bias has not been\ninvestigated. Motivated by this, the present study first carries out a\ntheoretical analysis in order to demonstrate the sources of algorithmic bias in\nGAT-based learning for node classification. Then, a novel algorithm, FairGAT,\nthat leverages a fairness-aware attention design is developed based on the\ntheoretical findings. Experimental results on real-world networks demonstrate\nthat FairGAT improves group fairness measures while also providing comparable\nutility to the fairness-aware baselines for node classification and link\nprediction.\n", "date": "2023-03-26", "categories": []}
{"id": "2303.14665", "title": "Achieving Counterfactual Fairness with Imperfect Structural Causal Model", "abstract": "  Counterfactual fairness alleviates the discrimination between the model\nprediction toward an individual in the actual world (observational data) and\nthat in counterfactual world (i.e., what if the individual belongs to other\nsensitive groups). The existing studies need to pre-define the structural\ncausal model that captures the correlations among variables for counterfactual\ninference; however, the underlying causal model is usually unknown and\ndifficult to be validated in real-world scenarios. Moreover, the\nmisspecification of the causal model potentially leads to poor performance in\nmodel prediction and thus makes unfair decisions. In this research, we propose\na novel minimax game-theoretic model for counterfactual fairness that can\nproduce accurate results meanwhile achieve a counterfactually fair decision\nwith the relaxation of strong assumptions of structural causal models. In\naddition, we also theoretically prove the error bound of the proposed minimax\nmodel. Empirical experiments on multiple real-world datasets illustrate our\nsuperior performance in both accuracy and fairness. Source code is available at\n\\url{https://github.com/tridungduong16/counterfactual_fairness_game_theoretic}.\n", "date": "2023-03-26", "categories": []}
{"id": "2302.01574", "title": "An Operational Perspective to Fairness Interventions: Where and How to\n  Intervene", "abstract": "  As AI-based decision systems proliferate, their successful operationalization\nrequires balancing multiple desiderata: predictive performance, disparity\nacross groups, safeguarding sensitive group attributes (e.g., race), and\nengineering cost. We present a holistic framework for evaluating and\ncontextualizing fairness interventions with respect to the above desiderata.\nThe two key points of practical consideration are \\emph{where} (pre-, in-,\npost-processing) and \\emph{how} (in what way the sensitive group data is used)\nthe intervention is introduced. We demonstrate our framework with a case study\non predictive parity. In it, we first propose a novel method for achieving\npredictive parity fairness without using group data at inference time via\ndistibutionally robust optimization. Then, we showcase the effectiveness of\nthese methods in a benchmarking study of close to 400 variations across two\nmajor model types (XGBoost vs. Neural Net), ten datasets, and over twenty\nunique methodologies. Methodological insights derived from our empirical study\ninform the practical design of ML workflow with fairness as a central concern.\nWe find predictive parity is difficult to achieve without using group data, and\ndespite requiring group data during model training (but not inference),\ndistributionally robust methods we develop provide significant Pareto\nimprovement. Moreover, a plain XGBoost model often Pareto-dominates neural\nnetworks with fairness interventions, highlighting the importance of model\ninductive bias.\n", "date": "2023-02-03", "categories": []}
{"id": "2303.12734", "title": "MultiModal Bias: Introducing a Framework for Stereotypical Bias\n  Assessment beyond Gender and Race in Vision Language Models", "abstract": "  Recent breakthroughs in self supervised training have led to a new class of\npretrained vision language models. While there have been investigations of bias\nin multimodal models, they have mostly focused on gender and racial bias,\ngiving much less attention to other relevant groups, such as minorities with\nregard to religion, nationality, sexual orientation, or disabilities. This is\nmainly due to lack of suitable benchmarks for such groups. We seek to address\nthis gap by providing a visual and textual bias benchmark called MMBias,\nconsisting of around 3,800 images and phrases covering 14 population subgroups.\nWe utilize this dataset to assess bias in several prominent self supervised\nmultimodal models, including CLIP, ALBEF, and ViLT. Our results show that these\nmodels demonstrate meaningful bias favoring certain groups. Finally, we\nintroduce a debiasing method designed specifically for such large pre-trained\nmodels that can be applied as a post-processing step to mitigate bias, while\npreserving the remaining accuracy of the model.\n", "date": "2023-03-16", "categories": []}
{"id": "2303.11449", "title": "Bias mitigation techniques in image classification: fair machine\n  learning in human heritage collections", "abstract": "  A major problem with using automated classification systems is that if they\nare not engineered correctly and with fairness considerations, they could be\ndetrimental to certain populations. Furthermore, while engineers have developed\ncutting-edge technologies for image classification, there is still a gap in the\napplication of these models in human heritage collections, where data sets\nusually consist of low-quality pictures of people with diverse ethnicity,\ngender, and age. In this work, we evaluate three bias mitigation techniques\nusing two state-of-the-art neural networks, Xception and EfficientNet, for\ngender classification. Moreover, we explore the use of transfer learning using\na fair data set to overcome the training data scarcity. We evaluated the\neffectiveness of the bias mitigation pipeline on a cultural heritage collection\nof photographs from the 19th and 20th centuries, and we used the FairFace data\nset for the transfer learning experiments. After the evaluation, we found that\ntransfer learning is a good technique that allows better performance when\nworking with a small data set. Moreover, the fairest classifier was found to be\naccomplished using transfer learning, threshold change, re-weighting and image\naugmentation as bias mitigation methods.\n", "date": "2023-03-20", "categories": []}
{"id": "2303.11459", "title": "Fairness-Aware Graph Filter Design", "abstract": "  Graphs are mathematical tools that can be used to represent complex\nreal-world systems, such as financial markets and social networks. Hence,\nmachine learning (ML) over graphs has attracted significant attention recently.\nHowever, it has been demonstrated that ML over graphs amplifies the already\nexisting bias towards certain under-represented groups in various\ndecision-making problems due to the information aggregation over biased graph\nstructures. Faced with this challenge, in this paper, we design a fair graph\nfilter that can be employed in a versatile manner for graph-based learning\ntasks. The design of the proposed filter is based on a bias analysis and its\noptimality in mitigating bias compared to its fairness-agnostic counterpart is\nestablished. Experiments on real-world networks for node classification\ndemonstrate the efficacy of the proposed filter design in mitigating bias,\nwhile attaining similar utility and better stability compared to baseline\nalgorithms.\n", "date": "2023-03-20", "categories": []}
{"id": "2303.11529", "title": "Counterfactually Fair Regression with Double Machine Learning", "abstract": "  Counterfactual fairness is an approach to AI fairness that tries to make\ndecisions based on the outcomes that an individual with some kind of sensitive\nstatus would have had without this status. This paper proposes Double Machine\nLearning (DML) Fairness which analogises this problem of counterfactual\nfairness in regression problems to that of estimating counterfactual outcomes\nin causal inference under the Potential Outcomes framework. It uses arbitrary\nmachine learning methods to partial out the effect of sensitive variables on\nnonsensitive variables and outcomes. Assuming that the effects of the two sets\nof variables are additively separable, outcomes will be approximately equalised\nand individual-level outcomes will be counterfactually fair. This paper\ndemonstrates the approach in a simulation study pertaining to discrimination in\nworkplace hiring and an application on real data estimating the GPAs of law\nschool students. It then discusses when it is appropriate to apply such a\nmethod to problems of real-world discrimination where constructs are\nconceptually complex and finally, whether DML Fairness can achieve justice in\nthese settings.\n", "date": "2023-03-21", "categories": []}
{"id": "2303.10431", "title": "DeAR: Debiasing Vision-Language Models with Additive Residuals", "abstract": "  Large pre-trained vision-language models (VLMs) reduce the time for\ndeveloping predictive models for various vision-grounded language downstream\ntasks by providing rich, adaptable image and text representations. However,\nthese models suffer from societal biases owing to the skewed distribution of\nvarious identity groups in the training data. These biases manifest as the\nskewed similarity between the representations for specific text concepts and\nimages of people of different identity groups and, therefore, limit the\nusefulness of such models in real-world high-stakes applications. In this work,\nwe present DeAR (Debiasing with Additive Residuals), a novel debiasing method\nthat learns additive residual image representations to offset the original\nrepresentations, ensuring fair output representations. In doing so, it reduces\nthe ability of the representations to distinguish between the different\nidentity groups. Further, we observe that the current fairness tests are\nperformed on limited face image datasets that fail to indicate why a specific\ntext concept should/should not apply to them. To bridge this gap and better\nevaluate DeAR, we introduce the Protected Attribute Tag Association (PATA)\ndataset - a new context-based bias benchmarking dataset for evaluating the\nfairness of large pre-trained VLMs. Additionally, PATA provides visual context\nfor a diverse human population in different scenarios with both positive and\nnegative connotations. Experimental results for fairness and zero-shot\nperformance preservation using multiple datasets demonstrate the efficacy of\nour framework.\n", "date": "2023-03-18", "categories": []}
{"id": "2212.10154", "title": "Human-Guided Fair Classification for Natural Language Processing", "abstract": "  Text classifiers have promising applications in high-stake tasks such as\nresume screening and content moderation. These classifiers must be fair and\navoid discriminatory decisions by being invariant to perturbations of sensitive\nattributes such as gender or ethnicity. However, there is a gap between human\nintuition about these perturbations and the formal similarity specifications\ncapturing them. While existing research has started to address this gap,\ncurrent methods are based on hardcoded word replacements, resulting in\nspecifications with limited expressivity or ones that fail to fully align with\nhuman intuition (e.g., in cases of asymmetric counterfactuals). This work\nproposes novel methods for bridging this gap by discovering expressive and\nintuitive individual fairness specifications. We show how to leverage\nunsupervised style transfer and GPT-3's zero-shot capabilities to automatically\ngenerate expressive candidate pairs of semantically similar sentences that\ndiffer along sensitive attributes. We then validate the generated pairs via an\nextensive crowdsourcing study, which confirms that a lot of these pairs align\nwith human intuition about fairness in the context of toxicity classification.\nFinally, we show how limited amounts of human feedback can be leveraged to\nlearn a similarity specification that can be used to train downstream\nfairness-aware models.\n", "date": "2022-12-20", "categories": []}
{"id": "2303.08157", "title": "Graph Neural Network Surrogates of Fair Graph Filtering", "abstract": "  Graph filters that transform prior node values to posterior scores via edge\npropagation often support graph mining tasks affecting humans, such as\nrecommendation and ranking. Thus, it is important to make them fair in terms of\nsatisfying statistical parity constraints between groups of nodes (e.g.,\ndistribute score mass between genders proportionally to their representation).\nTo achieve this while minimally perturbing the original posteriors, we\nintroduce a filter-aware universal approximation framework for posterior\nobjectives. This defines appropriate graph neural networks trained at runtime\nto be similar to filters but also locally optimize a large class of objectives,\nincluding fairness-aware ones. Experiments on a collection of 8 filters and 5\ngraphs show that our approach performs equally well or better than alternatives\nin meeting parity constraints while preserving the AUC of score-based community\nmember recommendation and creating minimal utility loss in prior diffusion.\n", "date": "2023-03-14", "categories": []}
{"id": "2303.09527", "title": "Fairness-aware Differentially Private Collaborative Filtering", "abstract": "  Recently, there has been an increasing adoption of differential privacy\nguided algorithms for privacy-preserving machine learning tasks. However, the\nuse of such algorithms comes with trade-offs in terms of algorithmic fairness,\nwhich has been widely acknowledged. Specifically, we have empirically observed\nthat the classical collaborative filtering method, trained by differentially\nprivate stochastic gradient descent (DP-SGD), results in a disparate impact on\nuser groups with respect to different user engagement levels. This, in turn,\ncauses the original unfair model to become even more biased against inactive\nusers. To address the above issues, we propose \\textbf{DP-Fair}, a two-stage\nframework for collaborative filtering based algorithms. Specifically, it\ncombines differential privacy mechanisms with fairness constraints to protect\nuser privacy while ensuring fair recommendations. The experimental results,\nbased on Amazon datasets, and user history logs collected from Etsy, one of the\nlargest e-commerce platforms, demonstrate that our proposed method exhibits\nsuperior performance in terms of both overall accuracy and user group fairness\non both shallow and deep recommendation models compared to vanilla DP-SGD.\n", "date": "2023-03-16", "categories": []}
{"id": "2209.13627", "title": "How GPT-3 responds to different publics on climate change and Black\n  Lives Matter: A critical appraisal of equity in conversational AI", "abstract": "  Autoregressive language models, which use deep learning to produce human-like\ntexts, have become increasingly widespread. Such models are powering popular\nvirtual assistants in areas like smart health, finance, and autonomous driving.\nWhile the parameters of these large language models are improving, concerns\npersist that these models might not work equally for all subgroups in society.\nDespite growing discussions of AI fairness across disciplines, there lacks\nsystemic metrics to assess what equity means in dialogue systems and how to\nengage different populations in the assessment loop. Grounded in theories of\ndeliberative democracy and science and technology studies, this paper proposes\nan analytical framework for unpacking the meaning of equity in human-AI\ndialogues. Using this framework, we conducted an auditing study to examine how\nGPT-3 responded to different sub-populations on crucial science and social\ntopics: climate change and the Black Lives Matter (BLM) movement. Our corpus\nconsists of over 20,000 rounds of dialogues between GPT-3 and 3290 individuals\nwho vary in gender, race and ethnicity, education level, English as a first\nlanguage, and opinions toward the issues. We found a substantively worse user\nexperience with GPT-3 among the opinion and the education minority\nsubpopulations; however, these two groups achieved the largest knowledge gain,\nchanging attitudes toward supporting BLM and climate change efforts after the\nchat. We traced these user experience divides to conversational differences and\nfound that GPT-3 used more negative expressions when it responded to the\neducation and opinion minority groups, compared to its responses to the\nmajority groups. We discuss the implications of our findings for a deliberative\nconversational AI system that centralizes diversity, equity, and inclusion.\n", "date": "2022-09-27", "categories": []}
{"id": "2303.08403", "title": "DualFair: Fair Representation Learning at Both Group and Individual\n  Levels via Contrastive Self-supervision", "abstract": "  Algorithmic fairness has become an important machine learning problem,\nespecially for mission-critical Web applications. This work presents a\nself-supervised model, called DualFair, that can debias sensitive attributes\nlike gender and race from learned representations. Unlike existing models that\ntarget a single type of fairness, our model jointly optimizes for two fairness\ncriteria - group fairness and counterfactual fairness - and hence makes fairer\npredictions at both the group and individual levels. Our model uses contrastive\nloss to generate embeddings that are indistinguishable for each protected\ngroup, while forcing the embeddings of counterfactual pairs to be similar. It\nthen uses a self-knowledge distillation method to maintain the quality of\nrepresentation for the downstream tasks. Extensive analysis over multiple\ndatasets confirms the model's validity and further shows the synergy of jointly\naddressing two fairness criteria, suggesting the model's potential value in\nfair intelligent Web applications.\n", "date": "2023-03-15", "categories": []}
{"id": "2207.10018", "title": "Mitigating Algorithmic Bias with Limited Annotations", "abstract": "  Existing work on fairness modeling commonly assumes that sensitive attributes\nfor all instances are fully available, which may not be true in many real-world\napplications due to the high cost of acquiring sensitive information. When\nsensitive attributes are not disclosed or available, it is needed to manually\nannotate a small part of the training data to mitigate bias. However, the\nskewed distribution across different sensitive groups preserves the skewness of\nthe original dataset in the annotated subset, which leads to non-optimal bias\nmitigation. To tackle this challenge, we propose Active Penalization Of\nDiscrimination (APOD), an interactive framework to guide the limited\nannotations towards maximally eliminating the effect of algorithmic bias. The\nproposed APOD integrates discrimination penalization with active instance\nselection to efficiently utilize the limited annotation budget, and it is\ntheoretically proved to be capable of bounding the algorithmic bias. According\nto the evaluation on five benchmark datasets, APOD outperforms the\nstate-of-the-arts baseline methods under the limited annotation budget, and\nshows comparable performance to fully annotated bias mitigation, which\ndemonstrates that APOD could benefit real-world applications when sensitive\ninformation is limited.\n", "date": "2022-07-20", "categories": []}
{"id": "2303.07615", "title": "Variation of Gender Biases in Visual Recognition Models Before and After\n  Finetuning", "abstract": "  We introduce a framework to measure how biases change before and after\nfine-tuning a large scale visual recognition model for a downstream task. Deep\nlearning models trained on increasing amounts of data are known to encode\nsocietal biases. Many computer vision systems today rely on models typically\npretrained on large scale datasets. While bias mitigation techniques have been\ndeveloped for tuning models for downstream tasks, it is currently unclear what\nare the effects of biases already encoded in a pretrained model. Our framework\nincorporates sets of canonical images representing individual and pairs of\nconcepts to highlight changes in biases for an array of off-the-shelf\npretrained models across model sizes, dataset sizes, and training objectives.\nThrough our analyses, we find that (1) supervised models trained on datasets\nsuch as ImageNet-21k are more likely to retain their pretraining biases\nregardless of the target dataset compared to self-supervised models. We also\nfind that (2) models finetuned on larger scale datasets are more likely to\nintroduce new biased associations. Our results also suggest that (3) biases can\ntransfer to finetuned models and the finetuning objective and dataset can\nimpact the extent of transferred biases.\n", "date": "2023-03-14", "categories": []}
{"id": "2302.02404", "title": "The Unfairness of Fair Machine Learning: Levelling down and strict\n  egalitarianism by default", "abstract": "  In recent years fairness in machine learning (ML) has emerged as a highly\nactive area of research and development. Most define fairness in simple terms,\nwhere fairness means reducing gaps in performance or outcomes between\ndemographic groups while preserving as much of the accuracy of the original\nsystem as possible. This oversimplification of equality through fairness\nmeasures is troubling. Many current fairness measures suffer from both fairness\nand performance degradation, or \"levelling down,\" where fairness is achieved by\nmaking every group worse off, or by bringing better performing groups down to\nthe level of the worst off. When fairness can only be achieved by making\neveryone worse off in material or relational terms through injuries of stigma,\nloss of solidarity, unequal concern, and missed opportunities for substantive\nequality, something would appear to have gone wrong in translating the vague\nconcept of 'fairness' into practice. This paper examines the causes and\nprevalence of levelling down across fairML, and explore possible justifications\nand criticisms based on philosophical and legal theories of equality and\ndistributive justice, as well as equality law jurisprudence. We find that\nfairML does not currently engage in the type of measurement, reporting, or\nanalysis necessary to justify levelling down in practice. We propose a first\nstep towards substantive equality in fairML: \"levelling up\" systems by design\nthrough enforcement of minimum acceptable harm thresholds, or \"minimum rate\nconstraints,\" as fairness constraints. We likewise propose an alternative\nharms-based framework to counter the oversimplified egalitarian framing\ncurrently dominant in the field and push future discussion more towards\nsubstantive equality opportunities and away from strict egalitarianism by\ndefault. N.B. Shortened abstract, see paper for full abstract.\n", "date": "2023-02-05", "categories": []}
{"id": "2303.06376", "title": "Assessing gender fairness in EEG-based machine learning detection of\n  Parkinson's disease: A multi-center study", "abstract": "  As the number of automatic tools based on machine learning (ML) and\nresting-state electroencephalography (rs-EEG) for Parkinson's disease (PD)\ndetection keeps growing, the assessment of possible exacerbation of health\ndisparities by means of fairness and bias analysis becomes more relevant.\nProtected attributes, such as gender, play an important role in PD diagnosis\ndevelopment. However, analysis of sub-group populations stemming from different\ngenders is seldom taken into consideration in ML models' development or the\nperformance assessment for PD detection. In this work, we perform a systematic\nanalysis of the detection ability for gender sub-groups in a multi-center\nsetting of a previously developed ML algorithm based on power spectral density\n(PSD) features of rs-EEG. We find significant differences in the PD detection\nability for males and females at testing time (80.5% vs. 63.7% accuracy) and\nsignificantly higher activity for a set of parietal and frontal EEG channels\nand frequency sub-bands for PD and non-PD males that might explain the\ndifferences in the PD detection ability for the gender sub-groups.\n", "date": "2023-03-11", "categories": []}
{"id": "2303.07024", "title": "Addressing Biases in the Texts using an End-to-End Pipeline Approach", "abstract": "  The concept of fairness is gaining popularity in academia and industry.\nSocial media is especially vulnerable to media biases and toxic language and\ncomments. We propose a fair ML pipeline that takes a text as input and\ndetermines whether it contains biases and toxic content. Then, based on\npre-trained word embeddings, it suggests a set of new words by substituting the\nbi-ased words, the idea is to lessen the effects of those biases by replacing\nthem with alternative words. We compare our approach to existing fairness\nmodels to determine its effectiveness. The results show that our proposed\npipeline can de-tect, identify, and mitigate biases in social media data\n", "date": "2023-03-13", "categories": []}
{"id": "2303.07207", "title": "Bias, diversity, and challenges to fairness in classification and\n  automated text analysis. From libraries to AI and back", "abstract": "  Libraries are increasingly relying on computational methods, including\nmethods from Artificial Intelligence (AI). This increasing usage raises\nconcerns about the risks of AI that are currently broadly discussed in\nscientific literature, the media and law-making. In this article we investigate\nthe risks surrounding bias and unfairness in AI usage in classification and\nautomated text analysis within the context of library applications. We describe\nexamples that show how the library community has been aware of such risks for a\nlong time, and how it has developed and deployed countermeasures. We take a\ncloser look at the notion of '(un)fairness' in relation to the notion of\n'diversity', and we investigate a formalisation of diversity that models both\ninclusion and distribution. We argue that many of the unfairness problems of\nautomated content analysis can also be regarded through the lens of diversity\nand the countermeasures taken to enhance diversity.\n", "date": "2023-03-07", "categories": []}
{"id": "2303.05698", "title": "Fairness-enhancing deep learning for ride-hailing demand prediction", "abstract": "  Short-term demand forecasting for on-demand ride-hailing services is one of\nthe fundamental issues in intelligent transportation systems. However, previous\ntravel demand forecasting research predominantly focused on improving\nprediction accuracy, ignoring fairness issues such as systematic\nunderestimations of travel demand in disadvantaged neighborhoods. This study\ninvestigates how to measure, evaluate, and enhance prediction fairness between\ndisadvantaged and privileged communities in spatial-temporal demand forecasting\nof ride-hailing services. A two-pronged approach is taken to reduce the demand\nprediction bias. First, we develop a novel deep learning model architecture,\nnamed socially aware neural network (SA-Net), to integrate the\nsocio-demographics and ridership information for fair demand prediction through\nan innovative socially-aware convolution operation. Second, we propose a\nbias-mitigation regularization method to mitigate the mean percentage\nprediction error gap between different groups. The experimental results,\nvalidated on the real-world Chicago Transportation Network Company (TNC) data,\nshow that the de-biasing SA-Net can achieve better predictive performance in\nboth prediction accuracy and fairness. Specifically, the SA-Net improves\nprediction accuracy for both the disadvantaged and privileged groups compared\nwith the state-of-the-art models. When coupled with the bias mitigation\nregularization method, the de-biasing SA-Net effectively bridges the mean\npercentage prediction error gap between the disadvantaged and privileged\ngroups, and also protects the disadvantaged regions against systematic\nunderestimation of TNC demand. Our proposed de-biasing method can be adopted in\nmany existing short-term travel demand estimation models, and can be utilized\nfor various other spatial-temporal prediction tasks such as crime incidents\npredictions.\n", "date": "2023-03-10", "categories": []}
{"id": "2302.01546", "title": "Group Fairness in Non-monotone Submodular Maximization", "abstract": "  Maximizing a submodular function has a wide range of applications in machine\nlearning and data mining. One such application is data summarization whose goal\nis to select a small set of representative and diverse data items from a large\ndataset. However, data items might have sensitive attributes such as race or\ngender, in this setting, it is important to design \\emph{fairness-aware}\nalgorithms to mitigate potential algorithmic bias that may cause over- or\nunder- representation of particular groups. Motivated by that, we propose and\nstudy the classic non-monotone submodular maximization problem subject to novel\ngroup fairness constraints. Our goal is to select a set of items that maximizes\na non-monotone submodular function, while ensuring that the number of selected\nitems from each group is proportionate to its size, to the extent specified by\nthe decision maker. We develop the first constant-factor approximation\nalgorithms for this problem. We also extend the basic model to incorporate an\nadditional global size constraint on the total number of selected items.\n", "date": "2023-02-03", "categories": []}
{"id": "2303.04838", "title": "The Casual Conversations v2 Dataset", "abstract": "  This paper introduces a new large consent-driven dataset aimed at assisting\nin the evaluation of algorithmic bias and robustness of computer vision and\naudio speech models in regards to 11 attributes that are self-provided or\nlabeled by trained annotators. The dataset includes 26,467 videos of 5,567\nunique paid participants, with an average of almost 5 videos per person,\nrecorded in Brazil, India, Indonesia, Mexico, Vietnam, Philippines, and the\nUSA, representing diverse demographic characteristics. The participants agreed\nfor their data to be used in assessing fairness of AI models and provided\nself-reported age, gender, language/dialect, disability status, physical\nadornments, physical attributes and geo-location information, while trained\nannotators labeled apparent skin tone using the Fitzpatrick Skin Type and Monk\nSkin Tone scales, and voice timbre. Annotators also labeled for different\nrecording setups and per-second activity annotations.\n", "date": "2023-03-08", "categories": []}
{"id": "2303.04896", "title": "Using Positive Matching Contrastive Loss with Facial Action Units to\n  mitigate bias in Facial Expression Recognition", "abstract": "  Machine learning models automatically learn discriminative features from the\ndata, and are therefore susceptible to learn strongly-correlated biases, such\nas using protected attributes like gender and race. Most existing bias\nmitigation approaches aim to explicitly reduce the model's focus on these\nprotected features. In this work, we propose to mitigate bias by explicitly\nguiding the model's focus towards task-relevant features using domain\nknowledge, and we hypothesize that this can indirectly reduce the dependence of\nthe model on spurious correlations it learns from the data. We explore bias\nmitigation in facial expression recognition systems using facial Action Units\n(AUs) as the task-relevant feature. To this end, we introduce Feature-based\nPositive Matching Contrastive Loss which learns the distances between the\npositives of a sample based on the similarity between their corresponding AU\nembeddings. We compare our approach with representative baselines and show that\nincorporating task-relevant features via our method can improve model fairness\nat minimal cost to classification performance.\n", "date": "2023-03-08", "categories": []}
{"id": "2303.04379", "title": "HappyMap: A Generalized Multi-calibration Method", "abstract": "  Multi-calibration is a powerful and evolving concept originating in the field\nof algorithmic fairness. For a predictor $f$ that estimates the outcome $y$\ngiven covariates $x$, and for a function class $\\mathcal{C}$, multi-calibration\nrequires that the predictor $f(x)$ and outcome $y$ are indistinguishable under\nthe class of auditors in $\\mathcal{C}$. Fairness is captured by incorporating\ndemographic subgroups into the class of functions~$\\mathcal{C}$. Recent work\nhas shown that, by enriching the class $\\mathcal{C}$ to incorporate appropriate\npropensity re-weighting functions, multi-calibration also yields\ntarget-independent learning, wherein a model trained on a source domain\nperforms well on unseen, future, target domains(approximately) captured by the\nre-weightings.\n  Formally, multi-calibration with respect to $\\mathcal{C}$ bounds\n$\\big|\\mathbb{E}_{(x,y)\\sim \\mathcal{D}}[c(f(x),x)\\cdot(f(x)-y)]\\big|$ for all\n$c \\in \\mathcal{C}$. In this work, we view the term $(f(x)-y)$ as just one\nspecific mapping, and explore the power of an enriched class of mappings. We\npropose \\textit{HappyMap}, a generalization of multi-calibration, which yields\na wide range of new applications, including a new fairness notion for\nuncertainty quantification (conformal prediction), a novel technique for\nconformal prediction under covariate shift, and a different approach to\nanalyzing missing data, while also yielding a unified understanding of several\nexisting seemingly disparate algorithmic fairness notions and\ntarget-independent learning approaches.\n  We give a single \\textit{HappyMap} meta-algorithm that captures all these\nresults, together with a sufficiency condition for its success.\n", "date": "2023-03-08", "categories": []}
{"id": "2302.07372", "title": "Same Same, But Different: Conditional Multi-Task Learning for\n  Demographic-Specific Toxicity Detection", "abstract": "  Algorithmic bias often arises as a result of differential subgroup validity,\nin which predictive relationships vary across groups. For example, in toxic\nlanguage detection, comments targeting different demographic groups can vary\nmarkedly across groups. In such settings, trained models can be dominated by\nthe relationships that best fit the majority group, leading to disparate\nperformance. We propose framing toxicity detection as multi-task learning\n(MTL), allowing a model to specialize on the relationships that are relevant to\neach demographic group while also leveraging shared properties across groups.\nWith toxicity detection, each task corresponds to identifying toxicity against\na particular demographic group. However, traditional MTL requires labels for\nall tasks to be present for every data point. To address this, we propose\nConditional MTL (CondMTL), wherein only training examples relevant to the given\ndemographic group are considered by the loss function. This lets us learn group\nspecific representations in each branch which are not cross contaminated by\nirrelevant labels. Results on synthetic and real data show that using CondMTL\nimproves predictive recall over various baselines in general and for the\nminority demographic group in particular, while having similar overall\naccuracy.\n", "date": "2023-02-14", "categories": []}
{"id": "2206.09895", "title": "Multiple Fairness and Cardinality constraints for Students-Topics\n  Grouping Problem", "abstract": "  Group work is a prevalent activity in educational settings, where students\nare often divided into topic-specific groups based on their preferences. The\ngrouping should reflect the students' aspirations as much as possible. Usually,\nthe resulting groups should also be balanced in terms of protected attributes\nlike gender or race since studies indicate that students might learn better in\na diverse group. Moreover, balancing the group cardinalities is also an\nessential requirement for fair workload distribution across the groups. In this\npaper, we introduce the multi-fair capacitated (MFC) grouping problem that\nfairly partitions students into non-overlapping groups while ensuring balanced\ngroup cardinalities (with a lower bound and an upper bound), and maximizing the\ndiversity of members in terms of protected attributes. We propose two\napproaches: a heuristic method and a knapsack-based method to obtain the MFC\ngrouping. The experiments on a real dataset and a semi-synthetic dataset show\nthat our proposed methods can satisfy students' preferences well and deliver\nbalanced and diverse groups regarding cardinality and the protected attribute,\nrespectively.\n", "date": "2022-06-20", "categories": []}
{"id": "2208.03843", "title": "Counterfactual Fairness Is Basically Demographic Parity", "abstract": "  Making fair decisions is crucial to ethically implementing machine learning\nalgorithms in social settings. In this work, we consider the celebrated\ndefinition of counterfactual fairness [Kusner et al., NeurIPS, 2017]. We begin\nby showing that an algorithm which satisfies counterfactual fairness also\nsatisfies demographic parity, a far simpler fairness constraint. Similarly, we\nshow that all algorithms satisfying demographic parity can be trivially\nmodified to satisfy counterfactual fairness. Together, our results indicate\nthat counterfactual fairness is basically equivalent to demographic parity,\nwhich has important implications for the growing body of work on counterfactual\nfairness. We then validate our theoretical findings empirically, analyzing\nthree existing algorithms for counterfactual fairness against three simple\nbenchmarks. We find that two simple benchmark algorithms outperform all three\nexisting algorithms -- in terms of fairness, accuracy, and efficiency -- on\nseveral data sets. Our analysis leads us to formalize a concrete fairness goal:\nto preserve the order of individuals within protected groups. We believe\ntransparency around the ordering of individuals within protected groups makes\nfair algorithms more trustworthy. By design, the two simple benchmark\nalgorithms satisfy this goal while the existing algorithms for counterfactual\nfairness do not.\n", "date": "2022-08-07", "categories": []}
{"id": "2303.02318", "title": "Achieving Counterfactual Fairness for Anomaly Detection", "abstract": "  Ensuring fairness in anomaly detection models has received much attention\nrecently as many anomaly detection applications involve human beings. However,\nexisting fair anomaly detection approaches mainly focus on association-based\nfairness notions. In this work, we target counterfactual fairness, which is a\nprevalent causation-based fairness notion. The goal of counterfactually fair\nanomaly detection is to ensure that the detection outcome of an individual in\nthe factual world is the same as that in the counterfactual world where the\nindividual had belonged to a different group. To this end, we propose a\ncounterfactually fair anomaly detection (CFAD) framework which consists of two\nphases, counterfactual data generation and fair anomaly detection. Experimental\nresults on a synthetic dataset and two real datasets show that CFAD can\neffectively detect anomalies as well as ensure counterfactual fairness.\n", "date": "2023-03-04", "categories": []}
{"id": "2303.02874", "title": "Adversarial Sampling for Fairness Testing in Deep Neural Network", "abstract": "  In this research, we focus on the usage of adversarial sampling to test for\nthe fairness in the prediction of deep neural network model across different\nclasses of image in a given dataset. While several framework had been proposed\nto ensure robustness of machine learning model against adversarial attack, some\nof which includes adversarial training algorithm. There is still the pitfall\nthat adversarial training algorithm tends to cause disparity in accuracy and\nrobustness among different group. Our research is aimed at using adversarial\nsampling to test for fairness in the prediction of deep neural network model\nacross different classes or categories of image in a given dataset. We\nsuccessfully demonstrated a new method of ensuring fairness across various\ngroup of input in deep neural network classifier. We trained our neural network\nmodel on the original image, and without training our model on the perturbed or\nattacked image. When we feed the adversarial samplings to our model, it was\nable to predict the original category/ class of the image the adversarial\nsample belongs to. We also introduced and used the separation of concern\nconcept from software engineering whereby there is an additional standalone\nfilter layer that filters perturbed image by heavily removing the noise or\nattack before automatically passing it to the network for classification, we\nwere able to have accuracy of 93.3%. Cifar-10 dataset have ten categories of\ndataset, and so, in order to account for fairness, we applied our hypothesis\nacross each categories of dataset and were able to get a consistent result and\naccuracy.\n", "date": "2023-03-06", "categories": []}
{"id": "2303.03242", "title": "Evaluating the Fairness of Deep Learning Uncertainty Estimates in\n  Medical Image Analysis", "abstract": "  Although deep learning (DL) models have shown great success in many medical\nimage analysis tasks, deployment of the resulting models into real clinical\ncontexts requires: (1) that they exhibit robustness and fairness across\ndifferent sub-populations, and (2) that the confidence in DL model predictions\nbe accurately expressed in the form of uncertainties. Unfortunately, recent\nstudies have indeed shown significant biases in DL models across demographic\nsubgroups (e.g., race, sex, age) in the context of medical image analysis,\nindicating a lack of fairness in the models. Although several methods have been\nproposed in the ML literature to mitigate a lack of fairness in DL models, they\nfocus entirely on the absolute performance between groups without considering\ntheir effect on uncertainty estimation. In this work, we present the first\nexploration of the effect of popular fairness models on overcoming biases\nacross subgroups in medical image analysis in terms of bottom-line performance,\nand their effects on uncertainty quantification. We perform extensive\nexperiments on three different clinically relevant tasks: (i) skin lesion\nclassification, (ii) brain tumour segmentation, and (iii) Alzheimer's disease\nclinical score regression. Our results indicate that popular ML methods, such\nas data-balancing and distributionally robust optimization, succeed in\nmitigating fairness issues in terms of the model performances for some of the\ntasks. However, this can come at the cost of poor uncertainty estimates\nassociated with the model predictions. This tradeoff must be mitigated if\nfairness models are to be adopted in medical image analysis.\n", "date": "2023-03-06", "categories": []}
{"id": "2209.07850", "title": "FairGBM: Gradient Boosting with Fairness Constraints", "abstract": "  Tabular data is prevalent in many high-stakes domains, such as financial\nservices or public policy. Gradient Boosted Decision Trees (GBDT) are popular\nin these settings due to their scalability, performance, and low training cost.\nWhile fairness in these domains is a foremost concern, existing in-processing\nFair ML methods are either incompatible with GBDT, or incur in significant\nperformance losses while taking considerably longer to train. We present\nFairGBM, a dual ascent learning framework for training GBDT under fairness\nconstraints, with little to no impact on predictive performance when compared\nto unconstrained GBDT. Since observational fairness metrics are\nnon-differentiable, we propose smooth convex error rate proxies for common\nfairness criteria, enabling gradient-based optimization using a\n``proxy-Lagrangian'' formulation. Our implementation shows an order of\nmagnitude speedup in training time relative to related work, a pivotal aspect\nto foster the widespread adoption of FairGBM by real-world practitioners.\n", "date": "2022-09-16", "categories": []}
{"id": "2303.00442", "title": "Re-weighting Based Group Fairness Regularization via Classwise Robust\n  Optimization", "abstract": "  Many existing group fairness-aware training methods aim to achieve the group\nfairness by either re-weighting underrepresented groups based on certain rules\nor using weakly approximated surrogates for the fairness metrics in the\nobjective as regularization terms. Although each of the learning schemes has\nits own strength in terms of applicability or performance, respectively, it is\ndifficult for any method in the either category to be considered as a gold\nstandard since their successful performances are typically limited to specific\ncases. To that end, we propose a principled method, dubbed as \\ours, which\nunifies the two learning schemes by incorporating a well-justified group\nfairness metric into the training objective using a class wise distributionally\nrobust optimization (DRO) framework. We then develop an iterative optimization\nalgorithm that minimizes the resulting objective by automatically producing the\ncorrect re-weights for each group. Our experiments show that FairDRO is\nscalable and easily adaptable to diverse applications, and consistently\nachieves the state-of-the-art performance on several benchmark datasets in\nterms of the accuracy-fairness trade-off, compared to recent strong baselines.\n", "date": "2023-03-01", "categories": []}
{"id": "2303.00673", "title": "Fairness Evaluation in Text Classification: Machine Learning\n  Practitioner Perspectives of Individual and Group Fairness", "abstract": "  Mitigating algorithmic bias is a critical task in the development and\ndeployment of machine learning models. While several toolkits exist to aid\nmachine learning practitioners in addressing fairness issues, little is known\nabout the strategies practitioners employ to evaluate model fairness and what\nfactors influence their assessment, particularly in the context of text\nclassification. Two common approaches of evaluating the fairness of a model are\ngroup fairness and individual fairness. We run a study with Machine Learning\npractitioners (n=24) to understand the strategies used to evaluate models.\nMetrics presented to practitioners (group vs. individual fairness) impact which\nmodels they consider fair. Participants focused on risks associated with\nunderpredicting/overpredicting and model sensitivity relative to identity token\nmanipulations. We discover fairness assessment strategies involving personal\nexperiences or how users form groups of identity tokens to test model fairness.\nWe provide recommendations for interactive tools for evaluating fairness in\ntext classification.\n", "date": "2023-03-01", "categories": []}
{"id": "2302.12578", "title": "Fairness in Language Models Beyond English: Gaps and Challenges", "abstract": "  With language models becoming increasingly ubiquitous, it has become\nessential to address their inequitable treatment of diverse demographic groups\nand factors. Most research on evaluating and mitigating fairness harms has been\nconcentrated on English, while multilingual models and non-English languages\nhave received comparatively little attention. This paper presents a survey of\nfairness in multilingual and non-English contexts, highlighting the\nshortcomings of current research and the difficulties faced by methods designed\nfor English. We contend that the multitude of diverse cultures and languages\nacross the world makes it infeasible to achieve comprehensive coverage in terms\nof constructing fairness datasets. Thus, the measurement and mitigation of\nbiases must evolve beyond the current dataset-driven practices that are\nnarrowly focused on specific dimensions and types of biases and, therefore,\nimpossible to scale across languages and cultures.\n", "date": "2023-02-24", "categories": []}
{"id": "2302.14063", "title": "How optimal transport can tackle gender biases in multi-class\n  neural-network classifiers for job recommendations?", "abstract": "  Automatic recommendation systems based on deep neural networks have become\nextremely popular during the last decade. Some of these systems can however be\nused for applications which are ranked as High Risk by the European Commission\nin the A.I. act, as for instance for online job candidate recommendation. When\nused in the European Union, commercial AI systems for this purpose will then be\nrequired to have to proper statistical properties with regard to potential\ndiscrimination they could engender. This motivated our contribution, where we\npresent a novel optimal transport strategy to mitigate undesirable algorithmic\nbiases in multi-class neural-network classification. Our stratey is model\nagnostic and can be used on any multi-class classification neural-network\nmodel. To anticipate the certification of recommendation systems using textual\ndata, we then used it on the Bios dataset, for which the learning task consists\nin predicting the occupation of female and male individuals, based on their\nLinkedIn biography. Results show that it can reduce undesired algorithmic\nbiases in this context to lower levels than a standard strategy.\n", "date": "2023-02-27", "categories": []}
{"id": "2302.13136", "title": "Toward Fairness in Text Generation via Mutual Information Minimization\n  based on Importance Sampling", "abstract": "  Pretrained language models (PLMs), such as GPT2, have achieved remarkable\nempirical performance in text generation tasks. However, pretrained on\nlarge-scale natural language corpora, the generated text from PLMs may exhibit\nsocial bias against disadvantaged demographic groups. To improve the fairness\nof PLMs in text generation, we propose to minimize the mutual information\nbetween the semantics in the generated text sentences and their demographic\npolarity, i.e., the demographic group to which the sentence is referring. In\nthis way, the mentioning of a demographic group (e.g., male or female) is\nencouraged to be independent from how it is described in the generated text,\nthus effectively alleviating the social bias. Moreover, we propose to\nefficiently estimate the upper bound of the above mutual information via\nimportance sampling, leveraging a natural language corpus. We also propose a\ndistillation mechanism that preserves the language modeling ability of the PLMs\nafter debiasing. Empirical results on real-world benchmarks demonstrate that\nthe proposed method yields superior performance in term of both fairness and\nlanguage modeling ability.\n", "date": "2023-02-25", "categories": []}
{"id": "2302.13319", "title": "Efficient fair PCA for fair representation learning", "abstract": "  We revisit the problem of fair principal component analysis (PCA), where the\ngoal is to learn the best low-rank linear approximation of the data that\nobfuscates demographic information. We propose a conceptually simple approach\nthat allows for an analytic solution similar to standard PCA and can be\nkernelized. Our methods have the same complexity as standard PCA, or kernel\nPCA, and run much faster than existing methods for fair PCA based on\nsemidefinite programming or manifold optimization, while achieving similar\nresults.\n", "date": "2023-02-26", "categories": []}
{"id": "2302.12333", "title": "Auditing for Spatial Fairness", "abstract": "  This paper studies algorithmic fairness when the protected attribute is\nlocation. To handle protected attributes that are continuous, such as age or\nincome, the standard approach is to discretize the domain into predefined\ngroups, and compare algorithmic outcomes across groups. However, applying this\nidea to location raises concerns of gerrymandering and may introduce\nstatistical bias. Prior work addresses these concerns but only for regularly\nspaced locations, while raising other issues, most notably its inability to\ndiscern regions that are likely to exhibit spatial unfairness. Similar to\nestablished notions of algorithmic fairness, we define spatial fairness as the\nstatistical independence of outcomes from location. This translates into\nrequiring that for each region of space, the distribution of outcomes is\nidentical inside and outside the region. To allow for localized discrepancies\nin the distribution of outcomes, we compare how well two competing hypotheses\nexplain the observed outcomes. The null hypothesis assumes spatial fairness,\nwhile the alternate allows different distributions inside and outside regions.\nTheir goodness of fit is then assessed by a likelihood ratio test. If there is\nno significant difference in how well the two hypotheses explain the observed\noutcomes, we conclude that the algorithm is spatially fair.\n", "date": "2023-02-23", "categories": []}
{"id": "2302.12640", "title": "In-Depth Look at Word Filling Societal Bias Measures", "abstract": "  Many measures of societal bias in language models have been proposed in\nrecent years. A popular approach is to use a set of word filling prompts to\nevaluate the behavior of the language models. In this work, we analyze the\nvalidity of two such measures -- StereoSet and CrowS-Pairs. We show that these\nmeasures produce unexpected and illogical results when appropriate control\ngroup samples are constructed. Based on this, we believe that they are\nproblematic and using them in the future should be reconsidered. We propose a\nway forward with an improved testing protocol. Finally, we also introduce a new\ngender bias dataset for Slovak.\n", "date": "2023-02-24", "categories": []}
{"id": "2302.12683", "title": "Intersectional Fairness: A Fractal Approach", "abstract": "  The issue of fairness in AI has received an increasing amount of attention in\nrecent years. The problem can be approached by looking at different protected\nattributes (e.g., ethnicity, gender, etc) independently, but fairness for\nindividual protected attributes does not imply intersectional fairness. In this\nwork, we frame the problem of intersectional fairness within a geometrical\nsetting. We project our data onto a hypercube, and split the analysis of\nfairness by levels, where each level encodes the number of protected attributes\nwe are intersecting over. We prove mathematically that, while fairness does not\npropagate \"down\" the levels, it does propagate \"up\" the levels. This means that\nensuring fairness for all subgroups at the lowest intersectional level (e.g.,\nblack women, white women, black men and white men), will necessarily result in\nfairness for all the above levels, including each of the protected attributes\n(e.g., ethnicity and gender) taken independently. We also derive a formula\ndescribing the variance of the set of estimated success rates on each level,\nunder the assumption of perfect fairness. Using this theoretical finding as a\nbenchmark, we define a family of metrics which capture overall intersectional\nbias. Finally, we propose that fairness can be metaphorically thought of as a\n\"fractal\" problem. In fractals, patterns at the smallest scale repeat at a\nlarger scale. We see from this example that tackling the problem at the lowest\npossible level, in a bottom-up manner, leads to the natural emergence of fair\nAI. We suggest that trustworthiness is necessarily an emergent, fractal and\nrelational property of the AI system.\n", "date": "2023-02-24", "categories": []}
{"id": "2212.12603", "title": "Stochastic Methods for AUC Optimization subject to AUC-based Fairness\n  Constraints", "abstract": "  As machine learning being used increasingly in making high-stakes decisions,\nan arising challenge is to avoid unfair AI systems that lead to discriminatory\ndecisions for protected population. A direct approach for obtaining a fair\npredictive model is to train the model through optimizing its prediction\nperformance subject to fairness constraints, which achieves Pareto efficiency\nwhen trading off performance against fairness. Among various fairness metrics,\nthe ones based on the area under the ROC curve (AUC) are emerging recently\nbecause they are threshold-agnostic and effective for unbalanced data. In this\nwork, we formulate the training problem of a fairness-aware machine learning\nmodel as an AUC optimization problem subject to a class of AUC-based fairness\nconstraints. This problem can be reformulated as a min-max optimization problem\nwith min-max constraints, which we solve by stochastic first-order methods\nbased on a new Bregman divergence designed for the special structure of the\nproblem. We numerically demonstrate the effectiveness of our approach on\nreal-world data under different fairness metrics.\n", "date": "2022-12-23", "categories": []}
{"id": "2302.08851", "title": "On (assessing) the fairness of risk score models", "abstract": "  Recent work on algorithmic fairness has largely focused on the fairness of\ndiscrete decisions, or classifications. While such decisions are often based on\nrisk score models, the fairness of the risk models themselves has received\nconsiderably less attention. Risk models are of interest for a number of\nreasons, including the fact that they communicate uncertainty about the\npotential outcomes to users, thus representing a way to enable meaningful human\noversight. Here, we address fairness desiderata for risk score models. We\nidentify the provision of similar epistemic value to different groups as a key\ndesideratum for risk score fairness. Further, we address how to assess the\nfairness of risk score models quantitatively, including a discussion of metric\nchoices and meaningful statistical comparisons between groups. In this context,\nwe also introduce a novel calibration error metric that is less sample\nsize-biased than previously proposed metrics, enabling meaningful comparisons\nbetween groups of different sizes. We illustrate our methodology - which is\nwidely applicable in many other settings - in two case studies, one in\nrecidivism risk prediction, and one in risk of major depressive disorder (MDD)\nprediction.\n", "date": "2023-02-17", "categories": []}
{"id": "2302.09683", "title": "SimFair: A Unified Framework for Fairness-Aware Multi-Label\n  Classification", "abstract": "  Recent years have witnessed increasing concerns towards unfair decisions made\nby machine learning algorithms. To improve fairness in model decisions, various\nfairness notions have been proposed and many fairness-aware methods are\ndeveloped. However, most of existing definitions and methods focus only on\nsingle-label classification. Fairness for multi-label classification, where\neach instance is associated with more than one labels, is still yet to\nestablish. To fill this gap, we study fairness-aware multi-label classification\nin this paper. We start by extending Demographic Parity (DP) and Equalized\nOpportunity (EOp), two popular fairness notions, to multi-label classification\nscenarios. Through a systematic study, we show that on multi-label data,\nbecause of unevenly distributed labels, EOp usually fails to construct a\nreliable estimate on labels with few instances. We then propose a new framework\nnamed Similarity $s$-induced Fairness ($s_\\gamma$-SimFair). This new framework\nutilizes data that have similar labels when estimating fairness on a particular\nlabel group for better stability, and can unify DP and EOp. Theoretical\nanalysis and experimental results on real-world datasets together demonstrate\nthe advantage of over existing methods $s_\\gamma$-SimFair on multi-label\nclassification tasks.\n", "date": "2023-02-19", "categories": []}
{"id": "2302.11295", "title": "Fair Correlation Clustering in Forests", "abstract": "  The study of algorithmic fairness received growing attention recently. This\nstems from the awareness that bias in the input data for machine learning\nsystems may result in discriminatory outputs. For clustering tasks, one of the\nmost central notions of fairness is the formalization by Chierichetti, Kumar,\nLattanzi, and Vassilvitskii [NeurIPS 2017]. A clustering is said to be fair, if\neach cluster has the same distribution of manifestations of a sensitive\nattribute as the whole input set. This is motivated by various applications\nwhere the objects to be clustered have sensitive attributes that should not be\nover- or underrepresented.\n  We discuss the applicability of this fairness notion to Correlation\nClustering. The existing literature on the resulting Fair Correlation\nClustering problem either presents approximation algorithms with poor\napproximation guarantees or severely limits the possible distributions of the\nsensitive attribute (often only two manifestations with a 1:1 ratio are\nconsidered). Our goal is to understand if there is hope for better results in\nbetween these two extremes. To this end, we consider restricted graph classes\nwhich allow us to characterize the distributions of sensitive attributes for\nwhich this form of fairness is tractable from a complexity point of view.\n  While existing work on Fair Correlation Clustering gives approximation\nalgorithms, we focus on exact solutions and investigate whether there are\nefficiently solvable instances. The unfair version of Correlation Clustering is\ntrivial on forests, but adding fairness creates a surprisingly rich picture of\ncomplexities. We give an overview of the distributions and types of forests\nwhere Fair Correlation Clustering turns from tractable to intractable. The most\nsurprising insight to us is the fact that the cause of the hardness of Fair\nCorrelation Clustering is not the strictness of the fairness condition.\n", "date": "2023-02-22", "categories": []}
{"id": "2302.11479", "title": "Drop Edges and Adapt: a Fairness Enforcing Fine-tuning for Graph Neural\n  Networks", "abstract": "  The rise of graph representation learning as the primary solution for many\ndifferent network science tasks led to a surge of interest in the fairness of\nthis family of methods. Link prediction, in particular, has a substantial\nsocial impact. However, link prediction algorithms tend to increase the\nsegregation in social networks by disfavoring the links between individuals in\nspecific demographic groups. This paper proposes a novel way to enforce\nfairness on graph neural networks with a fine-tuning strategy. We Drop the\nunfair Edges and, simultaneously, we Adapt the model's parameters to those\nmodifications, DEA in short. We introduce two covariance-based constraints\ndesigned explicitly for the link prediction task. We use these constraints to\nguide the optimization process responsible for learning the new \"fair\"\nadjacency matrix. One novelty of DEA is that we can use a discrete yet\nlearnable adjacency matrix in our fine-tuning. We demonstrate the effectiveness\nof our approach on five real-world datasets and show that we can improve both\nthe accuracy and the fairness of the link prediction tasks. In addition, we\npresent an in-depth ablation study demonstrating that our training algorithm\nfor the adjacency matrix can be used to improve link prediction performances\nduring training. Finally, we compute the relevance of each component of our\nframework to show that the combination of both the constraints and the training\nof the adjacency matrix leads to optimal performances.\n", "date": "2023-02-22", "categories": []}
{"id": "2210.01953", "title": "Robust Fair Clustering: A Novel Fairness Attack and Defense Framework", "abstract": "  Clustering algorithms are widely used in many societal resource allocation\napplications, such as loan approvals and candidate recruitment, among others,\nand hence, biased or unfair model outputs can adversely impact individuals that\nrely on these applications. To this end, many fair clustering approaches have\nbeen recently proposed to counteract this issue. Due to the potential for\nsignificant harm, it is essential to ensure that fair clustering algorithms\nprovide consistently fair outputs even under adversarial influence. However,\nfair clustering algorithms have not been studied from an adversarial attack\nperspective. In contrast to previous research, we seek to bridge this gap and\nconduct a robustness analysis against fair clustering by proposing a novel\nblack-box fairness attack. Through comprehensive experiments, we find that\nstate-of-the-art models are highly susceptible to our attack as it can reduce\ntheir fairness performance significantly. Finally, we propose Consensus Fair\nClustering (CFC), the first robust fair clustering approach that transforms\nconsensus clustering into a fair graph partitioning problem, and iteratively\nlearns to generate fair cluster outputs. Experimentally, we observe that CFC is\nhighly robust to the proposed attack and is thus a truly robust fair clustering\nalternative.\n", "date": "2022-10-04", "categories": []}
{"id": "2302.06333", "title": "Improving Recommendation Fairness via Data Augmentation", "abstract": "  Collaborative filtering based recommendation learns users' preferences from\nall users' historical behavior data, and has been popular to facilitate\ndecision making. R Recently, the fairness issue of recommendation has become\nmore and more essential. A recommender system is considered unfair when it does\nnot perform equally well for different user groups according to users'\nsensitive attributes~(e.g., gender, race). Plenty of methods have been proposed\nto alleviate unfairness by optimizing a predefined fairness goal or changing\nthe distribution of unbalanced training data. However, they either suffered\nfrom the specific fairness optimization metrics or relied on redesigning the\ncurrent recommendation architecture. In this paper, we study how to improve\nrecommendation fairness from the data augmentation perspective. The\nrecommendation model amplifies the inherent unfairness of imbalanced training\ndata. We augment imbalanced training data towards balanced data distribution to\nimprove fairness. The proposed framework is generally applicable to any\nembedding-based recommendation, and does not need to pre-define a fairness\nmetric. Extensive experiments on two real-world datasets clearly demonstrate\nthe superiority of our proposed framework. We publish the source code at\nhttps://github.com/newlei/FDA.\n", "date": "2023-02-13", "categories": []}
{"id": "2302.09298", "title": "Improving Fairness in Adaptive Social Exergames via Shapley Bandits", "abstract": "  Algorithmic fairness is an essential requirement as AI becomes integrated in\nsociety. In the case of social applications where AI distributes resources,\nalgorithms often must make decisions that will benefit a subset of users,\nsometimes repeatedly or exclusively, while attempting to maximize specific\noutcomes. How should we design such systems to serve users more fairly? This\npaper explores this question in the case where a group of users works toward a\nshared goal in a social exergame called Step Heroes. We identify adverse\noutcomes in traditional multi-armed bandits (MABs) and formalize the Greedy\nBandit Problem. We then propose a solution based on a new type of\nfairness-aware multi-armed bandit, Shapley Bandits. It uses the Shapley Value\nfor increasing overall player participation and intervention adherence rather\nthan the maximization of total group output, which is traditionally achieved by\nfavoring only high-performing participants. We evaluate our approach via a user\nstudy (n=46). Our results indicate that our Shapley Bandits effectively\nmediates the Greedy Bandit Problem and achieves better user retention and\nmotivation across the participants.\n", "date": "2023-02-18", "categories": []}
{"id": "2302.10409", "title": "Mean Parity Fair Regression in RKHS", "abstract": "  We study the fair regression problem under the notion of Mean Parity (MP)\nfairness, which requires the conditional mean of the learned function output to\nbe constant with respect to the sensitive attributes. We address this problem\nby leveraging reproducing kernel Hilbert space (RKHS) to construct the\nfunctional space whose members are guaranteed to satisfy the fairness\nconstraints. The proposed functional space suggests a closed-form solution for\nthe fair regression problem that is naturally compatible with multiple\nsensitive attributes. Furthermore, by formulating the fairness-accuracy\ntradeoff as a relaxed fair regression problem, we derive a corresponding\nregression function that can be implemented efficiently and provides\ninterpretable tradeoffs. More importantly, under some mild assumptions, the\nproposed method can be applied to regression problems with a covariance-based\nnotion of fairness. Experimental results on benchmark datasets show the\nproposed methods achieve competitive and even superior performance compared\nwith several state-of-the-art methods.\n", "date": "2023-02-21", "categories": []}
{"id": "2302.10575", "title": "Managing multi-facet bias in collaborative filtering recommender systems", "abstract": "  Due to the extensive growth of information available online, recommender\nsystems play a more significant role in serving people's interests. Traditional\nrecommender systems mostly use an accuracy-focused approach to produce\nrecommendations. Today's research suggests that this single-dimension approach\ncan lead the system to be biased against a series of items with certain\nattributes. Biased recommendations across groups of items can endanger the\ninterests of item providers along with causing user dissatisfaction with the\nsystem. This study aims to manage a new type of intersectional bias regarding\nthe geographical origin and popularity of items in the output of\nstate-of-the-art collaborative filtering recommender algorithms. We introduce\nan algorithm called MFAIR, a multi-facet post-processing bias mitigation\nalgorithm to alleviate these biases. Extensive experiments on two real-world\ndatasets of movies and books, enriched with the items' continents of\nproduction, show that the proposed algorithm strikes a reasonable balance\nbetween accuracy and both types of the mentioned biases. According to the\nresults, our proposed approach outperforms a well-known competitor with no or\nonly a slight loss of efficiency.\n", "date": "2023-02-21", "categories": []}
{"id": "2209.00617", "title": "Fair mapping", "abstract": "  To mitigate the effects of undesired biases in models, several approaches\npropose to pre-process the input dataset to reduce the risks of discrimination\nby preventing the inference of sensitive attributes. Unfortunately, most of\nthese pre-processing methods lead to the generation a new distribution that is\nvery different from the original one, thus often leading to unrealistic data.\nAs a side effect, this new data distribution implies that existing models need\nto be re-trained to be able to make accurate predictions. To address this\nissue, we propose a novel pre-processing method, that we coin as fair mapping,\nbased on the transformation of the distribution of protected groups onto a\nchosen target one, with additional privacy constraints whose objective is to\nprevent the inference of sensitive attributes. More precisely, we leverage on\nthe recent works of the Wasserstein GAN and AttGAN frameworks to achieve the\noptimal transport of data points coupled with a discriminator enforcing the\nprotection against attribute inference. Our proposed approach, preserves the\ninterpretability of data and can be used without defining exactly the sensitive\ngroups. In addition, our approach can be specialized to model existing\nstate-of-the-art approaches, thus proposing a unifying view on these methods.\nFinally, several experiments on real and synthetic datasets demonstrate that\nour approach is able to hide the sensitive attributes, while limiting the\ndistortion of the data and improving the fairness on subsequent data analysis\ntasks.\n", "date": "2022-09-01", "categories": []}
{"id": "2210.01725", "title": "MEDFAIR: Benchmarking Fairness for Medical Imaging", "abstract": "  A multitude of work has shown that machine learning-based medical diagnosis\nsystems can be biased against certain subgroups of people. This has motivated a\ngrowing number of bias mitigation algorithms that aim to address fairness\nissues in machine learning. However, it is difficult to compare their\neffectiveness in medical imaging for two reasons. First, there is little\nconsensus on the criteria to assess fairness. Second, existing bias mitigation\nalgorithms are developed under different settings, e.g., datasets, model\nselection strategies, backbones, and fairness metrics, making a direct\ncomparison and evaluation based on existing results impossible. In this work,\nwe introduce MEDFAIR, a framework to benchmark the fairness of machine learning\nmodels for medical imaging. MEDFAIR covers eleven algorithms from various\ncategories, nine datasets from different imaging modalities, and three model\nselection criteria. Through extensive experiments, we find that the\nunder-studied issue of model selection criterion can have a significant impact\non fairness outcomes; while in contrast, state-of-the-art bias mitigation\nalgorithms do not significantly improve fairness outcomes over empirical risk\nminimization (ERM) in both in-distribution and out-of-distribution settings. We\nevaluate fairness from various perspectives and make recommendations for\ndifferent medical application scenarios that require different ethical\nprinciples. Our framework provides a reproducible and easy-to-use entry point\nfor the development and evaluation of future bias mitigation algorithms in deep\nlearning. Code is available at https://github.com/ys-zong/MEDFAIR.\n", "date": "2022-10-04", "categories": []}
{"id": "2302.03881", "title": "On Generalized Degree Fairness in Graph Neural Networks", "abstract": "  Conventional graph neural networks (GNNs) are often confronted with fairness\nissues that may stem from their input, including node attributes and neighbors\nsurrounding a node. While several recent approaches have been proposed to\neliminate the bias rooted in sensitive attributes, they ignore the other key\ninput of GNNs, namely the neighbors of a node, which can introduce bias since\nGNNs hinge on neighborhood structures to generate node representations. In\nparticular, the varying neighborhood structures across nodes, manifesting\nthemselves in drastically different node degrees, give rise to the diverse\nbehaviors of nodes and biased outcomes. In this paper, we first define and\ngeneralize the degree bias using a generalized definition of node degree as a\nmanifestation and quantification of different multi-hop structures around\ndifferent nodes. To address the bias in the context of node classification, we\npropose a novel GNN framework called Generalized Degree Fairness-centric Graph\nNeural Network (Deg-FairGNN). Specifically, in each GNN layer, we employ a\nlearnable debiasing function to generate debiasing contexts, which modulate the\nlayer-wise neighborhood aggregation to eliminate the degree bias originating\nfrom the diverse degrees among nodes. Extensive experiments on three benchmark\ndatasets demonstrate the effectiveness of our model on both accuracy and\nfairness metrics.\n", "date": "2023-02-08", "categories": []}
{"id": "2302.09157", "title": "Designing Equitable Algorithms", "abstract": "  Predictive algorithms are now used to help distribute a large share of our\nsociety's resources and sanctions, such as healthcare, loans, criminal\ndetentions, and tax audits. Under the right circumstances, these algorithms can\nimprove the efficiency and equity of decision-making. At the same time, there\nis a danger that the algorithms themselves could entrench and exacerbate\ndisparities, particularly along racial, ethnic, and gender lines. To help\nensure their fairness, many researchers suggest that algorithms be subject to\nat least one of three constraints: (1) no use of legally protected features,\nsuch as race, ethnicity, and gender; (2) equal rates of \"positive\" decisions\nacross groups; and (3) equal error rates across groups. Here we show that these\nconstraints, while intuitively appealing, often worsen outcomes for individuals\nin marginalized groups, and can even leave all groups worse off. The inherent\ntrade-off we identify between formal fairness constraints and welfare\nimprovements -- particularly for the marginalized -- highlights the need for a\nmore robust discussion on what it means for an algorithm to be \"fair\". We\nillustrate these ideas with examples from healthcare and the criminal-legal\nsystem, and make several proposals to help practitioners design more equitable\nalgorithms.\n", "date": "2023-02-17", "categories": []}
{"id": "2302.09183", "title": "Learning with Impartiality to Walk on the Pareto Frontier of Fairness,\n  Privacy, and Utility", "abstract": "  Deploying machine learning (ML) models often requires both fairness and\nprivacy guarantees. Both of these objectives present unique trade-offs with the\nutility (e.g., accuracy) of the model. However, the mutual interactions between\nfairness, privacy, and utility are less well-understood. As a result, often\nonly one objective is optimized, while the others are tuned as\nhyper-parameters. Because they implicitly prioritize certain objectives, such\ndesigns bias the model in pernicious, undetectable ways. To address this, we\nadopt impartiality as a principle: design of ML pipelines should not favor one\nobjective over another. We propose impartially-specified models, which provide\nus with accurate Pareto frontiers that show the inherent trade-offs between the\nobjectives. Extending two canonical ML frameworks for privacy-preserving\nlearning, we provide two methods (FairDP-SGD and FairPATE) to train\nimpartially-specified models and recover the Pareto frontier. Through\ntheoretical privacy analysis and a comprehensive empirical study, we provide an\nanswer to the question of where fairness mitigation should be integrated within\na privacy-aware ML pipeline.\n", "date": "2023-02-17", "categories": []}
{"id": "2302.09400", "title": "Fairly Predicting Graft Failure in Liver Transplant for Organ Assigning", "abstract": "  Liver transplant is an essential therapy performed for severe liver diseases.\nThe fact of scarce liver resources makes the organ assigning crucial. Model for\nEnd-stage Liver Disease (MELD) score is a widely adopted criterion when making\norgan distribution decisions. However, it ignores post-transplant outcomes and\norgan/donor features. These limitations motivate the emergence of machine\nlearning (ML) models. Unfortunately, ML models could be unfair and trigger bias\nagainst certain groups of people. To tackle this problem, this work proposes a\nfair machine learning framework targeting graft failure prediction in liver\ntransplant. Specifically, knowledge distillation is employed to handle dense\nand sparse features by combining the advantages of tree models and neural\nnetworks. A two-step debiasing method is tailored for this framework to enhance\nfairness. Experiments are conducted to analyze unfairness issues in existing\nmodels and demonstrate the superiority of our method in both prediction and\nfairness performance.\n", "date": "2023-02-18", "categories": []}
{"id": "2302.09991", "title": "Towards Measuring and Scoring Speaker Diarization Fairness", "abstract": "  Speaker diarization, or the task of finding \"who spoke and when\", is now used\nin almost every speech processing application. Nevertheless, its fairness has\nnot yet been evaluated because there was no protocol to study its biases one by\none. In this paper we propose a protocol and a scoring method designed to\nevaluate speaker diarization fairness. This protocol is applied on a large\ndataset of spoken utterances and report the performances of speaker diarization\ndepending on the gender, the age, the accent of the speaker and the length of\nthe spoken sentence. Some biases induced by the gender, or the accent of the\nspeaker were identified when we applied a state-of-the-art speaker diarization\nmethod.\n", "date": "2023-02-20", "categories": []}
{"id": "2302.08572", "title": "Towards Reliable Assessments of Demographic Disparities in Multi-Label\n  Image Classifiers", "abstract": "  Disaggregated performance metrics across demographic groups are a hallmark of\nfairness assessments in computer vision. These metrics successfully\nincentivized performance improvements on person-centric tasks such as face\nanalysis and are used to understand risks of modern models. However, there is a\nlack of discussion on the vulnerabilities of these measurements for more\ncomplex computer vision tasks. In this paper, we consider multi-label image\nclassification and, specifically, object categorization tasks. First, we\nhighlight design choices and trade-offs for measurement that involve more\nnuance than discussed in prior computer vision literature. These challenges are\nrelated to the necessary scale of data, definition of groups for images, choice\nof metric, and dataset imbalances. Next, through two case studies using modern\nvision models, we demonstrate that naive implementations of these assessments\nare brittle. We identify several design choices that look merely like\nimplementation details but significantly impact the conclusions of assessments,\nboth in terms of magnitude and direction (on which group the classifiers work\nbest) of disparities. Based on ablation studies, we propose some\nrecommendations to increase the reliability of these assessments. Finally,\nthrough a qualitative analysis we find that concepts with large disparities\ntend to have varying definitions and representations between groups, with\ninconsistencies across datasets and annotators. While this result suggests\navenues for mitigation through more consistent data collection, it also\nhighlights that ambiguous label definitions remain a challenge when performing\nmodel assessments. Vision models are expanding and becoming more ubiquitous; it\nis even more important that our disparity assessments accurately reflect the\ntrue performance of models.\n", "date": "2023-02-16", "categories": []}
{"id": "2302.08704", "title": "The Unbearable Weight of Massive Privilege: Revisiting Bias-Variance\n  Trade-Offs in the Context of Fair Prediction", "abstract": "  In this paper we revisit the bias-variance decomposition of model error from\nthe perspective of designing a fair classifier: we are motivated by the widely\nheld socio-technical belief that noise variance in large datasets in social\ndomains tracks demographic characteristics such as gender, race, disability,\netc. We propose a conditional-iid (ciid) model built from group-specific\nclassifiers that seeks to improve on the trade-offs made by a single model (iid\nsetting). We theoretically analyze the bias-variance decomposition of different\nmodels in the Gaussian Mixture Model, and then empirically test our setup on\nthe COMPAS and folktables datasets. We instantiate the ciid model with two\nprocedures that improve \"fairness\" by conditioning out undesirable effects:\nfirst, by conditioning directly on sensitive attributes, and second, by\nclustering samples into groups and conditioning on cluster membership (blind to\nprotected group membership).\n  Our analysis suggests that there might be principled procedures and concrete\nreal-world use cases under which conditional models are preferred, and our\nstriking empirical results strongly indicate that non-iid settings, such as the\nciid setting proposed here, might be more suitable for big data applications in\nsocial contexts.\n", "date": "2023-02-17", "categories": []}
{"id": "2209.07463", "title": "Omnipredictors for Constrained Optimization", "abstract": "  The notion of omnipredictors (Gopalan, Kalai, Reingold, Sharan and Wieder\nITCS 2021), suggested a new paradigm for loss minimization. Rather than\nlearning a predictor based on a known loss function, omnipredictors can easily\nbe post-processed to minimize any one of a rich family of loss functions\ncompared with the loss of hypotheses in a class $\\mathcal C$. It has been shown\nthat such omnipredictors exist and are implied (for all convex and Lipschitz\nloss functions) by the notion of multicalibration from the algorithmic fairness\nliterature. In this paper, we introduce omnipredictors for constrained\noptimization and study their complexity and implications. The notion that we\nintroduce allows the learner to be unaware of the loss function that will be\nlater assigned as well as the constraints that will be later imposed, as long\nas the subpopulations that are used to define these constraints are known. We\nshow how to obtain omnipredictors for constrained optimization problems,\nrelying on appropriate variants of multicalibration. We also investigate the\nimplications of this notion when the constraints used are so-called group\nfairness notions.\n", "date": "2022-09-15", "categories": []}
{"id": "2210.07269", "title": "SODAPOP: Open-Ended Discovery of Social Biases in Social Commonsense\n  Reasoning Models", "abstract": "  A common limitation of diagnostic tests for detecting social biases in NLP\nmodels is that they may only detect stereotypic associations that are\npre-specified by the designer of the test. Since enumerating all possible\nproblematic associations is infeasible, it is likely these tests fail to detect\nbiases that are present in a model but not pre-specified by the designer. To\naddress this limitation, we propose SODAPOP (SOcial bias Discovery from Answers\nabout PeOPle) in social commonsense question-answering. Our pipeline generates\nmodified instances from the Social IQa dataset (Sap et al., 2019) by (1)\nsubstituting names associated with different demographic groups, and (2)\ngenerating many distractor answers from a masked language model. By using a\nsocial commonsense model to score the generated distractors, we are able to\nuncover the model's stereotypic associations between demographic groups and an\nopen set of words. We also test SODAPOP on debiased models and show the\nlimitations of multiple state-of-the-art debiasing algorithms.\n", "date": "2022-10-13", "categories": []}
{"id": "2302.08017", "title": "Preventing Discriminatory Decision-making in Evolving Data Streams", "abstract": "  Bias in machine learning has rightly received significant attention over the\nlast decade. However, most fair machine learning (fair-ML) work to address bias\nin decision-making systems has focused solely on the offline setting. Despite\nthe wide prevalence of online systems in the real world, work on identifying\nand correcting bias in the online setting is severely lacking. The unique\nchallenges of the online environment make addressing bias more difficult than\nin the offline setting. First, Streaming Machine Learning (SML) algorithms must\ndeal with the constantly evolving real-time data stream. Second, they need to\nadapt to changing data distributions (concept drift) to make accurate\npredictions on new incoming data. Adding fairness constraints to this already\ncomplicated task is not straightforward. In this work, we focus on the\nchallenges of achieving fairness in biased data streams while accounting for\nthe presence of concept drift, accessing one sample at a time. We present Fair\nSampling over Stream ($FS^2$), a novel fair rebalancing approach capable of\nbeing integrated with SML classification algorithms. Furthermore, we devise the\nfirst unified performance-fairness metric, Fairness Bonded Utility (FBU), to\nevaluate and compare the trade-off between performance and fairness of\ndifferent bias mitigation methods efficiently. FBU simplifies the comparison of\nfairness-performance trade-offs of multiple techniques through one unified and\nintuitive evaluation, allowing model designers to easily choose a technique.\nOverall, extensive evaluations show our measures surpass those of other fair\nonline techniques previously reported in the literature.\n", "date": "2023-02-16", "categories": []}
{"id": "2302.08158", "title": "Counterfactual Fair Opportunity: Measuring Decision Model Fairness with\n  Counterfactual Reasoning", "abstract": "  The increasing application of Artificial Intelligence and Machine Learning\nmodels poses potential risks of unfair behavior and, in light of recent\nregulations, has attracted the attention of the research community. Several\nresearchers focused on seeking new fairness definitions or developing\napproaches to identify biased predictions. However, none try to exploit the\ncounterfactual space to this aim. In that direction, the methodology proposed\nin this work aims to unveil unfair model behaviors using counterfactual\nreasoning in the case of fairness under unawareness setting. A counterfactual\nversion of equal opportunity named counterfactual fair opportunity is defined\nand two novel metrics that analyze the sensitive information of counterfactual\nsamples are introduced. Experimental results on three different datasets show\nthe efficacy of our methodologies and our metrics, disclosing the unfair\nbehavior of classic machine learning and debiasing models.\n", "date": "2023-02-16", "categories": []}
{"id": "2302.02463", "title": "Nationality Bias in Text Generation", "abstract": "  Little attention is placed on analyzing nationality bias in language models,\nespecially when nationality is highly used as a factor in increasing the\nperformance of social NLP models. This paper examines how a text generation\nmodel, GPT-2, accentuates pre-existing societal biases about country-based\ndemonyms. We generate stories using GPT-2 for various nationalities and use\nsensitivity analysis to explore how the number of internet users and the\ncountry's economic status impacts the sentiment of the stories. To reduce the\npropagation of biases through large language models (LLM), we explore the\ndebiasing method of adversarial triggering. Our results show that GPT-2\ndemonstrates significant bias against countries with lower internet users, and\nadversarial triggering effectively reduces the same.\n", "date": "2023-02-05", "categories": []}
{"id": "2206.14983", "title": "A Validity Perspective on Evaluating the Justified Use of Data-driven\n  Decision-making Algorithms", "abstract": "  Recent research increasingly brings to question the appropriateness of using\npredictive tools in complex, real-world tasks. While a growing body of work has\nexplored ways to improve value alignment in these tools, comparatively less\nwork has centered concerns around the fundamental justifiability of using these\ntools. This work seeks to center validity considerations in deliberations\naround whether and how to build data-driven algorithms in high-stakes domains.\nToward this end, we translate key concepts from validity theory to predictive\nalgorithms. We apply the lens of validity to re-examine common challenges in\nproblem formulation and data issues that jeopardize the justifiability of using\npredictive algorithms and connect these challenges to the social science\ndiscourse around validity. Our interdisciplinary exposition clarifies how these\nconcepts apply to algorithmic decision making contexts. We demonstrate how\nthese validity considerations could distill into a series of high-level\nquestions intended to promote and document reflections on the legitimacy of the\npredictive task and the suitability of the data.\n", "date": "2022-06-30", "categories": []}
{"id": "2302.06752", "title": "Provable Detection of Propagating Sampling Bias in Prediction Models", "abstract": "  With an increased focus on incorporating fairness in machine learning models,\nit becomes imperative not only to assess and mitigate bias at each stage of the\nmachine learning pipeline but also to understand the downstream impacts of bias\nacross stages. Here we consider a general, but realistic, scenario in which a\npredictive model is learned from (potentially biased) training data, and model\npredictions are assessed post-hoc for fairness by some auditing method. We\nprovide a theoretical analysis of how a specific form of data bias,\ndifferential sampling bias, propagates from the data stage to the prediction\nstage. Unlike prior work, we evaluate the downstream impacts of data biases\nquantitatively rather than qualitatively and prove theoretical guarantees for\ndetection. Under reasonable assumptions, we quantify how the amount of bias in\nthe model predictions varies as a function of the amount of differential\nsampling bias in the data, and at what point this bias becomes provably\ndetectable by the auditor. Through experiments on two criminal justice datasets\n-- the well-known COMPAS dataset and historical data from NYPD's stop and frisk\npolicy -- we demonstrate that the theoretical results hold in practice even\nwhen our assumptions are relaxed.\n", "date": "2023-02-13", "categories": []}
{"id": "2302.07159", "title": "A Friendly Face: Do Text-to-Image Systems Rely on Stereotypes when the\n  Input is Under-Specified?", "abstract": "  As text-to-image systems continue to grow in popularity with the general\npublic, questions have arisen about bias and diversity in the generated images.\nHere, we investigate properties of images generated in response to prompts\nwhich are visually under-specified, but contain salient social attributes\n(e.g., 'a portrait of a threatening person' versus 'a portrait of a friendly\nperson'). Grounding our work in social cognition theory, we find that in many\ncases, images contain similar demographic biases to those reported in the\nstereotype literature. However, trends are inconsistent across different models\nand further investigation is warranted.\n", "date": "2023-02-14", "categories": []}
{"id": "2211.14498", "title": "The Impact of Racial Distribution in Training Data on Face Recognition\n  Bias: A Closer Look", "abstract": "  Face recognition algorithms, when used in the real world, can be very useful,\nbut they can also be dangerous when biased toward certain demographics. So, it\nis essential to understand how these algorithms are trained and what factors\naffect their accuracy and fairness to build better ones. In this study, we shed\nsome light on the effect of racial distribution in the training data on the\nperformance of face recognition models. We conduct 16 different experiments\nwith varying racial distributions of faces in the training data. We analyze\nthese trained models using accuracy metrics, clustering metrics, UMAP\nprojections, face quality, and decision thresholds. We show that a uniform\ndistribution of races in the training datasets alone does not guarantee\nbias-free face recognition algorithms and how factors like face image quality\nplay a crucial role. We also study the correlation between the clustering\nmetrics and bias to understand whether clustering is a good indicator of bias.\nFinally, we introduce a metric called racial gradation to study the inter and\nintra race correlation in facial features and how they affect the learning\nability of the face recognition models. With this study, we try to bring more\nunderstanding to an essential element of face recognition training, the data. A\nbetter understanding of the impact of training data on the bias of face\nrecognition algorithms will aid in creating better datasets and, in turn,\nbetter face recognition systems.\n", "date": "2022-11-26", "categories": []}
{"id": "2302.03465", "title": "Robustness Implies Fairness in Causal Algorithmic Recourse", "abstract": "  Algorithmic recourse aims to disclose the inner workings of the black-box\ndecision process in situations where decisions have significant consequences,\nby providing recommendations to empower beneficiaries to achieve a more\nfavorable outcome. To ensure an effective remedy, suggested interventions must\nnot only be low-cost but also robust and fair. This goal is accomplished by\nproviding similar explanations to individuals who are alike. This study\nexplores the concept of individual fairness and adversarial robustness in\ncausal algorithmic recourse and addresses the challenge of achieving both. To\nresolve the challenges, we propose a new framework for defining adversarially\nrobust recourse. The new setting views the protected feature as a pseudometric\nand demonstrates that individual fairness is a special case of adversarial\nrobustness. Finally, we introduce the fair robust recourse problem to achieve\nboth desirable properties and show how it can be satisfied both theoretically\nand empirically.\n", "date": "2023-02-07", "categories": []}
{"id": "2302.05711", "title": "Fair Enough: Standardizing Evaluation and Model Selection for Fairness\n  Research in NLP", "abstract": "  Modern NLP systems exhibit a range of biases, which a growing literature on\nmodel debiasing attempts to correct. However current progress is hampered by a\nplurality of definitions of bias, means of quantification, and oftentimes vague\nrelation between debiasing algorithms and theoretical measures of bias. This\npaper seeks to clarify the current situation and plot a course for meaningful\nprogress in fair learning, with two key contributions: (1) making clear\ninter-relations among the current gamut of methods, and their relation to\nfairness theory; and (2) addressing the practical problem of model selection,\nwhich involves a trade-off between fairness and accuracy and has led to\nsystemic issues in fairness research. Putting them together, we make several\nrecommendations to help shape future work.\n", "date": "2023-02-11", "categories": []}
{"id": "2302.05788", "title": "Fairness-aware Multi-view Clustering", "abstract": "  In the era of big data, we are often facing the challenge of data\nheterogeneity and the lack of label information simultaneously. In the\nfinancial domain (e.g., fraud detection), the heterogeneous data may include\nnot only numerical data (e.g., total debt and yearly income), but also text and\nimages (e.g., financial statement and invoice images). At the same time, the\nlabel information (e.g., fraud transactions) may be missing for building\npredictive models. To address these challenges, many state-of-the-art\nmulti-view clustering methods have been proposed and achieved outstanding\nperformance. However, these methods typically do not take into consideration\nthe fairness aspect and are likely to generate biased results using sensitive\ninformation such as race and gender. Therefore, in this paper, we propose a\nfairness-aware multi-view clustering method named FairMVC. It incorporates the\ngroup fairness constraint into the soft membership assignment for each cluster\nto ensure that the fraction of different groups in each cluster is\napproximately identical to the entire data set. Meanwhile, we adopt the idea of\nboth contrastive learning and non-contrastive learning and propose novel\nregularizers to handle heterogeneous data in complex scenarios with missing\ndata or noisy features. Experimental results on real-world data sets\ndemonstrate the effectiveness and efficiency of the proposed framework. We also\nderive insights regarding the relative performance of the proposed regularizers\nin various scenarios.\n", "date": "2023-02-11", "categories": []}
{"id": "2302.05995", "title": "Multi-dimensional discrimination in Law and Machine Learning -- A\n  comparative overview", "abstract": "  AI-driven decision-making can lead to discrimination against certain\nindividuals or social groups based on protected characteristics/attributes such\nas race, gender, or age. The domain of fairness-aware machine learning focuses\non methods and algorithms for understanding, mitigating, and accounting for\nbias in AI/ML models. Still, thus far, the vast majority of the proposed\nmethods assess fairness based on a single protected attribute, e.g. only gender\nor race. In reality, though, human identities are multi-dimensional, and\ndiscrimination can occur based on more than one protected characteristic,\nleading to the so-called ``multi-dimensional discrimination'' or\n``multi-dimensional fairness'' problem. While well-elaborated in legal\nliterature, the multi-dimensionality of discrimination is less explored in the\nmachine learning community. Recent approaches in this direction mainly follow\nthe so-called intersectional fairness definition from the legal domain, whereas\nother notions like additive and sequential discrimination are less studied or\nnot considered thus far. In this work, we overview the different definitions of\nmulti-dimensional discrimination/fairness in the legal domain as well as how\nthey have been transferred/ operationalized (if) in the fairness-aware machine\nlearning domain. By juxtaposing these two domains, we draw the connections,\nidentify the limitations, and point out open research directions.\n", "date": "2023-02-12", "categories": []}
{"id": "2302.06347", "title": "The Possibility of Fairness: Revisiting the Impossibility Theorem in\n  Practice", "abstract": "  The ``impossibility theorem'' -- which is considered foundational in\nalgorithmic fairness literature -- asserts that there must be trade-offs\nbetween common notions of fairness and performance when fitting statistical\nmodels, except in two special cases: when the prevalence of the outcome being\npredicted is equal across groups, or when a perfectly accurate predictor is\nused. However, theory does not always translate to practice. In this work, we\nchallenge the implications of the impossibility theorem in practical settings.\nFirst, we show analytically that, by slightly relaxing the impossibility\ntheorem (to accommodate a \\textit{practitioner's} perspective of fairness), it\nbecomes possible to identify a large set of models that satisfy seemingly\nincompatible fairness constraints. Second, we demonstrate the existence of\nthese models through extensive experiments on five real-world datasets. We\nconclude by offering tools and guidance for practitioners to understand when --\nand to what degree -- fairness along multiple criteria can be achieved. For\nexample, if one allows only a small margin-of-error between metrics, there\nexists a large set of models simultaneously satisfying \\emph{False Negative\nRate Parity}, \\emph{False Positive Rate Parity}, and \\emph{Positive Predictive\nValue Parity}, even when there is a moderate prevalence difference between\ngroups. This work has an important implication for the community: achieving\nfairness along multiple metrics for multiple groups (and their intersections)\nis much more possible than was previously believed.\n", "date": "2023-02-13", "categories": []}
{"id": "2207.03277", "title": "A Comprehensive Empirical Study of Bias Mitigation Methods for Machine\n  Learning Classifiers", "abstract": "  Software bias is an increasingly important operational concern for software\nengineers. We present a large-scale, comprehensive empirical study of 17\nrepresentative bias mitigation methods for Machine Learning (ML) classifiers,\nevaluated with 11 ML performance metrics (e.g., accuracy), 4 fairness metrics,\nand 20 types of fairness-performance trade-off assessment, applied to 8\nwidely-adopted software decision tasks. The empirical coverage is much more\ncomprehensive, covering the largest numbers of bias mitigation methods,\nevaluation metrics, and fairness-performance trade-off measures compared to\nprevious work on this important software property. We find that (1) the bias\nmitigation methods significantly decrease ML performance in 53% of the studied\nscenarios (ranging between 42%~66% according to different ML performance\nmetrics); (2) the bias mitigation methods significantly improve fairness\nmeasured by the 4 used metrics in 46% of all the scenarios (ranging between\n24%~59% according to different fairness metrics); (3) the bias mitigation\nmethods even lead to decrease in both fairness and ML performance in 25% of the\nscenarios; (4) the effectiveness of the bias mitigation methods depends on\ntasks, models, the choice of protected attributes, and the set of metrics used\nto assess fairness and ML performance; (5) there is no bias mitigation method\nthat can achieve the best trade-off in all the scenarios. The best method that\nwe find outperforms other methods in 30% of the scenarios. Researchers and\npractitioners need to choose the bias mitigation method best suited to their\nintended application scenario(s).\n", "date": "2022-07-07", "categories": []}
{"id": "2208.01157", "title": "Global Performance Disparities Between English-Language Accents in\n  Automatic Speech Recognition", "abstract": "  Past research has identified discriminatory automatic speech recognition\n(ASR) performance as a function of the racial group and nationality of the\nspeaker. In this paper, we expand the discussion beyond bias as a function of\nthe individual national origin of the speaker to look for bias as a function of\nthe geopolitical orientation of their nation of origin. We audit some of the\nmost popular English language ASR services using a large and global data set of\nspeech from The Speech Accent Archive, which includes over 2,700 speakers of\nEnglish born in 171 different countries. We show that, even when controlling\nfor multiple linguistic covariates, ASR service performance has a statistically\nsignificant relationship to the political alignment of the speaker's birth\ncountry with respect to the United States' geopolitical power. This holds for\nall ASR services tested. We discuss this bias in the context of the historical\nuse of language to maintain global and political power.\n", "date": "2022-08-01", "categories": []}
{"id": "2302.04358", "title": "Mitigating Bias in Visual Transformers via Targeted Alignment", "abstract": "  As transformer architectures become increasingly prevalent in computer\nvision, it is critical to understand their fairness implications. We perform\nthe first study of the fairness of transformers applied to computer vision and\nbenchmark several bias mitigation approaches from prior work. We visualize the\nfeature space of the transformer self-attention modules and discover that a\nsignificant portion of the bias is encoded in the query matrix. With this\nknowledge, we propose TADeT, a targeted alignment strategy for debiasing\ntransformers that aims to discover and remove bias primarily from query matrix\nfeatures. We measure performance using Balanced Accuracy and Standard Accuracy,\nand fairness using Equalized Odds and Balanced Accuracy Difference. TADeT\nconsistently leads to improved fairness over prior work on multiple attribute\nprediction tasks on the CelebA dataset, without compromising performance.\n", "date": "2023-02-08", "categories": []}
{"id": "2209.05889", "title": "Investigating Bias with a Synthetic Data Generator: Empirical Evidence\n  and Philosophical Interpretation", "abstract": "  Machine learning applications are becoming increasingly pervasive in our\nsociety. Since these decision-making systems rely on data-driven learning, risk\nis that they will systematically spread the bias embedded in data. In this\npaper, we propose to analyze biases by introducing a framework for generating\nsynthetic data with specific types of bias and their combinations. We delve\ninto the nature of these biases discussing their relationship to moral and\njustice frameworks. Finally, we exploit our proposed synthetic data generator\nto perform experiments on different scenarios, with various bias combinations.\nWe thus analyze the impact of biases on performance and fairness metrics both\nin non-mitigated and mitigated machine learning models.\n", "date": "2022-09-13", "categories": []}
{"id": "2302.03151", "title": "Fair Minimum Representation Clustering", "abstract": "  Clustering is an unsupervised learning task that aims to partition data into\na set of clusters. In many applications, these clusters correspond to\nreal-world constructs (e.g. electoral districts) whose benefit can only be\nattained by groups when they reach a minimum level of representation (e.g. 50\\%\nto elect their desired candidate). This paper considers the problem of\nperforming k-means clustering while ensuring groups (e.g. demographic groups)\nhave that minimum level of representation in a specified number of clusters. We\nshow that the popular $k$-means algorithm, Lloyd's algorithm, can result in\nunfair outcomes where certain groups lack sufficient representation past the\nminimum threshold in a proportional number of clusters. We formulate the\nproblem through a mixed-integer optimization framework and present a variant of\nLloyd's algorithm, called MiniReL, that directly incorporates the fairness\nconstraints. We show that incorporating the fairness criteria leads to a\nNP-Hard sub-problem within Lloyd's algorithm, but we provide computational\napproaches that make the problem tractable for even large datasets. Numerical\nresults show that the approach is able to create fairer clusters with\npractically no increase in the k-means clustering cost across standard\nbenchmark datasets.\n", "date": "2023-02-06", "categories": []}
{"id": "2302.04119", "title": "Local Law 144: A Critical Analysis of Regression Metrics", "abstract": "  The use of automated decision tools in recruitment has received an increasing\namount of attention. In November 2021, the New York City Council passed a\nlegislation (Local Law 144) that mandates bias audits of Automated Employment\nDecision Tools. From 15th April 2023, companies that use automated tools for\nhiring or promoting employees are required to have these systems audited by an\nindependent entity. Auditors are asked to compute bias metrics that compare\noutcomes for different groups, based on sex/gender and race/ethnicity\ncategories at a minimum. Local Law 144 proposes novel bias metrics for\nregression tasks (scenarios where the automated system scores candidates with a\ncontinuous range of values). A previous version of the legislation proposed a\nbias metric that compared the mean scores of different groups. The new revised\nbias metric compares the proportion of candidates in each group that falls\nabove the median. In this paper, we argue that both metrics fail to capture\ndistributional differences over the whole domain, and therefore cannot reliably\ndetect bias. We first introduce two metrics, as possible alternatives to the\nlegislation metrics. We then compare these metrics over a range of theoretical\nexamples, for which the legislation proposed metrics seem to underestimate\nbias. Finally, we study real data and show that the legislation metrics can\nsimilarly fail in a real-world recruitment application.\n", "date": "2023-02-08", "categories": []}
{"id": "2302.03567", "title": "From Utilitarian to Rawlsian Designs for Algorithmic Fairness", "abstract": "  There is a lack of consensus within the literature as to how `fairness' of\nalgorithmic systems can be measured, and different metrics can often be at\nodds. In this paper, we approach this task by drawing on the ethical frameworks\nof utilitarianism and John Rawls. Informally, these two theories of\ndistributive justice measure the `good' as either a population's sum of\nutility, or worst-off outcomes, respectively. We present a parameterized class\nof objective functions that interpolates between these two (possibly)\nconflicting notions of the `good'. This class is shown to represent a\nrelaxation of the Rawlsian `veil of ignorance', and its sequence of optimal\nsolutions converges to both a utilitarian and Rawlsian optimum. Several other\nproperties of this class are studied, including: 1) a relationship to\nregularized optimization, 2) feasibility of consistent estimation, and 3)\nalgorithmic cost. In several real-world datasets, we compute optimal solutions\nand construct the tradeoff between utilitarian and Rawlsian notions of the\n`good'. Empirically, we demonstrate that increasing model complexity can\nmanifest strict improvements to both measures of the `good'. This work suggests\nthat the proper degree of `fairness' can be informed by a designer's\npreferences over the space of induced utilitarian and Rawlsian `good'.\n", "date": "2023-02-07", "categories": []}
{"id": "2302.02038", "title": "Rating Sentiment Analysis Systems for Bias through a Causal Lens", "abstract": "  Sentiment Analysis Systems (SASs) are data-driven Artificial Intelligence\n(AI) systems that, given a piece of text, assign one or more numbers conveying\nthe polarity and emotional intensity expressed in the input. Like other\nautomatic machine learning systems, they have also been known to exhibit model\nuncertainty where a (small) change in the input leads to drastic swings in the\noutput. This can be especially problematic when inputs are related to protected\nfeatures like gender or race since such behavior can be perceived as a lack of\nfairness, i.e., bias. We introduce a novel method to assess and rate SASs where\ninputs are perturbed in a controlled causal setting to test if the output\nsentiment is sensitive to protected variables even when other components of the\ntextual input, e.g., chosen emotion words, are fixed. We then use the result to\nassign labels (ratings) at fine-grained and overall levels to convey the\nrobustness of the SAS to input changes. The ratings serve as a principled basis\nto compare SASs and choose among them based on behavior. It benefits all users,\nespecially developers who reuse off-the-shelf SASs to build larger AI systems\nbut do not have access to their code or training data to compare.\n", "date": "2023-02-04", "categories": []}
{"id": "2302.02306", "title": "Fair Spatial Indexing: A paradigm for Group Spatial Fairness", "abstract": "  Machine learning (ML) is playing an increasing role in decision-making tasks\nthat directly affect individuals, e.g., loan approvals, or job applicant\nscreening. Significant concerns arise that, without special provisions,\nindividuals from under-privileged backgrounds may not get equitable access to\nservices and opportunities. Existing research studies fairness with respect to\nprotected attributes such as gender, race or income, but the impact of location\ndata on fairness has been largely overlooked. With the widespread adoption of\nmobile apps, geospatial attributes are increasingly used in ML, and their\npotential to introduce unfair bias is significant, given their high correlation\nwith protected attributes. We propose techniques to mitigate location bias in\nmachine learning. Specifically, we consider the issue of miscalibration when\ndealing with geospatial attributes. We focus on spatial group fairness and we\npropose a spatial indexing algorithm that accounts for fairness. Our KD-tree\ninspired approach significantly improves fairness while maintaining high\nlearning accuracy, as shown by extensive experimental results on real data.\n", "date": "2023-02-05", "categories": []}
{"id": "2302.02323", "title": "Improving Fair Training under Correlation Shifts", "abstract": "  Model fairness is an essential element for Trustworthy AI. While many\ntechniques for model fairness have been proposed, most of them assume that the\ntraining and deployment data distributions are identical, which is often not\ntrue in practice. In particular, when the bias between labels and sensitive\ngroups changes, the fairness of the trained model is directly influenced and\ncan worsen. We make two contributions for solving this problem. First, we\nanalytically show that existing in-processing fair algorithms have fundamental\nlimits in accuracy and group fairness. We introduce the notion of correlation\nshifts, which can explicitly capture the change of the above bias. Second, we\npropose a novel pre-processing step that samples the input data to reduce\ncorrelation shifts and thus enables the in-processing approaches to overcome\ntheir limitations. We formulate an optimization problem for adjusting the data\nratio among labels and sensitive groups to reflect the shifted correlation. A\nkey benefit of our approach lies in decoupling the roles of pre- and\nin-processing approaches: correlation adjustment via pre-processing and\nunfairness mitigation on the processed data via in-processing. Experiments show\nthat our framework effectively improves existing in-processing fair algorithms\nw.r.t. accuracy and fairness, both on synthetic and real datasets.\n", "date": "2023-02-05", "categories": []}
{"id": "2302.02453", "title": "FineDeb: A Debiasing Framework for Language Models", "abstract": "  As language models are increasingly included in human-facing machine learning\ntools, bias against demographic subgroups has gained attention. We propose\nFineDeb, a two-phase debiasing framework for language models that starts with\ncontextual debiasing of embeddings learned by pretrained language models. The\nmodel is then fine-tuned on a language modeling objective. Our results show\nthat FineDeb offers stronger debiasing in comparison to other methods which\noften result in models as biased as the original language model. Our framework\nis generalizable for demographics with multiple classes, and we demonstrate its\neffectiveness through extensive experiments and comparisons with state of the\nart techniques. We release our code and data on GitHub.\n", "date": "2023-02-05", "categories": []}
{"id": "2302.01448", "title": "Out of Context: Investigating the Bias and Fairness Concerns of\n  \"Artificial Intelligence as a Service\"", "abstract": "  \"AI as a Service\" (AIaaS) is a rapidly growing market, offering various\nplug-and-play AI services and tools. AIaaS enables its customers (users) - who\nmay lack the expertise, data, and/or resources to develop their own systems -\nto easily build and integrate AI capabilities into their applications. Yet, it\nis known that AI systems can encapsulate biases and inequalities that can have\nsocietal impact. This paper argues that the context-sensitive nature of\nfairness is often incompatible with AIaaS' 'one-size-fits-all' approach,\nleading to issues and tensions. Specifically, we review and systematise the\nAIaaS space by proposing a taxonomy of AI services based on the levels of\nautonomy afforded to the user. We then critically examine the different\ncategories of AIaaS, outlining how these services can lead to biases or be\notherwise harmful in the context of end-user applications. In doing so, we seek\nto draw research attention to the challenges of this emerging area.\n", "date": "2023-02-02", "categories": []}
{"id": "2208.08881", "title": "Modelling the long-term fairness dynamics of data-driven targeted help\n  on job seekers", "abstract": "  The use of data-driven decision support by public agencies is becoming more\nwidespread and already influences the allocation of public resources. This\nraises ethical concerns, as it has adversely affected minorities and\nhistorically discriminated groups. In this paper, we use an approach that\ncombines statistics and data-driven approaches with dynamical modeling to\nassess long-term fairness effects of labor market interventions. Specifically,\nwe develop and use a model to investigate the impact of decisions caused by a\npublic employment authority that selectively supports job-seekers through\ntargeted help. The selection of who receives what help is based on a\ndata-driven intervention model that estimates an individual's chances of\nfinding a job in a timely manner and rests upon data that describes a\npopulation in which skills relevant to the labor market are unevenly\ndistributed between two groups (e.g., males and females). The intervention\nmodel has incomplete access to the individual's actual skills and can augment\nthis with knowledge of the individual's group affiliation, thus using a\nprotected attribute to increase predictive accuracy. We assess this\nintervention model's dynamics -- especially fairness-related issues and\ntrade-offs between different fairness goals -- over time and compare it to an\nintervention model that does not use group affiliation as a predictive feature.\nWe conclude that in order to quantify the trade-off correctly and to assess the\nlong-term fairness effects of such a system in the real-world, careful modeling\nof the surrounding labor market is indispensable.\n", "date": "2022-08-17", "categories": []}
{"id": "2302.01079", "title": "Uncertainty in Fairness Assessment: Maintaining Stable Conclusions\n  Despite Fluctuations", "abstract": "  Several recent works encourage the use of a Bayesian framework when assessing\nperformance and fairness metrics of a classification algorithm in a supervised\nsetting. We propose the Uncertainty Matters (UM) framework that generalizes a\nBeta-Binomial approach to derive the posterior distribution of any criteria\ncombination, allowing stable performance assessment in a bias-aware setting.We\nsuggest modeling the confusion matrix of each demographic group using a\nMultinomial distribution updated through a Bayesian procedure. We extend UM to\nbe applicable under the popular K-fold cross-validation procedure. Experiments\nhighlight the benefits of UM over classical evaluation frameworks regarding\ninformativeness and stability.\n", "date": "2023-02-02", "categories": []}
{"id": "2210.03175", "title": "Weak Proxies are Sufficient and Preferable for Fairness with Missing\n  Sensitive Attributes", "abstract": "  Evaluating fairness can be challenging in practice because the sensitive\nattributes of data are often inaccessible due to privacy constraints. The go-to\napproach that the industry frequently adopts is using off-the-shelf proxy\nmodels to predict the missing sensitive attributes, e.g. Meta [Alao et al.,\n2021] and Twitter [Belli et al., 2022]. Despite its popularity, there are three\nimportant questions unanswered: (1) Is directly using proxies efficacious in\nmeasuring fairness? (2) If not, is it possible to accurately evaluate fairness\nusing proxies only? (3) Given the ethical controversy over inferring user\nprivate information, is it possible to only use weak (i.e. inaccurate) proxies\nin order to protect privacy? Our theoretical analyses show that directly using\nproxy models can give a false sense of (un)fairness. Second, we develop an\nalgorithm that is able to measure fairness (provably) accurately with only\nthree properly identified proxies. Third, we show that our algorithm allows the\nuse of only weak proxies (e.g. with only 68.85%accuracy on COMPAS), adding an\nextra layer of protection on user privacy. Experiments validate our theoretical\nanalyses and show our algorithm can effectively measure and mitigate bias. Our\nresults imply a set of practical guidelines for practitioners on how to use\nproxies properly. Code is available at github.com/UCSC-REAL/fair-eval.\n", "date": "2022-10-06", "categories": []}
{"id": "2301.13323", "title": "Fairness and Accuracy under Domain Generalization", "abstract": "  As machine learning (ML) algorithms are increasingly used in high-stakes\napplications, concerns have arisen that they may be biased against certain\nsocial groups. Although many approaches have been proposed to make ML models\nfair, they typically rely on the assumption that data distributions in training\nand deployment are identical. Unfortunately, this is commonly violated in\npractice and a model that is fair during training may lead to an unexpected\noutcome during its deployment. Although the problem of designing robust ML\nmodels under dataset shifts has been widely studied, most existing works focus\nonly on the transfer of accuracy. In this paper, we study the transfer of both\nfairness and accuracy under domain generalization where the data at test time\nmay be sampled from never-before-seen domains. We first develop theoretical\nbounds on the unfairness and expected loss at deployment, and then derive\nsufficient conditions under which fairness and accuracy can be perfectly\ntransferred via invariant representation learning. Guided by this, we design a\nlearning algorithm such that fair ML models learned with training data still\nhave high fairness and accuracy when deployment environments change.\nExperiments on real-world data validate the proposed algorithm. Model\nimplementation is available at https://github.com/pth1993/FATDM.\n", "date": "2023-01-30", "categories": []}
{"id": "2301.13420", "title": "Superhuman Fairness", "abstract": "  The fairness of machine learning-based decisions has become an increasingly\nimportant focus in the design of supervised machine learning methods. Most\nfairness approaches optimize a specified trade-off between performance\nmeasure(s) (e.g., accuracy, log loss, or AUC) and fairness metric(s) (e.g.,\ndemographic parity, equalized odds). This begs the question: are the right\nperformance-fairness trade-offs being specified? We instead re-cast fair\nmachine learning as an imitation learning task by introducing superhuman\nfairness, which seeks to simultaneously outperform human decisions on multiple\npredictive performance and fairness measures. We demonstrate the benefits of\nthis approach given suboptimal decisions.\n", "date": "2023-01-31", "categories": []}
{"id": "2301.13767", "title": "Multicalibration as Boosting for Regression", "abstract": "  We study the connection between multicalibration and boosting for squared\nerror regression. First we prove a useful characterization of multicalibration\nin terms of a ``swap regret'' like condition on squared error. Using this\ncharacterization, we give an exceedingly simple algorithm that can be analyzed\nboth as a boosting algorithm for regression and as a multicalibration algorithm\nfor a class H that makes use only of a standard squared error regression oracle\nfor H. We give a weak learning assumption on H that ensures convergence to\nBayes optimality without the need to make any realizability assumptions --\ngiving us an agnostic boosting algorithm for regression. We then show that our\nweak learning assumption on H is both necessary and sufficient for\nmulticalibration with respect to H to imply Bayes optimality. We also show that\nif H satisfies our weak learning condition relative to another class C then\nmulticalibration with respect to H implies multicalibration with respect to C.\nFinally we investigate the empirical performance of our algorithm\nexperimentally using an open source implementation that we make available. Our\ncode repository can be found at\nhttps://github.com/Declancharrison/Level-Set-Boosting.\n", "date": "2023-01-31", "categories": []}
{"id": "2206.00553", "title": "FETA: Fairness Enforced Verifying, Training, and Predicting Algorithms\n  for Neural Networks", "abstract": "  Algorithmic decision making driven by neural networks has become very\nprominent in applications that directly affect people's quality of life. In\nthis paper, we study the problem of verifying, training, and guaranteeing\nindividual fairness of neural network models. A popular approach for enforcing\nfairness is to translate a fairness notion into constraints over the parameters\nof the model. However, such a translation does not always guarantee fair\npredictions of the trained neural network model. To address this challenge, we\ndevelop a counterexample-guided post-processing technique to provably enforce\nfairness constraints at prediction time. Contrary to prior work that enforces\nfairness only on points around test or train data, we are able to enforce and\nguarantee fairness on all points in the input domain. Additionally, we propose\nan in-processing technique to use fairness as an inductive bias by iteratively\nincorporating fairness counterexamples in the learning process. We have\nimplemented these techniques in a tool called FETA. Empirical evaluation on\nreal-world datasets indicates that FETA is not only able to guarantee fairness\non-the-fly at prediction time but also is able to train accurate models\nexhibiting a much higher degree of individual fairness.\n", "date": "2022-06-01", "categories": []}
{"id": "2208.10625", "title": "Evaluation of group fairness measures in student performance prediction\n  problems", "abstract": "  Predicting students' academic performance is one of the key tasks of\neducational data mining (EDM). Traditionally, the high forecasting quality of\nsuch models was deemed critical. More recently, the issues of fairness and\ndiscrimination w.r.t. protected attributes, such as gender or race, have gained\nattention. Although there are several fairness-aware learning approaches in\nEDM, a comparative evaluation of these measures is still missing. In this\npaper, we evaluate different group fairness measures for student performance\nprediction problems on various educational datasets and fairness-aware learning\nmodels. Our study shows that the choice of the fairness measure is important,\nlikewise for the choice of the grade threshold.\n", "date": "2022-08-22", "categories": []}
{"id": "2301.12278", "title": "Pragmatic Fairness: Developing Policies with Outcome Disparity Control", "abstract": "  We introduce a causal framework for designing optimal policies that satisfy\nfairness constraints. We take a pragmatic approach asking what we can do with\nan action space available to us and only with access to historical data. We\npropose two different fairness constraints: a moderation breaking constraint\nwhich aims at blocking moderation paths from the action and sensitive attribute\nto the outcome, and by that at reducing disparity in outcome levels as much as\nthe provided action space permits; and an equal benefit constraint which aims\nat distributing gain from the new and maximized policy equally across sensitive\nattribute levels, and thus at keeping pre-existing preferential treatment in\nplace or avoiding the introduction of new disparity. We introduce practical\nmethods for implementing the constraints and illustrate their uses on\nexperiments with semi-synthetic models.\n", "date": "2023-01-28", "categories": []}
{"id": "2301.12364", "title": "Fair Decision-making Under Uncertainty", "abstract": "  There has been concern within the artificial intelligence (AI) community and\nthe broader society regarding the potential lack of fairness of AI-based\ndecision-making systems. Surprisingly, there is little work quantifying and\nguaranteeing fairness in the presence of uncertainty which is prevalent in many\nsocially sensitive applications, ranging from marketing analytics to actuarial\nanalysis and recidivism prediction instruments. To this end, we study a\nlongitudinal censored learning problem subject to fairness constraints, where\nwe require that algorithmic decisions made do not affect certain individuals or\nsocial groups negatively in the presence of uncertainty on class label due to\ncensorship. We argue that this formulation has a broader applicability to\npractical scenarios concerning fairness. We show how the newly devised fairness\nnotions involving censored information and the general framework for fair\npredictions in the presence of censorship allow us to measure and mitigate\ndiscrimination under uncertainty that bridges the gap with real-world\napplications. Empirical evaluations on real-world discriminated datasets with\ncensorship demonstrate the practicality of our approach.\n", "date": "2023-01-29", "categories": []}
{"id": "2301.12369", "title": "Preserving Fairness in AI under Domain Shift", "abstract": "  Existing algorithms for ensuring fairness in AI use a single-shot training\nstrategy, where an AI model is trained on an annotated training dataset with\nsensitive attributes and then fielded for utilization. This training strategy\nis effective in problems with stationary distributions, where both training and\ntesting data are drawn from the same distribution. However, it is vulnerable\nwith respect to distributional shifts in the input space that may occur after\nthe initial training phase. As a result, the time-dependent nature of data can\nintroduce biases into the model predictions. Model retraining from scratch\nusing a new annotated dataset is a naive solution that is expensive and\ntime-consuming. We develop an algorithm to adapt a fair model to remain fair\nunder domain shift using solely new unannotated data points. We recast this\nlearning setting as an unsupervised domain adaptation problem. Our algorithm is\nbased on updating the model such that the internal representation of data\nremains unbiased despite distributional shifts in the input space. We provide\nextensive empirical validation on three widely employed fairness datasets to\ndemonstrate the effectiveness of our algorithm.\n", "date": "2023-01-29", "categories": []}
{"id": "2208.12212", "title": "Sustaining Fairness via Incremental Learning", "abstract": "  Machine learning systems are often deployed for making critical decisions\nlike credit lending, hiring, etc. While making decisions, such systems often\nencode the user's demographic information (like gender, age) in their\nintermediate representations. This can lead to decisions that are biased\ntowards specific demographics. Prior work has focused on debiasing intermediate\nrepresentations to ensure fair decisions. However, these approaches fail to\nremain fair with changes in the task or demographic distribution. To ensure\nfairness in the wild, it is important for a system to adapt to such changes as\nit accesses new data in an incremental fashion. In this work, we propose to\naddress this issue by introducing the problem of learning fair representations\nin an incremental learning setting. To this end, we present Fairness-aware\nIncremental Representation Learning (FaIRL), a representation learning system\nthat can sustain fairness while incrementally learning new tasks. FaIRL is able\nto achieve fairness and learn new tasks by controlling the rate-distortion\nfunction of the learned representations. Our empirical evaluations show that\nFaIRL is able to make fair decisions while achieving high performance on the\ntarget task, outperforming several baselines.\n", "date": "2022-08-25", "categories": []}
{"id": "2301.08375", "title": "Within-group fairness: A guidance for more sound between-group fairness", "abstract": "  As they have a vital effect on social decision-making, AI algorithms not only\nshould be accurate and but also should not pose unfairness against certain\nsensitive groups (e.g., non-white, women). Various specially designed AI\nalgorithms to ensure trained AI models to be fair between sensitive groups have\nbeen developed. In this paper, we raise a new issue that between-group fair AI\nmodels could treat individuals in a same sensitive group unfairly. We introduce\na new concept of fairness so-called within-group fairness which requires that\nAI models should be fair for those in a same sensitive group as well as those\nin different sensitive groups. We materialize the concept of within-group\nfairness by proposing corresponding mathematical definitions and developing\nlearning algorithms to control within-group fairness and between-group fairness\nsimultaneously. Numerical studies show that the proposed learning algorithms\nimprove within-group fairness without sacrificing accuracy as well as\nbetween-group fairness.\n", "date": "2023-01-20", "categories": []}
{"id": "2301.08412", "title": "Fair Credit Scorer through Bayesian Approach", "abstract": "  Machine learning currently plays an increasingly important role in people's\nlives in areas such as credit scoring, auto-driving, disease diagnosing, and\ninsurance quoting. However, in many of these areas, machine learning models\nhave performed unfair behaviors against some sub-populations, such as some\nparticular groups of race, sex, and age. These unfair behaviors can be on\naccount of the pre-existing bias in the training dataset due to historical and\nsocial factors. In this paper, we focus on a real-world application of credit\nscoring and construct a fair prediction model by introducing latent variables\nto remove the correlation between protected attributes, such as sex and age,\nwith the observable feature inputs, including house and job. For detailed\nimplementation, we apply Bayesian approaches, including the Markov Chain Monte\nCarlo simulation, to estimate our proposed fair model.\n", "date": "2023-01-20", "categories": []}
{"id": "2301.07060", "title": "Monotonicity for AI ethics and society: An empirical study of the\n  monotonic neural additive model in criminology, education, health care, and\n  finance", "abstract": "  Algorithm fairness in the application of artificial intelligence (AI) is\nessential for a better society. As the foundational axiom of social mechanisms,\nfairness consists of multiple facets. Although the machine learning (ML)\ncommunity has focused on intersectionality as a matter of statistical parity,\nespecially in discrimination issues, an emerging body of literature addresses\nanother facet -- monotonicity. Based on domain expertise, monotonicity plays a\nvital role in numerous fairness-related areas, where violations could misguide\nhuman decisions and lead to disastrous consequences. In this paper, we first\nsystematically evaluate the significance of applying monotonic neural additive\nmodels (MNAMs), which use a fairness-aware ML algorithm to enforce both\nindividual and pairwise monotonicity principles, for the fairness of AI ethics\nand society. We have found, through a hybrid method of theoretical reasoning,\nsimulation, and extensive empirical analysis, that considering monotonicity\naxioms is essential in all areas of fairness, including criminology, education,\nhealth care, and finance. Our research contributes to the interdisciplinary\nresearch at the interface of AI ethics, explainable AI (XAI), and\nhuman-computer interactions (HCIs). By evidencing the catastrophic consequences\nif monotonicity is not met, we address the significance of monotonicity\nrequirements in AI applications. Furthermore, we demonstrate that MNAMs are an\neffective fairness-aware ML approach by imposing monotonicity restrictions\nintegrating human intelligence.\n", "date": "2023-01-17", "categories": []}
{"id": "2206.12796", "title": "Transferring Fairness under Distribution Shifts via Fair Consistency\n  Regularization", "abstract": "  The increasing reliance on ML models in high-stakes tasks has raised a major\nconcern on fairness violations. Although there has been a surge of work that\nimproves algorithmic fairness, most of them are under the assumption of an\nidentical training and test distribution. In many real-world applications,\nhowever, such an assumption is often violated as previously trained fair models\nare often deployed in a different environment, and the fairness of such models\nhas been observed to collapse. In this paper, we study how to transfer model\nfairness under distribution shifts, a widespread issue in practice. We conduct\na fine-grained analysis of how the fair model is affected under different types\nof distribution shifts and find that domain shifts are more challenging than\nsubpopulation shifts. Inspired by the success of self-training in transferring\naccuracy under domain shifts, we derive a sufficient condition for transferring\ngroup fairness. Guided by it, we propose a practical algorithm with a fair\nconsistency regularization as the key component. A synthetic dataset benchmark,\nwhich covers all types of distribution shifts, is deployed for experimental\nverification of the theoretical findings. Experiments on synthetic and real\ndatasets including image and tabular data demonstrate that our approach\neffectively transfers fairness and accuracy under various distribution shifts.\n", "date": "2022-06-26", "categories": []}
{"id": "2301.05753", "title": "Fairness and Sequential Decision Making: Limits, Lessons, and\n  Opportunities", "abstract": "  As automated decision making and decision assistance systems become common in\neveryday life, research on the prevention or mitigation of potential harms that\narise from decisions made by these systems has proliferated. However, various\nresearch communities have independently conceptualized these harms, envisioned\npotential applications, and proposed interventions. The result is a somewhat\nfractured landscape of literature focused generally on ensuring decision-making\nalgorithms \"do the right thing\". In this paper, we compare and discuss work\nacross two major subsets of this literature: algorithmic fairness, which\nfocuses primarily on predictive systems, and ethical decision making, which\nfocuses primarily on sequential decision making and planning. We explore how\neach of these settings has articulated its normative concerns, the viability of\ndifferent techniques for these different settings, and how ideas from each\nsetting may have utility for the other.\n", "date": "2023-01-13", "categories": []}
{"id": "2301.04713", "title": "How do \"technical\" design-choices made when building algorithmic\n  decision-making tools for criminal justice authorities create constitutional\n  dangers?", "abstract": "  This two part paper argues that seemingly \"technical\" choices made by\ndevelopers of machine-learning based algorithmic tools used to inform decisions\nby criminal justice authorities can create serious constitutional dangers,\nenhancing the likelihood of abuse of decision-making power and the scope and\nmagnitude of injustice. Drawing on three algorithmic tools in use, or recently\nused, to assess the \"risk\" posed by individuals to inform how they should be\ntreated by criminal justice authorities, we integrate insights from data\nscience and public law scholarship to show how public law principles and more\nspecific legal duties that are rooted in these principles, are routinely\noverlooked in algorithmic tool-building and implementation. We argue that\ntechnical developers must collaborate closely with public law experts to ensure\nthat if algorithmic decision-support tools are to inform criminal justice\ndecisions, those tools are configured and implemented in a manner that is\ndemonstrably compliant with public law principles and doctrine, including\nrespect for human rights, throughout the tool-building process.\n", "date": "2023-01-11", "categories": []}
{"id": "2301.04993", "title": "Against Algorithmic Exploitation of Human Vulnerabilities", "abstract": "  Decisions such as which movie to watch next, which song to listen to, or\nwhich product to buy online, are increasingly influenced by recommender systems\nand user models that incorporate information on users' past behaviours,\npreferences, and digitally created content. Machine learning models that enable\nrecommendations and that are trained on user data may unintentionally leverage\ninformation on human characteristics that are considered vulnerabilities, such\nas depression, young age, or gambling addiction. The use of algorithmic\ndecisions based on latent vulnerable state representations could be considered\nmanipulative and could have a deteriorating impact on the condition of\nvulnerable individuals. In this paper, we are concerned with the problem of\nmachine learning models inadvertently modelling vulnerabilities, and want to\nraise awareness for this issue to be considered in legislation and AI ethics.\nHence, we define and describe common vulnerabilities, and illustrate cases\nwhere they are likely to play a role in algorithmic decision-making. We propose\na set of requirements for methods to detect the potential for vulnerability\nmodelling, detect whether vulnerable groups are treated differently by a model,\nand detect whether a model has created an internal representation of\nvulnerability. We conclude that explainable artificial intelligence methods may\nbe necessary for detecting vulnerability exploitation by machine learning-based\nrecommendation systems.\n", "date": "2023-01-12", "categories": []}
{"id": "2301.05012", "title": "Fairly Private: Investigating The Fairness of Visual Privacy\n  Preservation Algorithms", "abstract": "  As the privacy risks posed by camera surveillance and facial recognition have\ngrown, so has the research into privacy preservation algorithms. Among these,\nvisual privacy preservation algorithms attempt to impart bodily privacy to\nsubjects in visuals by obfuscating privacy-sensitive areas. While disparate\nperformances of facial recognition systems across phenotypes are the subject of\nmuch study, its counterpart, privacy preservation, is not commonly analysed\nfrom a fairness perspective. In this paper, the fairness of commonly used\nvisual privacy preservation algorithms is investigated through the performances\nof facial recognition models on obfuscated images. Experiments on the PubFig\ndataset clearly show that the privacy protection provided is unequal across\ngroups.\n", "date": "2023-01-12", "categories": []}
{"id": "2301.03791", "title": "Fair Recommendation by Geometric Interpretation and Analysis of Matrix\n  Factorization", "abstract": "  Matrix factorization-based recommender system is in effect an angle\npreserving dimensionality reduction technique. Since the frequency of items\nfollows power-law distribution, most vectors in the original dimension of user\nfeature vectors and item feature vectors lie on the same hyperplane. However,\nit is very difficult to reconstruct the embeddings in the original dimension\nanalytically, so we reformulate the original angle preserving dimensionality\nreduction problem into a distance preserving dimensionality reduction problem.\nWe show that the geometric shape of input data of recommender system in its\noriginal higher dimension are distributed on co-centric circles with\ninteresting properties, and design a paraboloid-based matrix factorization\nnamed ParaMat to solve the recommendation problem. In the experiment section,\nwe compare our algorithm with 8 other algorithms and prove our new method is\nthe most fair algorithm compared with modern day recommender systems such as\nZeroMat and DotMat Hybrid.\n", "date": "2023-01-10", "categories": []}
{"id": "2212.14467", "title": "Cluster-level Group Representativity Fairness in $k$-means Clustering", "abstract": "  There has been much interest recently in developing fair clustering\nalgorithms that seek to do justice to the representation of groups defined\nalong sensitive attributes such as race and gender. We observe that clustering\nalgorithms could generate clusters such that different groups are disadvantaged\nwithin different clusters. We develop a clustering algorithm, building upon the\ncentroid clustering paradigm pioneered by classical algorithms such as\n$k$-means, where we focus on mitigating the unfairness experienced by the\nmost-disadvantaged group within each cluster. Our method uses an iterative\noptimisation paradigm whereby an initial cluster assignment is modified by\nreassigning objects to clusters such that the worst-off sensitive group within\neach cluster is benefitted. We demonstrate the effectiveness of our method\nthrough extensive empirical evaluations over a novel evaluation metric on\nreal-world datasets. Specifically, we show that our method is effective in\nenhancing cluster-level group representativity fairness significantly at low\nimpact on cluster coherence.\n", "date": "2022-12-29", "categories": []}
{"id": "2206.03985", "title": "How unfair is private learning ?", "abstract": "  As machine learning algorithms are deployed on sensitive data in critical\ndecision making processes, it is becoming increasingly important that they are\nalso private and fair. In this paper, we show that, when the data has a\nlong-tailed structure, it is not possible to build accurate learning algorithms\nthat are both private and results in higher accuracy on minority\nsubpopulations. We further show that relaxing overall accuracy can lead to good\nfairness even with strict privacy requirements. To corroborate our theoretical\nresults in practice, we provide an extensive set of experimental results using\na variety of synthetic, vision (CIFAR10 and CelebA), and tabular (Law School)\ndatasets and learning algorithms.\n", "date": "2022-06-08", "categories": []}
{"id": "2212.13014", "title": "Bias Mitigation Framework for Intersectional Subgroups in Neural\n  Networks", "abstract": "  We propose a fairness-aware learning framework that mitigates intersectional\nsubgroup bias associated with protected attributes. Prior research has\nprimarily focused on mitigating one kind of bias by incorporating complex\nfairness-driven constraints into optimization objectives or designing\nadditional layers that focus on specific protected attributes. We introduce a\nsimple and generic bias mitigation approach that prevents models from learning\nrelationships between protected attributes and output variable by reducing\nmutual information between them. We demonstrate that our approach is effective\nin reducing bias with little or no drop in accuracy. We also show that the\nmodels trained with our learning framework become causally fair and insensitive\nto the values of protected attributes. Finally, we validate our approach by\nstudying feature interactions between protected and non-protected attributes.\nWe demonstrate that these interactions are significantly reduced when applying\nour bias mitigation.\n", "date": "2022-12-26", "categories": []}
{"id": "2212.12040", "title": "Graph Learning with Localized Neighborhood Fairness", "abstract": "  Learning fair graph representations for downstream applications is becoming\nincreasingly important, but existing work has mostly focused on improving\nfairness at the global level by either modifying the graph structure or\nobjective function without taking into account the local neighborhood of a\nnode. In this work, we formally introduce the notion of neighborhood fairness\nand develop a computational framework for learning such locally fair\nembeddings. We argue that the notion of neighborhood fairness is more\nappropriate since GNN-based models operate at the local neighborhood level of a\nnode. Our neighborhood fairness framework has two main components that are\nflexible for learning fair graph representations from arbitrary data: the first\naims to construct fair neighborhoods for any arbitrary node in a graph and the\nsecond enables adaption of these fair neighborhoods to better capture certain\napplication or data-dependent constraints, such as allowing neighborhoods to be\nmore biased towards certain attributes or neighbors in the graph.Furthermore,\nwhile link prediction has been extensively studied, we are the first to\ninvestigate the graph representation learning task of fair link classification.\nWe demonstrate the effectiveness of the proposed neighborhood fairness\nframework for a variety of graph machine learning tasks including fair link\nprediction, link classification, and learning fair graph embeddings. Notably,\nour approach achieves not only better fairness but also increases the accuracy\nin the majority of cases across a wide variety of graphs, problem settings, and\nmetrics.\n", "date": "2022-12-22", "categories": []}
{"id": "2212.11207", "title": "A Seven-Layer Model for Standardising AI Fairness Assessment", "abstract": "  Problem statement: Standardisation of AI fairness rules and benchmarks is\nchallenging because AI fairness and other ethical requirements depend on\nmultiple factors such as context, use case, type of the AI system, and so on.\nIn this paper, we elaborate that the AI system is prone to biases at every\nstage of its lifecycle, from inception to its usage, and that all stages\nrequire due attention for mitigating AI bias. We need a standardised approach\nto handle AI fairness at every stage. Gap analysis: While AI fairness is a hot\nresearch topic, a holistic strategy for AI fairness is generally missing. Most\nresearchers focus only on a few facets of AI model-building. Peer review shows\nexcessive focus on biases in the datasets, fairness metrics, and algorithmic\nbias. In the process, other aspects affecting AI fairness get ignored. The\nsolution proposed: We propose a comprehensive approach in the form of a novel\nseven-layer model, inspired by the Open System Interconnection (OSI) model, to\nstandardise AI fairness handling. Despite the differences in the various\naspects, most AI systems have similar model-building stages. The proposed model\nsplits the AI system lifecycle into seven abstraction layers, each\ncorresponding to a well-defined AI model-building or usage stage. We also\nprovide checklists for each layer and deliberate on potential sources of bias\nin each layer and their mitigation methodologies. This work will facilitate\nlayer-wise standardisation of AI fairness rules and benchmarking parameters.\n", "date": "2022-12-21", "categories": []}
{"id": "2206.03150", "title": "Group Meritocratic Fairness in Linear Contextual Bandits", "abstract": "  We study the linear contextual bandit problem where an agent has to select\none candidate from a pool and each candidate belongs to a sensitive group. In\nthis setting, candidates' rewards may not be directly comparable between\ngroups, for example when the agent is an employer hiring candidates from\ndifferent ethnic groups and some groups have a lower reward due to\ndiscriminatory bias and/or social injustice. We propose a notion of fairness\nthat states that the agent's policy is fair when it selects a candidate with\nhighest relative rank, which measures how good the reward is when compared to\ncandidates from the same group. This is a very strong notion of fairness, since\nthe relative rank is not directly observed by the agent and depends on the\nunderlying reward model and on the distribution of rewards. Thus we study the\nproblem of learning a policy which approximates a fair policy under the\ncondition that the contexts are independent between groups and the distribution\nof rewards of each group is absolutely continuous. In particular, we design a\ngreedy policy which at each round constructs a ridge regression estimate from\nthe observed context-reward pairs, and then computes an estimate of the\nrelative rank of each candidate using the empirical cumulative distribution\nfunction. We prove that, despite its simplicity and the lack of an initial\nexploration phase, the greedy policy achieves, up to log factors and with high\nprobability, a fair pseudo-regret of order $\\sqrt{dT}$ after $T$ rounds, where\n$d$ is the dimension of the context vectors. The policy also satisfies\ndemographic parity at each round when averaged over all possible information\navailable before the selection. Finally, we use simulated settings and\nexperiments on the US census data to show that our policy achieves sub-linear\nfair pseudo-regret also in practice.\n", "date": "2022-06-07", "categories": []}
{"id": "2212.08578", "title": "Provable Fairness for Neural Network Models using Formal Verification", "abstract": "  Machine learning models are increasingly deployed for critical\ndecision-making tasks, making it important to verify that they do not contain\ngender or racial biases picked up from training data. Typical approaches to\nachieve fairness revolve around efforts to clean or curate training data, with\npost-hoc statistical evaluation of the fairness of the model on evaluation\ndata. In contrast, we propose techniques to \\emph{prove} fairness using\nrecently developed formal methods that verify properties of neural network\nmodels.Beyond the strength of guarantee implied by a formal proof, our methods\nhave the advantage that we do not need explicit training or evaluation data\n(which is often proprietary) in order to analyze a given trained model. In\nexperiments on two familiar datasets in the fairness literature (COMPAS and\nADULTS), we show that through proper training, we can reduce unfairness by an\naverage of 65.4\\% at a cost of less than 1\\% in AUC score.\n", "date": "2022-12-16", "categories": []}
{"id": "2212.06803", "title": "Fair Infinitesimal Jackknife: Mitigating the Influence of Biased\n  Training Data Points Without Refitting", "abstract": "  In consequential decision-making applications, mitigating unwanted biases in\nmachine learning models that yield systematic disadvantage to members of groups\ndelineated by sensitive attributes such as race and gender is one key\nintervention to strive for equity. Focusing on demographic parity and equality\nof opportunity, in this paper we propose an algorithm that improves the\nfairness of a pre-trained classifier by simply dropping carefully selected\ntraining data points. We select instances based on their influence on the\nfairness metric of interest, computed using an infinitesimal jackknife-based\napproach. The dropping of training points is done in principle, but in practice\ndoes not require the model to be refit. Crucially, we find that such an\nintervention does not substantially reduce the predictive performance of the\nmodel but drastically improves the fairness metric. Through careful\nexperiments, we evaluate the effectiveness of the proposed approach on diverse\ntasks and find that it consistently improves upon existing alternatives.\n", "date": "2022-12-13", "categories": []}
{"id": "2209.10222", "title": "Fairness Reprogramming", "abstract": "  Despite a surge of recent advances in promoting machine Learning (ML)\nfairness, the existing mainstream approaches mostly require retraining or\nfinetuning the entire weights of the neural network to meet the fairness\ncriteria. However, this is often infeasible in practice for those large-scale\ntrained models due to large computational and storage costs, low data\nefficiency, and model privacy issues. In this paper, we propose a new generic\nfairness learning paradigm, called FairReprogram, which incorporates the model\nreprogramming technique. Specifically, FairReprogram considers the case where\nmodels can not be changed and appends to the input a set of perturbations,\ncalled the fairness trigger, which is tuned towards the fairness criteria under\na min-max formulation. We further introduce an information-theoretic framework\nthat explains why and under what conditions fairness goals can be achieved\nusing the fairness trigger. We show both theoretically and empirically that the\nfairness trigger can effectively obscure demographic biases in the output\nprediction of fixed ML models by providing false demographic information that\nhinders the model from utilizing the correct demographic information to make\nthe prediction. Extensive experiments on both NLP and CV datasets demonstrate\nthat our method can achieve better fairness improvements than retraining-based\nmethods with far less data dependency under two widely-used fairness criteria.\nCodes are available at\nhttps://github.com/UCSB-NLP-Chang/Fairness-Reprogramming.git.\n", "date": "2022-09-21", "categories": []}
{"id": "2210.08649", "title": "Loss Minimization through the Lens of Outcome Indistinguishability", "abstract": "  We present a new perspective on loss minimization and the recent notion of\nOmniprediction through the lens of Outcome Indistingusihability. For a\ncollection of losses and hypothesis class, omniprediction requires that a\npredictor provide a loss-minimization guarantee simultaneously for every loss\nin the collection compared to the best (loss-specific) hypothesis in the class.\nWe present a generic template to learn predictors satisfying a guarantee we\ncall Loss Outcome Indistinguishability. For a set of statistical tests--based\non a collection of losses and hypothesis class--a predictor is Loss OI if it is\nindistinguishable (according to the tests) from Nature's true probabilities\nover outcomes. By design, Loss OI implies omniprediction in a direct and\nintuitive manner. We simplify Loss OI further, decomposing it into a\ncalibration condition plus multiaccuracy for a class of functions derived from\nthe loss and hypothesis classes. By careful analysis of this class, we give\nefficient constructions of omnipredictors for interesting classes of loss\nfunctions, including non-convex losses.\n  This decomposition highlights the utility of a new multi-group fairness\nnotion that we call calibrated multiaccuracy, which lies in between\nmultiaccuracy and multicalibration. We show that calibrated multiaccuracy\nimplies Loss OI for the important set of convex losses arising from Generalized\nLinear Models, without requiring full multicalibration. For such losses, we\nshow an equivalence between our computational notion of Loss OI and a geometric\nnotion of indistinguishability, formulated as Pythagorean theorems in the\nassociated Bregman divergence. We give an efficient algorithm for calibrated\nmultiaccuracy with computational complexity comparable to that of\nmultiaccuracy. In all, calibrated multiaccuracy offers an interesting tradeoff\npoint between efficiency and generality in the omniprediction landscape.\n", "date": "2022-10-16", "categories": []}
{"id": "2212.03840", "title": "Fairness and Explainability: Bridging the Gap Towards Fair Model\n  Explanations", "abstract": "  While machine learning models have achieved unprecedented success in\nreal-world applications, they might make biased/unfair decisions for specific\ndemographic groups and hence result in discriminative outcomes. Although\nresearch efforts have been devoted to measuring and mitigating bias, they\nmainly study bias from the result-oriented perspective while neglecting the\nbias encoded in the decision-making procedure. This results in their inability\nto capture procedure-oriented bias, which therefore limits the ability to have\na fully debiasing method. Fortunately, with the rapid development of\nexplainable machine learning, explanations for predictions are now available to\ngain insights into the procedure. In this work, we bridge the gap between\nfairness and explainability by presenting a novel perspective of\nprocedure-oriented fairness based on explanations. We identify the\nprocedure-based bias by measuring the gap of explanation quality between\ndifferent groups with Ratio-based and Value-based Explanation Fairness. The new\nmetrics further motivate us to design an optimization objective to mitigate the\nprocedure-based bias where we observe that it will also mitigate bias from the\nprediction. Based on our designed optimization objective, we propose a\nComprehensive Fairness Algorithm (CFA), which simultaneously fulfills multiple\nobjectives - improving traditional fairness, satisfying explanation fairness,\nand maintaining the utility performance. Extensive experiments on real-world\ndatasets demonstrate the effectiveness of our proposed CFA and highlight the\nimportance of considering fairness from the explainability perspective. Our\ncode is publicly available at\nhttps://github.com/YuyingZhao/FairExplanations-CFA .\n", "date": "2022-12-07", "categories": []}
{"id": "2212.02614", "title": "Can Ensembling Pre-processing Algorithms Lead to Better Machine Learning\n  Fairness?", "abstract": "  As machine learning (ML) systems get adopted in more critical areas, it has\nbecome increasingly crucial to address the bias that could occur in these\nsystems. Several fairness pre-processing algorithms are available to alleviate\nimplicit biases during model training. These algorithms employ different\nconcepts of fairness, often leading to conflicting strategies with\nconsequential trade-offs between fairness and accuracy. In this work, we\nevaluate three popular fairness pre-processing algorithms and investigate the\npotential for combining all algorithms into a more robust pre-processing\nensemble. We report on lessons learned that can help practitioners better\nselect fairness algorithms for their models.\n", "date": "2022-12-05", "categories": []}
{"id": "2211.14489", "title": "Mitigating Relational Bias on Knowledge Graphs", "abstract": "  Knowledge graph data are prevalent in real-world applications, and knowledge\ngraph neural networks (KGNNs) are essential techniques for knowledge graph\nrepresentation learning. Although KGNN effectively models the structural\ninformation from knowledge graphs, these frameworks amplify the underlying data\nbias that leads to discrimination towards certain groups or individuals in\nresulting applications. Additionally, as existing debiasing approaches mainly\nfocus on the entity-wise bias, eliminating the multi-hop relational bias that\npervasively exists in knowledge graphs remains an open question. However, it is\nvery challenging to eliminate relational bias due to the sparsity of the paths\nthat generate the bias and the non-linear proximity structure of knowledge\ngraphs. To tackle the challenges, we propose Fair-KGNN, a KGNN framework that\nsimultaneously alleviates multi-hop bias and preserves the proximity\ninformation of entity-to-relation in knowledge graphs. The proposed framework\nis generalizable to mitigate the relational bias for all types of KGNN. We\ndevelop two instances of Fair-KGNN incorporating with two state-of-the-art KGNN\nmodels, RGCN and CompGCN, to mitigate gender-occupation and nationality-salary\nbias. The experiments carried out on three benchmark knowledge graph datasets\ndemonstrate that the Fair-KGNN can effectively mitigate unfair situations\nduring representation learning while preserving the predictive performance of\nKGNN models.\n", "date": "2022-11-26", "categories": []}
{"id": "2212.01725", "title": "Fairness in Contextual Resource Allocation Systems: Metrics and\n  Incompatibility Results", "abstract": "  We study critical systems that allocate scarce resources to satisfy basic\nneeds, such as homeless services that provide housing. These systems often\nsupport communities disproportionately affected by systemic racial, gender, or\nother injustices, so it is crucial to design these systems with fairness\nconsiderations in mind. To address this problem, we propose a framework for\nevaluating fairness in contextual resource allocation systems that is inspired\nby fairness metrics in machine learning. This framework can be applied to\nevaluate the fairness properties of a historical policy, as well as to impose\nconstraints in the design of new (counterfactual) allocation policies. Our work\nculminates with a set of incompatibility results that investigate the interplay\nbetween the different fairness metrics we propose. Notably, we demonstrate\nthat: 1) fairness in allocation and fairness in outcomes are usually\nincompatible; 2) policies that prioritize based on a vulnerability score will\nusually result in unequal outcomes across groups, even if the score is\nperfectly calibrated; 3) policies using contextual information beyond what is\nneeded to characterize baseline risk and treatment effects can be fairer in\ntheir outcomes than those using just baseline risk and treatment effects; and\n4) policies using group status in addition to baseline risk and treatment\neffects are as fair as possible given all available information. Our framework\ncan help guide the discussion among stakeholders in deciding which fairness\nmetrics to impose when allocating scarce resources.\n", "date": "2022-12-04", "categories": []}
{"id": "2212.02474", "title": "Certifying Fairness of Probabilistic Circuits", "abstract": "  With the increased use of machine learning systems for decision making,\nquestions about the fairness properties of such systems start to take center\nstage. Most existing work on algorithmic fairness assume complete observation\nof features at prediction time, as is the case for popular notions like\nstatistical parity and equal opportunity. However, this is not sufficient for\nmodels that can make predictions with partial observation as we could miss\npatterns of bias and incorrectly certify a model to be fair. To address this, a\nrecently introduced notion of fairness asks whether the model exhibits any\ndiscrimination pattern, in which an individual characterized by (partial)\nfeature observations, receives vastly different decisions merely by disclosing\none or more sensitive attributes such as gender and race. By explicitly\naccounting for partial observations, this provides a much more fine-grained\nnotion of fairness.\n  In this paper, we propose an algorithm to search for discrimination patterns\nin a general class of probabilistic models, namely probabilistic circuits.\nPreviously, such algorithms were limited to naive Bayes classifiers which make\nstrong independence assumptions; by contrast, probabilistic circuits provide a\nunifying framework for a wide range of tractable probabilistic models and can\neven be compiled from certain classes of Bayesian networks and probabilistic\nprograms, making our method much more broadly applicable. Furthermore, for an\nunfair model, it may be useful to quickly find discrimination patterns and\ndistill them for better interpretability. As such, we also propose a\nsampling-based approach to more efficiently mine discrimination patterns, and\nintroduce new classes of patterns such as minimal, maximal, and Pareto optimal\npatterns that can effectively summarize exponentially many discrimination\npatterns\n", "date": "2022-12-05", "categories": []}
{"id": "2212.00926", "title": "Fair Generative Models via Transfer Learning", "abstract": "  This work addresses fair generative models. Dataset biases have been a major\ncause of unfairness in deep generative models. Previous work had proposed to\naugment large, biased datasets with small, unbiased reference datasets. Under\nthis setup, a weakly-supervised approach has been proposed, which achieves\nstate-of-the-art quality and fairness in generated samples. In our work, based\non this setup, we propose a simple yet effective approach. Specifically, first,\nwe propose fairTL, a transfer learning approach to learn fair generative\nmodels. Under fairTL, we pre-train the generative model with the available\nlarge, biased datasets and subsequently adapt the model using the small,\nunbiased reference dataset. We find that our fairTL can learn expressive sample\ngeneration during pre-training, thanks to the large (biased) dataset. This\nknowledge is then transferred to the target model during adaptation, which also\nlearns to capture the underlying fair distribution of the small reference\ndataset. Second, we propose fairTL++, where we introduce two additional\ninnovations to improve upon fairTL: (i) multiple feedback and (ii)\nLinear-Probing followed by Fine-Tuning (LP-FT). Taking one step further, we\nconsider an alternative, challenging setup when only a pre-trained (potentially\nbiased) model is available but the dataset that was used to pre-train the model\nis inaccessible. We demonstrate that our proposed fairTL and fairTL++ remain\nvery effective under this setup. We note that previous work requires access to\nthe large, biased datasets and is incapable of handling this more challenging\nsetup. Extensive experiments show that fairTL and fairTL++ achieve\nstate-of-the-art in both quality and fairness of generated samples. The code\nand additional resources can be found at bearwithchris.github.io/fairTL/.\n", "date": "2022-12-02", "categories": []}
{"id": "2210.13023", "title": "FairGen: Fair Synthetic Data Generation", "abstract": "  With the rising adoption of Machine Learning across the domains like banking,\npharmaceutical, ed-tech, etc, it has become utmost important to adopt\nresponsible AI methods to ensure models are not unfairly discriminating against\nany group. Given the lack of clean training data, generative adversarial\ntechniques are preferred to generate synthetic data with several\nstate-of-the-art architectures readily available across various domains from\nunstructured data such as text, images to structured datasets modelling fraud\ndetection and many more. These techniques overcome several challenges such as\nclass imbalance, limited training data, restricted access to data due to\nprivacy issues. Existing work focusing on generating fair data either works for\na certain GAN architecture or is very difficult to tune across the GANs. In\nthis paper, we propose a pipeline to generate fairer synthetic data independent\nof the GAN architecture. The proposed paper utilizes a pre-processing algorithm\nto identify and remove bias inducing samples. In particular, we claim that\nwhile generating synthetic data most GANs amplify bias present in the training\ndata but by removing these bias inducing samples, GANs essentially focuses more\non real informative samples. Our experimental evaluation on two open-source\ndatasets demonstrates how the proposed pipeline is generating fair data along\nwith improved performance in some cases.\n", "date": "2022-10-24", "categories": []}
{"id": "2212.00630", "title": "Probably Approximate Shapley Fairness with Applications in Machine\n  Learning", "abstract": "  The Shapley value (SV) is adopted in various scenarios in machine learning\n(ML), including data valuation, agent valuation, and feature attribution, as it\nsatisfies their fairness requirements. However, as exact SVs are infeasible to\ncompute in practice, SV estimates are approximated instead. This approximation\nstep raises an important question: do the SV estimates preserve the fairness\nguarantees of exact SVs? We observe that the fairness guarantees of exact SVs\nare too restrictive for SV estimates. Thus, we generalise Shapley fairness to\nprobably approximate Shapley fairness and propose fidelity score, a metric to\nmeasure the variation of SV estimates, that determines how probable the\nfairness guarantees hold. Our last theoretical contribution is a novel greedy\nactive estimation (GAE) algorithm that will maximise the lowest fidelity score\nand achieve a better fairness guarantee than the de facto Monte-Carlo\nestimation. We empirically verify GAE outperforms several existing methods in\nguaranteeing fairness while remaining competitive in estimation accuracy in\nvarious ML scenarios using real-world datasets.\n", "date": "2022-12-01", "categories": []}
{"id": "2211.17067", "title": "Fair Ranking with Noisy Protected Attributes", "abstract": "  The fair-ranking problem, which asks to rank a given set of items to maximize\nutility subject to group fairness constraints, has received attention in the\nfairness, information retrieval, and machine learning literature. Recent works,\nhowever, observe that errors in socially-salient (including protected)\nattributes of items can significantly undermine fairness guarantees of existing\nfair-ranking algorithms and raise the problem of mitigating the effect of such\nerrors. We study the fair-ranking problem under a model where socially-salient\nattributes of items are randomly and independently perturbed. We present a\nfair-ranking framework that incorporates group fairness requirements along with\nprobabilistic information about perturbations in socially-salient attributes.\nWe provide provable guarantees on the fairness and utility attainable by our\nframework and show that it is information-theoretically impossible to\nsignificantly beat these guarantees. Our framework works for multiple\nnon-disjoint attributes and a general class of fairness constraints that\nincludes proportional and equal representation. Empirically, we observe that,\ncompared to baselines, our algorithm outputs rankings with higher fairness, and\nhas a similar or better fairness-utility trade-off compared to baselines.\n", "date": "2022-11-30", "categories": []}
{"id": "2208.10451", "title": "Minimax AUC Fairness: Efficient Algorithm with Provable Convergence", "abstract": "  The use of machine learning models in consequential decision making often\nexacerbates societal inequity, in particular yielding disparate impact on\nmembers of marginalized groups defined by race and gender. The area under the\nROC curve (AUC) is widely used to evaluate the performance of a scoring\nfunction in machine learning, but is studied in algorithmic fairness less than\nother performance metrics. Due to the pairwise nature of the AUC, defining an\nAUC-based group fairness metric is pairwise-dependent and may involve both\n\\emph{intra-group} and \\emph{inter-group} AUCs. Importantly, considering only\none category of AUCs is not sufficient to mitigate unfairness in AUC\noptimization. In this paper, we propose a minimax learning and bias mitigation\nframework that incorporates both intra-group and inter-group AUCs while\nmaintaining utility. Based on this Rawlsian framework, we design an efficient\nstochastic optimization algorithm and prove its convergence to the minimum\ngroup-level AUC. We conduct numerical experiments on both synthetic and\nreal-world datasets to validate the effectiveness of the minimax framework and\nthe proposed optimization algorithm.\n", "date": "2022-08-22", "categories": []}
{"id": "2210.10837", "title": "On Learning Fairness and Accuracy on Multiple Subgroups", "abstract": "  We propose an analysis in fair learning that preserves the utility of the\ndata while reducing prediction disparities under the criteria of group\nsufficiency. We focus on the scenario where the data contains multiple or even\nmany subgroups, each with limited number of samples. As a result, we present a\nprincipled method for learning a fair predictor for all subgroups via\nformulating it as a bilevel objective. Specifically, the subgroup specific\npredictors are learned in the lower-level through a small amount of data and\nthe fair predictor. In the upper-level, the fair predictor is updated to be\nclose to all subgroup specific predictors. We further prove that such a bilevel\nobjective can effectively control the group sufficiency and generalization\nerror. We evaluate the proposed framework on real-world datasets. Empirical\nevidence suggests the consistently improved fair predictions, as well as the\ncomparable accuracy to the baselines.\n", "date": "2022-10-19", "categories": []}
{"id": "2211.04442", "title": "Algorithmic Bias in Machine Learning Based Delirium Prediction", "abstract": "  Although prediction models for delirium, a commonly occurring condition\nduring general hospitalization or post-surgery, have not gained huge\npopularity, their algorithmic bias evaluation is crucial due to the existing\nassociation between social determinants of health and delirium risk. In this\ncontext, using MIMIC-III and another academic hospital dataset, we present some\ninitial experimental evidence showing how sociodemographic features such as sex\nand race can impact the model performance across subgroups. With this work, our\nintent is to initiate a discussion about the intersectionality effects of old\nage, race and socioeconomic factors on the early-stage detection and prevention\nof delirium using ML.\n", "date": "2022-11-08", "categories": []}
{"id": "2211.14639", "title": "Gender Biases Unexpectedly Fluctuate in the Pre-training Stage of Masked\n  Language Models", "abstract": "  Masked language models pick up gender biases during pre-training. Such biases\nare usually attributed to a certain model architecture and its pre-training\ncorpora, with the implicit assumption that other variations in the pre-training\nprocess, such as the choices of the random seed or the stopping point, have no\neffect on the biases measured. However, we show that severe fluctuations exist\nat the fundamental level of individual templates, invalidating the assumption.\nFurther against the intuition of how humans acquire biases, these fluctuations\nare not correlated with the certainty of the predicted pronouns or the\nprofession frequencies in pre-training corpora. We release our code and data to\nbenefit future research.\n", "date": "2022-11-26", "categories": []}
{"id": "2211.14858", "title": "\"Explain it in the Same Way!\" -- Model-Agnostic Group Fairness of\n  Counterfactual Explanations", "abstract": "  Counterfactual explanations are a popular type of explanation for making the\noutcomes of a decision making system transparent to the user. Counterfactual\nexplanations tell the user what to do in order to change the outcome of the\nsystem in a desirable way. However, it was recently discovered that the\nrecommendations of what to do can differ significantly in their complexity\nbetween protected groups of individuals. Providing more difficult\nrecommendations of actions to one group leads to a disadvantage of this group\ncompared to other groups.\n  In this work we propose a model-agnostic method for computing counterfactual\nexplanations that do not differ significantly in their complexity between\nprotected groups.\n", "date": "2022-11-27", "categories": []}
{"id": "2211.15181", "title": "MixFairFace: Towards Ultimate Fairness via MixFair Adapter in Face\n  Recognition", "abstract": "  Although significant progress has been made in face recognition, demographic\nbias still exists in face recognition systems. For instance, it usually happens\nthat the face recognition performance for a certain demographic group is lower\nthan the others. In this paper, we propose MixFairFace framework to improve the\nfairness in face recognition models. First of all, we argue that the commonly\nused attribute-based fairness metric is not appropriate for face recognition. A\nface recognition system can only be considered fair while every person has a\nclose performance. Hence, we propose a new evaluation protocol to fairly\nevaluate the fairness performance of different approaches. Different from\nprevious approaches that require sensitive attribute labels such as race and\ngender for reducing the demographic bias, we aim at addressing the identity\nbias in face representation, i.e., the performance inconsistency between\ndifferent identities, without the need for sensitive attribute labels. To this\nend, we propose MixFair Adapter to determine and reduce the identity bias of\ntraining samples. Our extensive experiments demonstrate that our MixFairFace\napproach achieves state-of-the-art fairness performance on all benchmark\ndatasets.\n", "date": "2022-11-28", "categories": []}
{"id": "2211.11109", "title": "Deep Learning on a Healthy Data Diet: Finding Important Examples for\n  Fairness", "abstract": "  Data-driven predictive solutions predominant in commercial applications tend\nto suffer from biases and stereotypes, which raises equity concerns. Prediction\nmodels may discover, use, or amplify spurious correlations based on gender or\nother protected personal characteristics, thus discriminating against\nmarginalized groups. Mitigating gender bias has become an important research\nfocus in natural language processing (NLP) and is an area where annotated\ncorpora are available. Data augmentation reduces gender bias by adding\ncounterfactual examples to the training dataset. In this work, we show that\nsome of the examples in the augmented dataset can be not important or even\nharmful for fairness. We hence propose a general method for pruning both the\nfactual and counterfactual examples to maximize the model's fairness as\nmeasured by the demographic parity, equality of opportunity, and equality of\nodds. The fairness achieved by our method surpasses that of data augmentation\non three text classification datasets, using no more than half of the examples\nin the augmented dataset. Our experiments are conducted using models of varying\nsizes and pre-training settings.\n", "date": "2022-11-20", "categories": []}
{"id": "2211.11892", "title": "Equality of Effort via Algorithmic Recourse", "abstract": "  This paper proposes a method for measuring fairness through equality of\neffort by applying algorithmic recourse through minimal interventions. Equality\nof effort is a property that can be quantified at both the individual and the\ngroup level. It answers the counterfactual question: what is the minimal cost\nfor a protected individual or the average minimal cost for a protected group of\nindividuals to reverse the outcome computed by an automated system? Algorithmic\nrecourse increases the flexibility and applicability of the notion of equal\neffort: it overcomes its previous limitations by reconciling multiple treatment\nvariables, introducing feasibility and plausibility constraints, and\nintegrating the actual relative costs of interventions. We extend the existing\ndefinition of equality of effort and present an algorithm for its assessment\nvia algorithmic recourse. We validate our approach both on synthetic data and\non the German credit dataset.\n", "date": "2022-11-21", "categories": []}
{"id": "2211.13972", "title": "Picking on the Same Person: Does Algorithmic Monoculture lead to Outcome\n  Homogenization?", "abstract": "  As the scope of machine learning broadens, we observe a recurring theme of\nalgorithmic monoculture: the same systems, or systems that share components\n(e.g. training data), are deployed by multiple decision-makers. While sharing\noffers clear advantages (e.g. amortizing costs), does it bear risks? We\nintroduce and formalize one such risk, outcome homogenization: the extent to\nwhich particular individuals or groups experience negative outcomes from all\ndecision-makers. If the same individuals or groups exclusively experience\nundesirable outcomes, this may institutionalize systemic exclusion and\nreinscribe social hierarchy. To relate algorithmic monoculture and outcome\nhomogenization, we propose the component-sharing hypothesis: if decision-makers\nshare components like training data or specific models, then they will produce\nmore homogeneous outcomes. We test this hypothesis on algorithmic fairness\nbenchmarks, demonstrating that sharing training data reliably exacerbates\nhomogenization, with individual-level effects generally exceeding group-level\neffects. Further, given the dominant paradigm in AI of foundation models, i.e.\nmodels that can be adapted for myriad downstream tasks, we test whether model\nsharing homogenizes outcomes across tasks. We observe mixed results: we find\nthat for both vision and language settings, the specific methods for adapting a\nfoundation model significantly influence the degree of outcome homogenization.\nWe conclude with philosophical analyses of and societal challenges for outcome\nhomogenization, with an eye towards implications for deployed machine learning\nsystems.\n", "date": "2022-11-25", "categories": []}
{"id": "2211.11835", "title": "Fairness Increases Adversarial Vulnerability", "abstract": "  The remarkable performance of deep learning models and their applications in\nconsequential domains (e.g., facial recognition) introduces important\nchallenges at the intersection of equity and security. Fairness and robustness\nare two desired notions often required in learning models. Fairness ensures\nthat models do not disproportionately harm (or benefit) some groups over\nothers, while robustness measures the models' resilience against small input\nperturbations.\n  This paper shows the existence of a dichotomy between fairness and\nrobustness, and analyzes when achieving fairness decreases the model robustness\nto adversarial samples. The reported analysis sheds light on the factors\ncausing such contrasting behavior, suggesting that distance to the decision\nboundary across groups as a key explainer for this behavior. Extensive\nexperiments on non-linear models and different architectures validate the\ntheoretical findings in multiple vision domains. Finally, the paper proposes a\nsimple, yet effective, solution to construct models achieving good tradeoffs\nbetween fairness and robustness.\n", "date": "2022-11-21", "categories": []}
{"id": "2209.12226", "title": "Re-contextualizing Fairness in NLP: The Case of India", "abstract": "  Recent research has revealed undesirable biases in NLP data and models.\nHowever, these efforts focus on social disparities in West, and are not\ndirectly portable to other geo-cultural contexts. In this paper, we focus on\nNLP fair-ness in the context of India. We start with a brief account of the\nprominent axes of social disparities in India. We build resources for fairness\nevaluation in the Indian context and use them to demonstrate prediction biases\nalong some of the axes. We then delve deeper into social stereotypes for Region\nandReligion, demonstrating its prevalence in corpora and models. Finally, we\noutline a holistic research agenda to re-contextualize NLP fairness research\nfor the Indian context, ac-counting for Indian societal context, bridging\ntechnological gaps in NLP capabilities and re-sources, and adapting to Indian\ncultural values. While we focus on India, this framework can be generalized to\nother geo-cultural contexts.\n", "date": "2022-09-25", "categories": []}
{"id": "2211.10508", "title": "Distributionally Robust Survival Analysis: A Novel Fairness Loss Without\n  Demographics", "abstract": "  We propose a general approach for training survival analysis models that\nminimizes a worst-case error across all subpopulations that are large enough\n(occurring with at least a user-specified minimum probability). This approach\nuses a training loss function that does not know any demographic information to\ntreat as sensitive. Despite this, we demonstrate that our proposed approach\noften scores better on recently established fairness metrics (without a\nsignificant drop in prediction accuracy) compared to various baselines,\nincluding ones which directly use sensitive demographic information in their\ntraining loss. Our code is available at: https://github.com/discovershu/DRO_COX\n", "date": "2022-11-18", "categories": []}
{"id": "2211.11206", "title": "Cultural Re-contextualization of Fairness Research in Language\n  Technologies in India", "abstract": "  Recent research has revealed undesirable biases in NLP data and models.\nHowever, these efforts largely focus on social disparities in the West, and are\nnot directly portable to other geo-cultural contexts. In this position paper,\nwe outline a holistic research agenda to re-contextualize NLP fairness research\nfor the Indian context, accounting for Indian societal context, bridging\ntechnological gaps in capability and resources, and adapting to Indian cultural\nvalues. We also summarize findings from an empirical study on various social\nbiases along different axes of disparities relevant to India, demonstrating\ntheir prevalence in corpora and models.\n", "date": "2022-11-21", "categories": []}
{"id": "2211.11512", "title": "Bursting the Burden Bubble? An Assessment of Sharma et al.'s\n  Counterfactual-based Fairness Metric", "abstract": "  Machine learning has seen an increase in negative publicity in recent years,\ndue to biased, unfair, and uninterpretable models. There is a rising interest\nin making machine learning models more fair for unprivileged communities, such\nas women or people of color. Metrics are needed to evaluate the fairness of a\nmodel. A novel metric for evaluating fairness between groups is Burden, which\nuses counterfactuals to approximate the average distance of negatively\nclassified individuals in a group to the decision boundary of the model. The\ngoal of this study is to compare Burden to statistical parity, a well-known\nfairness metric, and discover Burden's advantages and disadvantages. We do this\nby calculating the Burden and statistical parity of a sensitive attribute in\nthree datasets: two synthetic datasets are created to display differences\nbetween the two metrics, and one real-world dataset is used. We show that\nBurden can show unfairness where statistical parity can not, and that the two\nmetrics can even disagree on which group is treated unfairly. We conclude that\nBurden is a valuable metric, but does not replace statistical parity: it rather\nis valuable to use both.\n", "date": "2022-11-21", "categories": []}
{"id": "2209.10729", "title": "Fair Robust Active Learning by Joint Inconsistency", "abstract": "  Fairness and robustness play vital roles in trustworthy machine learning.\nObserving safety-critical needs in various annotation-expensive vision\napplications, we introduce a novel learning framework, Fair Robust Active\nLearning (FRAL), generalizing conventional active learning to fair and\nadversarial robust scenarios. This framework allows us to achieve standard and\nrobust minimax fairness with limited acquired labels. In FRAL, we then observe\nexisting fairness-aware data selection strategies suffer from either\nineffectiveness under severe data imbalance or inefficiency due to huge\ncomputations of adversarial training. To address these two problems, we develop\na novel Joint INconsistency (JIN) method exploiting prediction inconsistencies\nbetween benign and adversarial inputs as well as between standard and robust\nmodels. These two inconsistencies can be used to identify potential fairness\ngains and data imbalance mitigations. Thus, by performing label acquisition\nwith our inconsistency-based ranking metrics, we can alleviate the class\nimbalance issue and enhance minimax fairness with limited computation.\nExtensive experiments on diverse datasets and sensitive groups demonstrate that\nour method obtains the best results in standard and robust fairness under\nwhite-box PGD attacks compared with existing active data selection baselines.\n", "date": "2022-09-22", "categories": []}
{"id": "2207.13596", "title": "Fairness and Randomness in Machine Learning: Statistical Independence\n  and Relativization", "abstract": "  Fair Machine Learning endeavors to prevent unfairness arising in the context\nof machine learning applications embedded in society. Despite the variety of\ndefinitions of fairness and proposed \"fair algorithms\", there remain unresolved\nconceptual problems regarding fairness. In this paper, we dissect the role of\nstatistical independence in fairness and randomness notions regularly used in\nmachine learning. Thereby, we are led to a suprising hypothesis: randomness and\nfairness can be considered equivalent concepts in machine learning.\n  In particular, we obtain a relativized notion of randomness expressed as\nstatistical independence by appealing to Von Mises' century-old foundations for\nprobability. This notion turns out to be \"orthogonal\" in an abstract sense to\nthe commonly used i.i.d.-randomness. Using standard fairness notions in machine\nlearning, which are defined via statistical independence, we then link the ex\nante randomness assumptions about the data to the ex post requirements for fair\npredictions. This connection proves fruitful: we use it to argue that\nrandomness and fairness are essentially relative and that both concepts should\nreflect their nature as modeling assumptions in machine learning.\n", "date": "2022-07-27", "categories": []}
{"id": "2211.08742", "title": "Auditing Algorithmic Fairness in Machine Learning for Health with\n  Severity-Based LOGAN", "abstract": "  Auditing machine learning-based (ML) healthcare tools for bias is critical to\npreventing patient harm, especially in communities that disproportionately face\nhealth inequities. General frameworks are becoming increasingly available to\nmeasure ML fairness gaps between groups. However, ML for health (ML4H) auditing\nprinciples call for a contextual, patient-centered approach to model\nassessment. Therefore, ML auditing tools must be (1) better aligned with ML4H\nauditing principles and (2) able to illuminate and characterize communities\nvulnerable to the most harm. To address this gap, we propose supplementing ML4H\nauditing frameworks with SLOGAN (patient Severity-based LOcal Group biAs\ndetectioN), an automatic tool for capturing local biases in a clinical\nprediction task. SLOGAN adapts an existing tool, LOGAN (LOcal Group biAs\ndetectioN), by contextualizing group bias detection in patient illness severity\nand past medical history. We investigate and compare SLOGAN's bias detection\ncapabilities to LOGAN and other clustering techniques across patient subgroups\nin the MIMIC-III dataset. On average, SLOGAN identifies larger fairness\ndisparities in over 75% of patient groups than LOGAN while maintaining\nclustering quality. Furthermore, in a diabetes case study, health disparity\nliterature corroborates the characterizations of the most biased clusters\nidentified by SLOGAN. Our results contribute to the broader discussion of how\nmachine learning biases may perpetuate existing healthcare disparities.\n", "date": "2022-11-16", "categories": []}
{"id": "2207.04376", "title": "On Graph Neural Network Fairness in the Presence of Heterophilous\n  Neighborhoods", "abstract": "  We study the task of node classification for graph neural networks (GNNs) and\nestablish a connection between group fairness, as measured by statistical\nparity and equal opportunity, and local assortativity, i.e., the tendency of\nlinked nodes to have similar attributes. Such assortativity is often induced by\nhomophily, the tendency for nodes of similar properties to connect. Homophily\ncan be common in social networks where systemic factors have forced individuals\ninto communities which share a sensitive attribute. Through synthetic graphs,\nwe study the interplay between locally occurring homophily and fair\npredictions, finding that not all node neighborhoods are equal in this respect\n-- neighborhoods dominated by one category of a sensitive attribute often\nstruggle to obtain fair treatment, especially in the case of diverging local\nclass and sensitive attribute homophily. After determining that a relationship\nbetween local homophily and fairness exists, we investigate if the issue of\nunfairness can be associated to the design of the applied GNN model. We show\nthat by adopting heterophilous GNN designs capable of handling disassortative\ngroup labels, group fairness in locally heterophilous neighborhoods can be\nimproved by up to 25% over homophilous designs in real and synthetic datasets.\n", "date": "2022-07-10", "categories": []}
{"id": "2211.07530", "title": "A Survey on Preserving Fairness Guarantees in Changing Environments", "abstract": "  Human lives are increasingly being affected by the outcomes of automated\ndecision-making systems and it is essential for the latter to be, not only\naccurate, but also fair. The literature of algorithmic fairness has grown\nconsiderably over the last decade, where most of the approaches are evaluated\nunder the strong assumption that the train and test samples are independently\nand identically drawn from the same underlying distribution. However, in\npractice, dissimilarity between the training and deployment environments\nexists, which compromises the performance of the decision-making algorithm as\nwell as its fairness guarantees in the deployment data. There is an emergent\nresearch line that studies how to preserve fairness guarantees when the data\ngenerating processes differ between the source (train) and target (test)\ndomains, which is growing remarkably. With this survey, we aim to provide a\nwide and unifying overview on the topic. For such purpose, we propose a\ntaxonomy of the existing approaches for fair classification under distribution\nshift, highlight benchmarking alternatives, point out the relation with other\nsimilar research fields and eventually, identify future venues of research.\n", "date": "2022-11-14", "categories": []}
{"id": "2211.05809", "title": "Casual Conversations v2: Designing a large consent-driven dataset to\n  measure algorithmic bias and robustness", "abstract": "  Developing robust and fair AI systems require datasets with comprehensive set\nof labels that can help ensure the validity and legitimacy of relevant\nmeasurements. Recent efforts, therefore, focus on collecting person-related\ndatasets that have carefully selected labels, including sensitive\ncharacteristics, and consent forms in place to use those attributes for model\ntesting and development. Responsible data collection involves several stages,\nincluding but not limited to determining use-case scenarios, selecting\ncategories (annotations) such that the data are fit for the purpose of\nmeasuring algorithmic bias for subgroups and most importantly ensure that the\nselected categories/subcategories are robust to regional diversities and\ninclusive of as many subgroups as possible.\n  Meta, in a continuation of our efforts to measure AI algorithmic bias and\nrobustness\n(https://ai.facebook.com/blog/shedding-light-on-fairness-in-ai-with-a-new-data-set),\nis working on collecting a large consent-driven dataset with a comprehensive\nlist of categories. This paper describes our proposed design of such categories\nand subcategories for Casual Conversations v2.\n", "date": "2022-11-10", "categories": []}
{"id": "2211.06106", "title": "Identifying, measuring, and mitigating individual unfairness for\n  supervised learning models and application to credit risk models", "abstract": "  In the past few years, Artificial Intelligence (AI) has garnered attention\nfrom various industries including financial services (FS). AI has made a\npositive impact in financial services by enhancing productivity and improving\nrisk management. While AI can offer efficient solutions, it has the potential\nto bring unintended consequences. One such consequence is the pronounced effect\nof AI-related unfairness and attendant fairness-related harms. These\nfairness-related harms could involve differential treatment of individuals; for\nexample, unfairly denying a loan to certain individuals or groups of\nindividuals. In this paper, we focus on identifying and mitigating individual\nunfairness and leveraging some of the recently published techniques in this\ndomain, especially as applicable to the credit adjudication use case. We also\ninvestigate the extent to which techniques for achieving individual fairness\nare effective at achieving group fairness. Our main contribution in this work\nis functionalizing a two-step training process which involves learning a fair\nsimilarity metric from a group sense using a small portion of the raw data and\ntraining an individually \"fair\" classifier using the rest of the data where the\nsensitive features are excluded. The key characteristic of this two-step\ntechnique is related to its flexibility, i.e., the fair metric obtained in the\nfirst step can be used with any other individual fairness algorithms in the\nsecond step. Furthermore, we developed a second metric (distinct from the fair\nsimilarity metric) to determine how fairly a model is treating similar\nindividuals. We use this metric to compare a \"fair\" model against its baseline\nmodel in terms of their individual fairness value. Finally, some experimental\nresults corresponding to the individual unfairness mitigation techniques are\npresented.\n", "date": "2022-11-11", "categories": []}
{"id": "2211.06138", "title": "Practical Approaches for Fair Learning with Multitype and Multivariate\n  Sensitive Attributes", "abstract": "  It is important to guarantee that machine learning algorithms deployed in the\nreal world do not result in unfairness or unintended social consequences. Fair\nML has largely focused on the protection of single attributes in the simpler\nsetting where both attributes and target outcomes are binary. However, the\npractical application in many a real-world problem entails the simultaneous\nprotection of multiple sensitive attributes, which are often not simply binary,\nbut continuous or categorical. To address this more challenging task, we\nintroduce FairCOCCO, a fairness measure built on cross-covariance operators on\nreproducing kernel Hilbert Spaces. This leads to two practical tools: first,\nthe FairCOCCO Score, a normalised metric that can quantify fairness in settings\nwith single or multiple sensitive attributes of arbitrary type; and second, a\nsubsequent regularisation term that can be incorporated into arbitrary learning\nobjectives to obtain fair predictors. These contributions address crucial gaps\nin the algorithmic fairness literature, and we empirically demonstrate\nconsistent improvements against state-of-the-art techniques in balancing\npredictive power and fairness on real-world datasets.\n", "date": "2022-11-11", "categories": []}
{"id": "2211.06398", "title": "Investigating Fairness Disparities in Peer Review: A Language Model\n  Enhanced Approach", "abstract": "  Double-blind peer review mechanism has become the skeleton of academic\nresearch across multiple disciplines including computer science, yet several\nstudies have questioned the quality of peer reviews and raised concerns on\npotential biases in the process. In this paper, we conduct a thorough and\nrigorous study on fairness disparities in peer review with the help of large\nlanguage models (LMs). We collect, assemble, and maintain a comprehensive\nrelational database for the International Conference on Learning\nRepresentations (ICLR) conference from 2017 to date by aggregating data from\nOpenReview, Google Scholar, arXiv, and CSRanking, and extracting high-level\nfeatures using language models. We postulate and study fairness disparities on\nmultiple protective attributes of interest, including author gender, geography,\nauthor, and institutional prestige. We observe that the level of disparity\ndiffers and textual features are essential in reducing biases in the predictive\nmodeling. We distill several insights from our analysis on study the peer\nreview process with the help of large LMs. Our database also provides avenues\nfor studying new natural language processing (NLP) methods that facilitate the\nunderstanding of the peer review mechanism. We study a concrete example towards\nautomatic machine review systems and provide baseline models for the review\ngeneration and scoring tasks such that the database can be used as a benchmark.\n", "date": "2022-11-07", "categories": []}
{"id": "2211.04568", "title": "Towards Algorithmic Fairness in Space-Time: Filling in Black Holes", "abstract": "  New technologies and the availability of geospatial data have drawn attention\nto spatio-temporal biases present in society. For example: the COVID-19\npandemic highlighted disparities in the availability of broadband service and\nits role in the digital divide; the environmental justice movement in the\nUnited States has raised awareness to health implications for minority\npopulations stemming from historical redlining practices; and studies have\nfound varying quality and coverage in the collection and sharing of open-source\ngeospatial data. Despite the extensive literature on machine learning (ML)\nfairness, few algorithmic strategies have been proposed to mitigate such\nbiases. In this paper we highlight the unique challenges for quantifying and\naddressing spatio-temporal biases, through the lens of use cases presented in\nthe scientific literature and media. We envision a roadmap of ML strategies\nthat need to be developed or adapted to quantify and overcome these challenges\n-- including transfer learning, active learning, and reinforcement learning\ntechniques. Further, we discuss the potential role of ML in providing guidance\nto policy makers on issues related to spatial fairness.\n", "date": "2022-11-08", "categories": []}
{"id": "2211.04812", "title": "Discrimination and Class Imbalance Aware Online Naive Bayes", "abstract": "  Fairness-aware mining of massive data streams is a growing and challenging\nconcern in the contemporary domain of machine learning. Many stream learning\nalgorithms are used to replace humans at critical decision-making points e.g.,\nhiring staff, assessing credit risk, etc. This calls for handling massive\nincoming information with minimum response delay while ensuring fair and high\nquality decisions. Recent discrimination-aware learning methods are optimized\nbased on overall accuracy. However, the overall accuracy is biased in favor of\nthe majority class; therefore, state-of-the-art methods mainly diminish\ndiscrimination by partially or completely ignoring the minority class. In this\ncontext, we propose a novel adaptation of Na\\\"ive Bayes to mitigate\ndiscrimination embedded in the streams while maintaining high predictive\nperformance for both the majority and minority classes. Our proposed algorithm\nis simple, fast, and attains multi-objective optimization goals. To handle\nclass imbalance and concept drifts, a dynamic instance weighting module is\nproposed, which gives more importance to recent instances and less importance\nto obsolete instances based on their membership in minority or majority class.\nWe conducted experiments on a range of streaming and static datasets and\ndeduced that our proposed methodology outperforms existing state-of-the-art\nfairness-aware methods in terms of both discrimination score and balanced\naccuracy.\n", "date": "2022-11-09", "categories": []}
{"id": "2211.03994", "title": "Reinforcement Learning with Stepwise Fairness Constraints", "abstract": "  AI methods are used in societally important settings, ranging from credit to\nemployment to housing, and it is crucial to provide fairness in regard to\nalgorithmic decision making. Moreover, many settings are dynamic, with\npopulations responding to sequential decision policies. We introduce the study\nof reinforcement learning (RL) with stepwise fairness constraints, requiring\ngroup fairness at each time step. Our focus is on tabular episodic RL, and we\nprovide learning algorithms with strong theoretical guarantees in regard to\npolicy optimality and fairness violation. Our framework provides useful tools\nto study the impact of fairness constraints in sequential settings and brings\nup new challenges in RL.\n", "date": "2022-11-08", "categories": []}
{"id": "2211.04256", "title": "Bridging Fairness and Environmental Sustainability in Natural Language\n  Processing", "abstract": "  Fairness and environmental impact are important research directions for the\nsustainable development of artificial intelligence. However, while each topic\nis an active research area in natural language processing (NLP), there is a\nsurprising lack of research on the interplay between the two fields. This\nlacuna is highly problematic, since there is increasing evidence that an\nexclusive focus on fairness can actually hinder environmental sustainability,\nand vice versa. In this work, we shed light on this crucial intersection in NLP\nby (1) investigating the efficiency of current fairness approaches through\nsurveying example methods for reducing unfair stereotypical bias from the\nliterature, and (2) evaluating a common technique to reduce energy consumption\n(and thus environmental impact) of English NLP models, knowledge distillation\n(KD), for its impact on fairness. In this case study, we evaluate the effect of\nimportant KD factors, including layer and dimensionality reduction, with\nrespect to: (a) performance on the distillation task (natural language\ninference and semantic similarity prediction), and (b) multiple measures and\ndimensions of stereotypical bias (e.g., gender bias measured via the Word\nEmbedding Association Test). Our results lead us to clarify current assumptions\nregarding the effect of KD on unfair bias: contrary to other findings, we show\nthat KD can actually decrease model fairness.\n", "date": "2022-11-08", "categories": []}
{"id": "2211.04449", "title": "Fairness-aware Regression Robust to Adversarial Attacks", "abstract": "  In this paper, we take a first step towards answering the question of how to\ndesign fair machine learning algorithms that are robust to adversarial attacks.\nUsing a minimax framework, we aim to design an adversarially robust fair\nregression model that achieves optimal performance in the presence of an\nattacker who is able to add a carefully designed adversarial data point to the\ndataset or perform a rank-one attack on the dataset. By solving the proposed\nnonsmooth nonconvex-nonconcave minimax problem, the optimal adversary as well\nas the robust fairness-aware regression model are obtained. For both synthetic\ndata and real-world datasets, numerical results illustrate that the proposed\nadversarially robust fair models have better performance on poisoned datasets\nthan other fair machine learning models in both prediction accuracy and\ngroup-based fairness measure.\n", "date": "2022-11-04", "categories": []}
{"id": "2211.02730", "title": "Uncertainty-aware predictive modeling for fair data-driven decisions", "abstract": "  Both industry and academia have made considerable progress in developing\ntrustworthy and responsible machine learning (ML) systems. While critical\nconcepts like fairness and explainability are often addressed, the safety of\nsystems is typically not sufficiently taken into account. By viewing\ndata-driven decision systems as socio-technical systems, we draw on the\nuncertainty in ML literature to show how fairML systems can also be safeML\nsystems. We posit that a fair model needs to be an uncertainty-aware model,\ne.g. by drawing on distributional regression. For fair decisions, we argue that\na safe fail option should be used for individuals with uncertain\ncategorization. We introduce semi-structured deep distributional regression as\na modeling framework which addresses multiple concerns brought against standard\nML models and show its use in a real-world example of algorithmic profiling of\njob seekers.\n", "date": "2022-11-04", "categories": []}
{"id": "2211.02882", "title": "HERB: Measuring Hierarchical Regional Bias in Pre-trained Language\n  Models", "abstract": "  Fairness has become a trending topic in natural language processing (NLP),\nwhich addresses biases targeting certain social groups such as genders and\nreligions. However, regional bias in language models (LMs), a long-standing\nglobal discrimination problem, still remains unexplored. This paper bridges the\ngap by analysing the regional bias learned by the pre-trained language models\nthat are broadly used in NLP tasks. In addition to verifying the existence of\nregional bias in LMs, we find that the biases on regional groups can be\nstrongly influenced by the geographical clustering of the groups. We\naccordingly propose a HiErarchical Regional Bias evaluation method (HERB)\nutilising the information from the sub-region clusters to quantify the bias in\npre-trained LMs. Experiments show that our hierarchical metric can effectively\nevaluate the regional bias with respect to comprehensive topics and measure the\npotential regional bias that can be propagated to downstream tasks. Our codes\nare available at https://github.com/Bernard-Yang/HERB.\n", "date": "2022-11-05", "categories": []}
{"id": "2211.01253", "title": "Fair Visual Recognition via Intervention with Proxy Features", "abstract": "  Deep learning models often learn to make predictions that rely on sensitive\nsocial attributes like gender and race, which poses significant fairness risks,\nespecially in societal applications, e.g., hiring, banking, and criminal\njustice. Existing work tackles this issue by minimizing information about\nsocial attributes in models for debiasing. However, the high correlation\nbetween target task and social attributes makes bias mitigation incompatible\nwith target task accuracy. Recalling that model bias arises because the\nlearning of features in regard to bias attributes (i.e., bias features) helps\ntarget task optimization, we explore the following research question: \\emph{Can\nwe leverage proxy features to replace the role of bias feature in target task\noptimization for debiasing?} To this end, we propose \\emph{Proxy Debiasing}, to\nfirst transfer the target task's learning of bias information from bias\nfeatures to artificial proxy features, and then employ causal intervention to\neliminate proxy features in inference. The key idea of \\emph{Proxy Debiasing}\nis to design controllable proxy features to on one hand replace bias features\nin contributing to target task during the training stage, and on the other hand\neasily to be removed by intervention during the inference stage. This\nguarantees the elimination of bias features without affecting the target\ninformation, thus addressing the fairness-accuracy paradox in previous\ndebiasing solutions. We apply \\emph{Proxy Debiasing} to several benchmark\ndatasets, and achieve significant improvements over the state-of-the-art\ndebiasing methods in both of accuracy and fairness.\n", "date": "2022-11-02", "categories": []}
{"id": "2211.01333", "title": "Item-based Variational Auto-encoder for Fair Music Recommendation", "abstract": "  We present our solution for the EvalRS DataChallenge. The EvalRS\nDataChallenge aims to build a more realistic recommender system considering\naccuracy, fairness, and diversity in evaluation. Our proposed system is based\non an ensemble between an item-based variational auto-encoder (VAE) and a\nBayesian personalized ranking matrix factorization (BPRMF). To mitigate the\nbias in popularity, we use an item-based VAE for each popularity group with an\nadditional fairness regularization. To make a reasonable recommendation even\nthe predictions are inaccurate, we combine the recommended list of BPRMF and\nthat of item-based VAE. Through the experiments, we demonstrate that the\nitem-based VAE with fairness regularization significantly reduces popularity\nbias compared to the user-based VAE. The ensemble between the item-based VAE\nand BPRMF makes the top-1 item similar to the ground truth even the predictions\nare inaccurate. Finally, we propose a `Coefficient Variance based Fairness' as\na novel evaluation metric based on our reflections from the extensive\nexperiments.\n", "date": "2022-10-24", "categories": []}
{"id": "2211.00129", "title": "Is Facial Recognition Biased at Near-Infrared Spectrum As Well?", "abstract": "  Published academic research and media articles suggest face recognition is\nbiased across demographics. Specifically, unequal performance is obtained for\nwomen, dark-skinned people, and older adults. However, these published studies\nhave examined the bias of facial recognition in the visible spectrum (VIS).\nFactors such as facial makeup, facial hair, skin color, and illumination\nvariation have been attributed to the bias of this technology at the VIS. The\nnear-infrared (NIR) spectrum offers an advantage over the VIS in terms of\nrobustness to factors such as illumination changes, facial makeup, and skin\ncolor. Therefore, it is worthwhile to investigate the bias of facial\nrecognition at the near-infrared spectrum (NIR). This first study investigates\nthe bias of the face recognition systems at the NIR spectrum. To this aim, two\npopular NIR facial image datasets namely, CASIA-Face-Africa and Notre-Dame-NIVL\nconsisting of African and Caucasian subjects, respectively, are used to\ninvestigate the bias of facial recognition technology across gender and race.\nInterestingly, experimental results suggest equitable face recognition\nperformance across gender and race at the NIR spectrum.\n", "date": "2022-10-31", "categories": []}
{"id": "2211.00168", "title": "Improving Fairness in Image Classification via Sketching", "abstract": "  Fairness is a fundamental requirement for trustworthy and human-centered\nArtificial Intelligence (AI) system. However, deep neural networks (DNNs) tend\nto make unfair predictions when the training data are collected from different\nsub-populations with different attributes (i.e. color, sex, age), leading to\nbiased DNN predictions. We notice that such a troubling phenomenon is often\ncaused by data itself, which means that bias information is encoded to the DNN\nalong with the useful information (i.e. class information, semantic\ninformation). Therefore, we propose to use sketching to handle this phenomenon.\nWithout losing the utility of data, we explore the image-to-sketching methods\nthat can maintain useful semantic information for the target classification\nwhile filtering out the useless bias information. In addition, we design a fair\nloss to further improve the model fairness. We evaluate our method through\nextensive experiments on both general scene dataset and medical scene dataset.\nOur results show that the desired image-to-sketching method improves model\nfairness and achieves satisfactory results among state-of-the-art.\n", "date": "2022-10-31", "categories": []}
{"id": "2210.16590", "title": "Track2Vec: fairness music recommendation with a GPU-free\n  customizable-driven framework", "abstract": "  Recommendation systems have illustrated the significant progress made in\ncharacterizing users' preferences based on their past behaviors. Despite the\neffectiveness of recommending accurately, there exist several factors that are\nessential but unexplored for evaluating various facets of recommendation\nsystems, e.g., fairness, diversity, and limited resources. To address these\nissues, we propose Track2Vec, a GPU-free customizable-driven framework for\nfairness music recommendation. In order to take both accuracy and fairness into\naccount, our solution consists of three modules, a customized fairness-aware\ngroups for modeling different features based on configurable settings, a track\nrepresentation learning module for learning better user embedding, and an\nensemble module for ranking the recommendation results from different track\nrepresentation learning modules. Moreover, inspired by TF-IDF which has been\nwidely used in natural language processing, we introduce a metric called Miss\nRate - Inverse Ground Truth Frequency (MR-ITF) to measure the fairness.\nExtensive experiments demonstrate that our model achieves a 4th price ranking\nin a GPU-free environment on the leaderboard in the EvalRS @ CIKM 2022\nchallenge, which is superior to the official baseline by about 200% in terms of\nthe official scores. In addition, the ablation study illustrates the necessity\nof ensembling each group to acquire both accurate and fair recommendations.\n", "date": "2022-10-29", "categories": []}
{"id": "2210.15901", "title": "Mitigating Health Disparities in EHR via Deconfounder", "abstract": "  Health disparities, or inequalities between different patient demographics,\nare becoming crucial in medical decision-making, especially in Electronic\nHealth Record (EHR) predictive modeling. To ensure the fairness of sensitive\nattributes, conventional studies mainly adopt calibration or re-weighting\nmethods to balance the performance on among different demographic groups.\nHowever, we argue that these methods have some limitations. First, these\nmethods usually mean a trade-off between the model's performance and fairness.\nSecond, many methods completely attribute unfairness to the data collection\nprocess, which lacks substantial evidence. In this paper, we provide an\nempirical study to discover the possibility of using deconfounder to address\nthe disparity issue in healthcare. Our study can be summarized in two parts.\nThe first part is a pilot study demonstrating the exacerbation of disparity\nwhen unobserved confounders exist. The second part proposed a novel framework,\nParity Medical Deconfounder (PriMeD), to deal with the disparity issue in\nhealthcare datasets. Inspired by the deconfounder theory, PriMeD adopts a\nConditional Variational Autoencoder (CVAE) to learn latent factors (substitute\nconfounders) for observational data, and extensive experiments are provided to\nshow its effectiveness.\n", "date": "2022-10-28", "categories": []}
{"id": "2210.15183", "title": "Outlier-Aware Training for Improving Group Accuracy Disparities", "abstract": "  Methods addressing spurious correlations such as Just Train Twice (JTT,\narXiv:2107.09044v2) involve reweighting a subset of the training set to\nmaximize the worst-group accuracy. However, the reweighted set of examples may\npotentially contain unlearnable examples that hamper the model's learning. We\npropose mitigating this by detecting outliers to the training set and removing\nthem before reweighting. Our experiments show that our method achieves\ncompetitive or better accuracy compared with JTT and can detect and remove\nannotation errors in the subset being reweighted in JTT.\n", "date": "2022-10-27", "categories": []}
{"id": "2210.15230", "title": "How well can Text-to-Image Generative Models understand Ethical Natural\n  Language Interventions?", "abstract": "  Text-to-image generative models have achieved unprecedented success in\ngenerating high-quality images based on natural language descriptions. However,\nit is shown that these models tend to favor specific social groups when\nprompted with neutral text descriptions (e.g., 'a photo of a lawyer').\nFollowing Zhao et al. (2021), we study the effect on the diversity of the\ngenerated images when adding ethical intervention that supports equitable\njudgment (e.g., 'if all individuals can be a lawyer irrespective of their\ngender') in the input prompts. To this end, we introduce an Ethical NaTural\nLanguage Interventions in Text-to-Image GENeration (ENTIGEN) benchmark dataset\nto evaluate the change in image generations conditional on ethical\ninterventions across three social axes -- gender, skin color, and culture.\nThrough ENTIGEN framework, we find that the generations from minDALL.E,\nDALL.E-mini and Stable Diffusion cover diverse social groups while preserving\nthe image quality. Preliminary studies indicate that a large change in the\nmodel predictions is triggered by certain phrases such as 'irrespective of\ngender' in the context of gender bias in the ethical interventions. We release\ncode and annotated data at https://github.com/Hritikbansal/entigen_emnlp.\n", "date": "2022-10-27", "categories": []}
{"id": "2210.14552", "title": "A Robust Bias Mitigation Procedure Based on the Stereotype Content Model", "abstract": "  The Stereotype Content model (SCM) states that we tend to perceive minority\ngroups as cold, incompetent or both. In this paper we adapt existing work to\ndemonstrate that the Stereotype Content model holds for contextualised word\nembeddings, then use these results to evaluate a fine-tuning process designed\nto drive a language model away from stereotyped portrayals of minority groups.\nWe find the SCM terms are better able to capture bias than demographic agnostic\nterms related to pleasantness. Further, we were able to reduce the presence of\nstereotypes in the model through a simple fine-tuning procedure that required\nminimal human and computer resources, without harming downstream performance.\nWe present this work as a prototype of a debiasing procedure that aims to\nremove the need for a priori knowledge of the specifics of bias in the model.\n", "date": "2022-10-26", "categories": []}
{"id": "2210.07455", "title": "Controlling Bias Exposure for Fair Interpretable Predictions", "abstract": "  Recent work on reducing bias in NLP models usually focuses on protecting or\nisolating information related to a sensitive attribute (like gender or race).\nHowever, when sensitive information is semantically entangled with the task\ninformation of the input, e.g., gender information is predictive for a\nprofession, a fair trade-off between task performance and bias mitigation is\ndifficult to achieve. Existing approaches perform this trade-off by eliminating\nbias information from the latent space, lacking control over how much bias is\nnecessarily required to be removed. We argue that a favorable debiasing method\nshould use sensitive information 'fairly', rather than blindly eliminating it\n(Caliskan et al., 2017; Sun et al., 2019; Bogen et al., 2020). In this work, we\nprovide a novel debiasing algorithm by adjusting the predictive model's belief\nto (1) ignore the sensitive information if it is not useful for the task; (2)\nuse sensitive information minimally as necessary for the prediction (while also\nincurring a penalty). Experimental results on two text classification tasks\n(influenced by gender) and an open-ended generation task (influenced by race)\nindicate that our model achieves a desirable trade-off between debiasing and\ntask performance along with producing debiased rationales as evidence.\n", "date": "2022-10-14", "categories": []}
{"id": "2210.12546", "title": "Policy Optimization with Advantage Regularization for Long-Term Fairness\n  in Decision Systems", "abstract": "  Long-term fairness is an important factor of consideration in designing and\ndeploying learning-based decision systems in high-stake decision-making\ncontexts. Recent work has proposed the use of Markov Decision Processes (MDPs)\nto formulate decision-making with long-term fairness requirements in\ndynamically changing environments, and demonstrated major challenges in\ndirectly deploying heuristic and rule-based policies that worked well in static\nenvironments. We show that policy optimization methods from deep reinforcement\nlearning can be used to find strictly better decision policies that can often\nachieve both higher overall utility and less violation of the fairness\nrequirements, compared to previously-known strategies. In particular, we\npropose new methods for imposing fairness requirements in policy optimization\nby regularizing the advantage evaluation of different actions. Our proposed\nmethods make it easy to impose fairness constraints without reward engineering\nor sacrificing training efficiency. We perform detailed analyses in three\nestablished case studies, including attention allocation in incident\nmonitoring, bank loan approval, and vaccine distribution in population\nnetworks.\n", "date": "2022-10-22", "categories": []}
{"id": "2210.13182", "title": "Simultaneous Improvement of ML Model Fairness and Performance by\n  Identifying Bias in Data", "abstract": "  Machine learning models built on datasets containing discriminative instances\nattributed to various underlying factors result in biased and unfair outcomes.\nIt's a well founded and intuitive fact that existing bias mitigation strategies\noften sacrifice accuracy in order to ensure fairness. But when AI engine's\nprediction is used for decision making which reflects on revenue or operational\nefficiency such as credit risk modelling, it would be desirable by the business\nif accuracy can be somehow reasonably preserved. This conflicting requirement\nof maintaining accuracy and fairness in AI motivates our research. In this\npaper, we propose a fresh approach for simultaneous improvement of fairness and\naccuracy of ML models within a realistic paradigm. The essence of our work is a\ndata preprocessing technique that can detect instances ascribing a specific\nkind of bias that should be removed from the dataset before training and we\nfurther show that such instance removal will have no adverse impact on model\naccuracy. In particular, we claim that in the problem settings where instances\nexist with similar feature but different labels caused by variation in\nprotected attributes , an inherent bias gets induced in the dataset, which can\nbe identified and mitigated through our novel scheme. Our experimental\nevaluation on two open-source datasets demonstrates how the proposed method can\nmitigate bias along with improving rather than degrading accuracy, while\noffering certain set of control for end user.\n", "date": "2022-10-24", "categories": []}
{"id": "2206.01295", "title": "Rashomon Capacity: A Metric for Predictive Multiplicity in\n  Classification", "abstract": "  Predictive multiplicity occurs when classification models with statistically\nindistinguishable performances assign conflicting predictions to individual\nsamples. When used for decision-making in applications of consequence (e.g.,\nlending, education, criminal justice), models developed without regard for\npredictive multiplicity may result in unjustified and arbitrary decisions for\nspecific individuals. We introduce a new metric, called Rashomon Capacity, to\nmeasure predictive multiplicity in probabilistic classification. Prior metrics\nfor predictive multiplicity focus on classifiers that output thresholded (i.e.,\n0-1) predicted classes. In contrast, Rashomon Capacity applies to probabilistic\nclassifiers, capturing more nuanced score variations for individual samples. We\nprovide a rigorous derivation for Rashomon Capacity, argue its intuitive\nappeal, and demonstrate how to estimate it in practice. We show that Rashomon\nCapacity yields principled strategies for disclosing conflicting models to\nstakeholders. Our numerical experiments illustrate how Rashomon Capacity\ncaptures predictive multiplicity in various datasets and learning models,\nincluding neural networks. The tools introduced in this paper can help data\nscientists measure and report predictive multiplicity prior to model\ndeployment.\n", "date": "2022-06-02", "categories": []}
{"id": "2210.11471", "title": "Choose Your Lenses: Flaws in Gender Bias Evaluation", "abstract": "  Considerable efforts to measure and mitigate gender bias in recent years have\nled to the introduction of an abundance of tasks, datasets, and metrics used in\nthis vein. In this position paper, we assess the current paradigm of gender\nbias evaluation and identify several flaws in it. First, we highlight the\nimportance of extrinsic bias metrics that measure how a model's performance on\nsome task is affected by gender, as opposed to intrinsic evaluations of model\nrepresentations, which are less strongly connected to specific harms to people\ninteracting with systems. We find that only a few extrinsic metrics are\nmeasured in most studies, although more can be measured. Second, we find that\ndatasets and metrics are often coupled, and discuss how their coupling hinders\nthe ability to obtain reliable conclusions, and how one may decouple them. We\nthen investigate how the choice of the dataset and its composition, as well as\nthe choice of the metric, affect bias measurement, finding significant\nvariations across each of them. Finally, we propose several guidelines for more\nreliable gender bias evaluation.\n", "date": "2022-10-20", "categories": []}
{"id": "2207.12043", "title": "Representational Ethical Model Calibration", "abstract": "  Equity is widely held to be fundamental to the ethics of healthcare. In the\ncontext of clinical decision-making, it rests on the comparative fidelity of\nthe intelligence -- evidence-based or intuitive -- guiding the management of\neach individual patient. Though brought to recent attention by the\nindividuating power of contemporary machine learning, such epistemic equity\narises in the context of any decision guidance, whether traditional or\ninnovative. Yet no general framework for its quantification, let alone\nassurance, currently exists. Here we formulate epistemic equity in terms of\nmodel fidelity evaluated over learnt multi-dimensional representations of\nidentity crafted to maximise the captured diversity of the population,\nintroducing a comprehensive framework for Representational Ethical Model\nCalibration. We demonstrate use of the framework on large-scale multimodal data\nfrom UK Biobank to derive diverse representations of the population, quantify\nmodel performance, and institute responsive remediation. We offer our approach\nas a principled solution to quantifying and assuring epistemic equity in\nhealthcare, with applications across the research, clinical, and regulatory\ndomains.\n", "date": "2022-07-25", "categories": []}
{"id": "2210.10456", "title": "Group Fairness in Prediction-Based Decision Making: From Moral\n  Assessment to Implementation", "abstract": "  Ensuring fairness of prediction-based decision making is based on statistical\ngroup fairness criteria. Which one of these criteria is the morally most\nappropriate one depends on the context, and its choice requires an ethical\nanalysis. In this paper, we present a step-by-step procedure integrating three\nelements: (a) a framework for the moral assessment of what fairness means in a\ngiven context, based on the recently proposed general principle of \"Fair\nequality of chances\" (FEC) (b) a mapping of the assessment's results to\nestablished statistical group fairness criteria, and (c) a method for\nintegrating the thus-defined fairness into optimal decision making. As a second\ncontribution, we show new applications of the FEC principle and show that, with\nthis extension, the FEC framework covers all types of group fairness criteria:\nindependence, separation, and sufficiency. Third, we introduce an extended\nversion of the FEC principle, which additionally allows accounting for morally\nirrelevant elements of the fairness assessment and links to well-known\nrelaxations of the fairness criteria. This paper presents a framework to\ndevelop fair decision systems in a conceptually sound way, combining the moral\nand the computational elements of fair prediction-based decision-making in an\nintegrated approach. Data and code to reproduce our results are available at\nhttps://github.com/joebaumann/fair-prediction-based-decision-making.\n", "date": "2022-10-19", "categories": []}
{"id": "2210.10689", "title": "Towards Procedural Fairness: Uncovering Biases in How a Toxic Language\n  Classifier Uses Sentiment Information", "abstract": "  Previous works on the fairness of toxic language classifiers compare the\noutput of models with different identity terms as input features but do not\nconsider the impact of other important concepts present in the context. Here,\nbesides identity terms, we take into account high-level latent features learned\nby the classifier and investigate the interaction between these features and\nidentity terms. For a multi-class toxic language classifier, we leverage a\nconcept-based explanation framework to calculate the sensitivity of the model\nto the concept of sentiment, which has been used before as a salient feature\nfor toxic language detection. Our results show that although for some classes,\nthe classifier has learned the sentiment information as expected, this\ninformation is outweighed by the influence of identity terms as input features.\nThis work is a step towards evaluating procedural fairness, where unfair\nprocesses lead to unfair outcomes. The produced knowledge can guide debiasing\ntechniques to ensure that important concepts besides identity terms are\nwell-represented in training datasets.\n", "date": "2022-10-19", "categories": []}
{"id": "2210.09503", "title": "Towards Fair Classification against Poisoning Attacks", "abstract": "  Fair classification aims to stress the classification models to achieve the\nequality (treatment or prediction quality) among different sensitive groups.\nHowever, fair classification can be under the risk of poisoning attacks that\ndeliberately insert malicious training samples to manipulate the trained\nclassifiers' performance. In this work, we study the poisoning scenario where\nthe attacker can insert a small fraction of samples into training data, with\narbitrary sensitive attributes as well as other predictive features. We\ndemonstrate that the fairly trained classifiers can be greatly vulnerable to\nsuch poisoning attacks, with much worse accuracy & fairness trade-off, even\nwhen we apply some of the most effective defenses (originally proposed to\ndefend traditional classification tasks). As countermeasures to defend fair\nclassification tasks, we propose a general and theoretically guaranteed\nframework which accommodates traditional defense methods to fair classification\nagainst poisoning attacks. Through extensive experiments, the results validate\nthat the proposed defense framework obtains better robustness in terms of\naccuracy and fairness than representative baseline methods.\n", "date": "2022-10-18", "categories": []}
{"id": "2210.08758", "title": "Systematic Evaluation of Predictive Fairness", "abstract": "  Mitigating bias in training on biased datasets is an important open problem.\nSeveral techniques have been proposed, however the typical evaluation regime is\nvery limited, considering very narrow data conditions. For instance, the effect\nof target class imbalance and stereotyping is under-studied. To address this\ngap, we examine the performance of various debiasing methods across multiple\ntasks, spanning binary classification (Twitter sentiment), multi-class\nclassification (profession prediction), and regression (valence prediction).\nThrough extensive experimentation, we find that data conditions have a strong\ninfluence on relative model performance, and that general conclusions cannot be\ndrawn about method efficacy when evaluating only on standard datasets, as is\ncurrent practice in fairness research.\n", "date": "2022-10-17", "categories": []}
{"id": "2210.02015", "title": "Conformalized Fairness via Quantile Regression", "abstract": "  Algorithmic fairness has received increased attention in socially sensitive\ndomains. While rich literature on mean fairness has been established, research\non quantile fairness remains sparse but vital. To fulfill great needs and\nadvocate the significance of quantile fairness, we propose a novel framework to\nlearn a real-valued quantile function under the fairness requirement of\nDemographic Parity with respect to sensitive attributes, such as race or\ngender, and thereby derive a reliable fair prediction interval. Using optimal\ntransport and functional synchronization techniques, we establish theoretical\nguarantees of distribution-free coverage and exact fairness for the induced\nprediction interval constructed by fair quantiles. A hands-on pipeline is\nprovided to incorporate flexible quantile regressions with an efficient\nfairness adjustment post-processing algorithm. We demonstrate the superior\nempirical performance of this approach on several benchmark datasets. Our\nresults show the model's ability to uncover the mechanism underlying the\nfairness-accuracy trade-off in a wide range of societal and medical\napplications.\n", "date": "2022-10-05", "categories": []}
{"id": "2210.07755", "title": "Simpson's Paradox in Recommender Fairness: Reconciling differences\n  between per-user and aggregated evaluations", "abstract": "  There has been a flurry of research in recent years on notions of fairness in\nranking and recommender systems, particularly on how to evaluate if a\nrecommender allocates exposure equally across groups of relevant items (also\nknown as provider fairness). While this research has laid an important\nfoundation, it gave rise to different approaches depending on whether relevant\nitems are compared per-user/per-query or aggregated across users. Despite both\nbeing established and intuitive, we discover that these two notions can lead to\nopposite conclusions, a form of Simpson's Paradox. We reconcile these notions\nand show that the tension is due to differences in distributions of users where\nitems are relevant, and break down the important factors of the user's\nrecommendations. Based on this new understanding, practitioners might be\ninterested in either notions, but might face challenges with the per-user\nmetric due to partial observability of the relevance and user satisfaction,\ntypical in real-world recommenders. We describe a technique based on\ndistribution matching to estimate it in such a scenario. We demonstrate on\nsimulated and real-world recommender data the effectiveness and usefulness of\nsuch an approach.\n", "date": "2022-10-14", "categories": []}
{"id": "2210.06630", "title": "Fairness via Adversarial Attribute Neighbourhood Robust Learning", "abstract": "  Improving fairness between privileged and less-privileged sensitive attribute\ngroups (e.g, {race, gender}) has attracted lots of attention. To enhance the\nmodel performs uniformly well in different sensitive attributes, we propose a\nprincipled \\underline{R}obust \\underline{A}dversarial \\underline{A}ttribute\n\\underline{N}eighbourhood (RAAN) loss to debias the classification head and\npromote a fairer representation distribution across different sensitive\nattribute groups. The key idea of RAAN is to mitigate the differences of biased\nrepresentations between different sensitive attribute groups by assigning each\nsample an adversarial robust weight, which is defined on the representations of\nadversarial attribute neighbors, i.e, the samples from different protected\ngroups. To provide efficient optimization algorithms, we cast the RAAN into a\nsum of coupled compositional functions and propose a stochastic adaptive\n(Adam-style) and non-adaptive (SGD-style) algorithm framework SCRAAN with\nprovable theoretical guarantee. Extensive empirical studies on fairness-related\nbenchmark datasets verify the effectiveness of the proposed method.\n", "date": "2022-10-12", "categories": []}
{"id": "2210.06680", "title": "Walk a Mile in Their Shoes: a New Fairness Criterion for Machine\n  Learning", "abstract": "  The old empathetic adage, ``Walk a mile in their shoes,'' asks that one\nimagine the difficulties others may face. This suggests a new ML counterfactual\nfairness criterion, based on a \\textit{group} level: How would members of a\nnonprotected group fare if their group were subject to conditions in some\nprotected group? Instead of asking what sentence would a particular Caucasian\nconvict receive if he were Black, take that notion to entire groups; e.g. how\nwould the average sentence for all White convicts change if they were Black,\nbut with their same White characteristics, e.g. same number of prior\nconvictions? We frame the problem and study it empirically, for different\ndatasets. Our approach also is a solution to the problem of covariate\ncorrelation with sensitive attributes.\n", "date": "2022-10-13", "categories": []}
{"id": "2210.05936", "title": "Equal Experience in Recommender Systems", "abstract": "  We explore the fairness issue that arises in recommender systems. Biased data\ndue to inherent stereotypes of particular groups (e.g., male students' average\nrating on mathematics is often higher than that on humanities, and vice versa\nfor females) may yield a limited scope of suggested items to a certain group of\nusers. Our main contribution lies in the introduction of a novel fairness\nnotion (that we call equal experience), which can serve to regulate such\nunfairness in the presence of biased data. The notion captures the degree of\nthe equal experience of item recommendations across distinct groups. We propose\nan optimization framework that incorporates the fairness notion as a\nregularization term, as well as introduce computationally-efficient algorithms\nthat solve the optimization. Experiments on synthetic and benchmark real\ndatasets demonstrate that the proposed framework can indeed mitigate such\nunfairness while exhibiting a minor degradation of recommendation accuracy.\n", "date": "2022-10-12", "categories": []}
{"id": "2210.06288", "title": "fAux: Testing Individual Fairness via Gradient Alignment", "abstract": "  Machine learning models are vulnerable to biases that result in unfair\ntreatment of individuals from different populations. Recent work that aims to\ntest a model's fairness at the individual level either relies on domain\nknowledge to choose metrics, or on input transformations that risk generating\nout-of-domain samples. We describe a new approach for testing individual\nfairness that does not have either requirement. We propose a novel criterion\nfor evaluating individual fairness and develop a practical testing method based\non this criterion which we call fAux (pronounced fox). This is based on\ncomparing the derivatives of the predictions of the model to be tested with\nthose of an auxiliary model, which predicts the protected variable from the\nobserved data. We show that the proposed method effectively identifies\ndiscrimination on both synthetic and real-world datasets, and has quantitative\nand qualitative advantages over contemporary methods.\n", "date": "2022-10-10", "categories": []}
{"id": "2210.06351", "title": "A Keyword Based Approach to Understanding the Overpenalization of\n  Marginalized Groups by English Marginal Abuse Models on Twitter", "abstract": "  Harmful content detection models tend to have higher false positive rates for\ncontent from marginalized groups. In the context of marginal abuse modeling on\nTwitter, such disproportionate penalization poses the risk of reduced\nvisibility, where marginalized communities lose the opportunity to voice their\nopinion on the platform. Current approaches to algorithmic harm mitigation, and\nbias detection for NLP models are often very ad hoc and subject to human bias.\nWe make two main contributions in this paper. First, we design a novel\nmethodology, which provides a principled approach to detecting and measuring\nthe severity of potential harms associated with a text-based model. Second, we\napply our methodology to audit Twitter's English marginal abuse model, which is\nused for removing amplification eligibility of marginally abusive content.\nWithout utilizing demographic labels or dialect classifiers, we are still able\nto detect and measure the severity of issues related to the over-penalization\nof the speech of marginalized communities, such as the use of reclaimed speech,\ncounterspeech, and identity related terms. In order to mitigate the associated\nharms, we experiment with adding additional true negative examples and find\nthat doing so provides improvements to our fairness metrics without large\ndegradations in model performance.\n", "date": "2022-10-07", "categories": []}
{"id": "2210.04995", "title": "FEAMOE: Fair, Explainable and Adaptive Mixture of Experts", "abstract": "  Three key properties that are desired of trustworthy machine learning models\ndeployed in high-stakes environments are fairness, explainability, and an\nability to account for various kinds of \"drift\". While drifts in model\naccuracy, for example due to covariate shift, have been widely investigated,\ndrifts in fairness metrics over time remain largely unexplored. In this paper,\nwe propose FEAMOE, a novel \"mixture-of-experts\" inspired framework aimed at\nlearning fairer, more explainable/interpretable models that can also rapidly\nadjust to drifts in both the accuracy and the fairness of a classifier. We\nillustrate our framework for three popular fairness measures and demonstrate\nhow drift can be handled with respect to these fairness constraints.\nExperiments on multiple datasets show that our framework as applied to a\nmixture of linear experts is able to perform comparably to neural networks in\nterms of accuracy while producing fairer models. We then use the large-scale\nHMDA dataset and show that while various models trained on HMDA demonstrate\ndrift with respect to both accuracy and fairness, FEAMOE can ably handle these\ndrifts with respect to all the considered fairness measures and maintain model\naccuracy as well. We also prove that the proposed framework allows for\nproducing fast Shapley value explanations, which makes computationally\nefficient feature attribution based explanations of model decisions readily\navailable via FEAMOE.\n", "date": "2022-10-10", "categories": []}
{"id": "2210.05366", "title": "Race Bias Analysis of Bona Fide Errors in face anti-spoofing", "abstract": "  The study of bias in Machine Learning is receiving a lot of attention in\nrecent years, however, few only papers deal explicitly with the problem of race\nbias in face anti-spoofing. In this paper, we present a systematic study of\nrace bias in face anti-spoofing with three key characteristics: the focus is on\nanalysing potential bias in the bona fide errors, where significant ethical and\nlegal issues lie; the analysis is not restricted to the final binary outcomes\nof the classifier, but also covers the classifier's scalar responses and its\nlatent space; the threshold determining the operating point of the classifier\nis considered a variable. We demonstrate the proposed bias analysis process on\na VQ-VAE based face anti-spoofing algorithm, trained on the Replay Attack and\nthe Spoof in the Wild (SiW) databases, and analysed for bias on the SiW and\nRacial Faces in the Wild (RFW), databases. The results demonstrate that race\nbias is not necessarily the result of different mean response values among the\nvarious populations. Instead, it can be better understood as the combined\neffect of several possible characteristics of the response distributions:\ndifferent means; different variances; bimodal behaviour; existence of outliers.\n", "date": "2022-10-11", "categories": []}
{"id": "2210.05594", "title": "Navigating Ensemble Configurations for Algorithmic Fairness", "abstract": "  Bias mitigators can improve algorithmic fairness in machine learning models,\nbut their effect on fairness is often not stable across data splits. A popular\napproach to train more stable models is ensemble learning, but unfortunately,\nit is unclear how to combine ensembles with mitigators to best navigate\ntrade-offs between fairness and predictive performance. To that end, we built\nan open-source library enabling the modular composition of 8 mitigators, 4\nensembles, and their corresponding hyperparameters, and we empirically explored\nthe space of configurations on 13 datasets. We distilled our insights from this\nexploration in the form of a guidance diagram for practitioners that we\ndemonstrate is robust and reproducible.\n", "date": "2022-10-11", "categories": []}
{"id": "2210.03826", "title": "An Analysis of the Effects of Decoding Algorithms on Fairness in\n  Open-Ended Language Generation", "abstract": "  Several prior works have shown that language models (LMs) can generate text\ncontaining harmful social biases and stereotypes. While decoding algorithms\nplay a central role in determining properties of LM generated text, their\nimpact on the fairness of the generations has not been studied. We present a\nsystematic analysis of the impact of decoding algorithms on LM fairness, and\nanalyze the trade-off between fairness, diversity and quality. Our experiments\nwith top-$p$, top-$k$ and temperature decoding algorithms, in open-ended\nlanguage generation, show that fairness across demographic groups changes\nsignificantly with change in decoding algorithm's hyper-parameters. Notably,\ndecoding algorithms that output more diverse text also output more texts with\nnegative sentiment and regard. We present several findings and provide\nrecommendations on standardized reporting of decoding details in fairness\nevaluations and optimization of decoding algorithms for fairness alongside\nquality and diversity.\n", "date": "2022-10-07", "categories": []}
{"id": "2210.04337", "title": "Quantifying Social Biases Using Templates is Unreliable", "abstract": "  Recently, there has been an increase in efforts to understand how large\nlanguage models (LLMs) propagate and amplify social biases. Several works have\nutilized templates for fairness evaluation, which allow researchers to quantify\nsocial biases in the absence of test sets with protected attribute labels.\nWhile template evaluation can be a convenient and helpful diagnostic tool to\nunderstand model deficiencies, it often uses a simplistic and limited set of\ntemplates. In this paper, we study whether bias measurements are sensitive to\nthe choice of templates used for benchmarking. Specifically, we investigate the\ninstability of bias measurements by manually modifying templates proposed in\nprevious works in a semantically-preserving manner and measuring bias across\nthese modifications. We find that bias values and resulting conclusions vary\nconsiderably across template modifications on four tasks, ranging from an 81%\nreduction (NLI) to a 162% increase (MLM) in (task-specific) bias measurements.\nOur results indicate that quantifying fairness in LLMs, as done in current\npractice, can be brittle and needs to be approached with more care and caution.\n", "date": "2022-10-09", "categories": []}
{"id": "2210.04369", "title": "A Differentiable Distance Approximation for Fairer Image Classification", "abstract": "  Naively trained AI models can be heavily biased. This can be particularly\nproblematic when the biases involve legally or morally protected attributes\nsuch as ethnic background, age or gender. Existing solutions to this problem\ncome at the cost of extra computation, unstable adversarial optimisation or\nhave losses on the feature space structure that are disconnected from fairness\nmeasures and only loosely generalise to fairness. In this work we propose a\ndifferentiable approximation of the variance of demographics, a metric that can\nbe used to measure the bias, or unfairness, in an AI model. Our approximation\ncan be optimised alongside the regular training objective which eliminates the\nneed for any extra models during training and directly improves the fairness of\nthe regularised models. We demonstrate that our approach improves the fairness\nof AI models in varied task and dataset scenarios, whilst still maintaining a\nhigh level of classification accuracy. Code is available at\nhttps://bitbucket.org/nelliottrosa/base_fairness.\n", "date": "2022-10-09", "categories": []}
{"id": "2210.04491", "title": "A survey of Identification and mitigation of Machine Learning\n  algorithmic biases in Image Analysis", "abstract": "  The problem of algorithmic bias in machine learning has gained a lot of\nattention in recent years due to its concrete and potentially hazardous\nimplications in society. In much the same manner, biases can also alter modern\nindustrial and safety-critical applications where machine learning are based on\nhigh dimensional inputs such as images. This issue has however been mostly left\nout of the spotlight in the machine learning literature. Contrarily to societal\napplications where a set of proxy variables can be provided by the common sense\nor by regulations to draw the attention on potential risks, industrial and\nsafety-critical applications are most of the times sailing blind. The variables\nrelated to undesired biases can indeed be indirectly represented in the input\ndata, or can be unknown, thus making them harder to tackle. This raises serious\nand well-founded concerns towards the commercial deployment of AI-based\nsolutions, especially in a context where new regulations clearly address the\nissues opened by undesired biases in AI. Consequently, we propose here to make\nan overview of recent advances in this area, firstly by presenting how such\nbiases can demonstrate themselves, then by exploring different ways to bring\nthem to light, and by probing different possibilities to mitigate them. We\nfinally present a practical remote sensing use-case of industrial Fairness.\n", "date": "2022-10-10", "categories": []}
{"id": "2210.02516", "title": "Equalizing Credit Opportunity in Algorithms: Aligning Algorithmic\n  Fairness Research with U.S. Fair Lending Regulation", "abstract": "  Credit is an essential component of financial wellbeing in America, and\nunequal access to it is a large factor in the economic disparities between\ndemographic groups that exist today. Today, machine learning algorithms,\nsometimes trained on alternative data, are increasingly being used to determine\naccess to credit, yet research has shown that machine learning can encode many\ndifferent versions of \"unfairness,\" thus raising the concern that banks and\nother financial institutions could -- potentially unwittingly -- engage in\nillegal discrimination through the use of this technology. In the US, there are\nlaws in place to make sure discrimination does not happen in lending and\nagencies charged with enforcing them. However, conversations around fair credit\nmodels in computer science and in policy are often misaligned: fair machine\nlearning research often lacks legal and practical considerations specific to\nexisting fair lending policy, and regulators have yet to issue new guidance on\nhow, if at all, credit risk models should be utilizing practices and techniques\nfrom the research community. This paper aims to better align these sides of the\nconversation. We describe the current state of credit discrimination regulation\nin the United States, contextualize results from fair ML research to identify\nthe specific fairness concerns raised by the use of machine learning in\nlending, and discuss regulatory opportunities to address these concerns.\n", "date": "2022-10-05", "categories": []}
{"id": "2207.05727", "title": "Enhancing Fairness of Visual Attribute Predictors", "abstract": "  The performance of deep neural networks for image recognition tasks such as\npredicting a smiling face is known to degrade with under-represented classes of\nsensitive attributes. We address this problem by introducing fairness-aware\nregularization losses based on batch estimates of Demographic Parity, Equalized\nOdds, and a novel Intersection-over-Union measure. The experiments performed on\nfacial and medical images from CelebA, UTKFace, and the SIIM-ISIC melanoma\nclassification challenge show the effectiveness of our proposed fairness losses\nfor bias mitigation as they improve model fairness while maintaining high\nclassification performance. To the best of our knowledge, our work is the first\nattempt to incorporate these types of losses in an end-to-end training scheme\nfor mitigating biases of visual attribute predictors. Our code is available at\nhttps://github.com/nish03/FVAP.\n", "date": "2022-07-07", "categories": []}
{"id": "2209.15550", "title": "The More Secure, The Less Equally Usable: Gender and Ethnicity\n  (Un)fairness of Deep Face Recognition along Security Thresholds", "abstract": "  Face biometrics are playing a key role in making modern smart city\napplications more secure and usable. Commonly, the recognition threshold of a\nface recognition system is adjusted based on the degree of security for the\nconsidered use case. The likelihood of a match can be for instance decreased by\nsetting a high threshold in case of a payment transaction verification. Prior\nwork in face recognition has unfortunately showed that error rates are usually\nhigher for certain demographic groups. These disparities have hence brought\ninto question the fairness of systems empowered with face biometrics. In this\npaper, we investigate the extent to which disparities among demographic groups\nchange under different security levels. Our analysis includes ten face\nrecognition models, three security thresholds, and six demographic groups based\non gender and ethnicity. Experiments show that the higher the security of the\nsystem is, the higher the disparities in usability among demographic groups\nare. Compelling unfairness issues hence exist and urge countermeasures in\nreal-world high-stakes environments requiring severe security levels.\n", "date": "2022-09-30", "categories": []}
{"id": "2209.14670", "title": "Towards Equalised Odds as Fairness Metric in Academic Performance\n  Prediction", "abstract": "  The literature for fairness-aware machine learning knows a plethora of\ndifferent fairness notions. It is however wellknown, that it is impossible to\nsatisfy all of them, as certain notions contradict each other. In this paper,\nwe take a closer look at academic performance prediction (APP) systems and try\nto distil which fairness notions suit this task most. For this, we scan recent\nliterature proposing guidelines as to which fairness notion to use and apply\nthese guidelines onto APP. Our findings suggest equalised odds as most suitable\nnotion for APP, based on APP's WYSIWYG worldview as well as potential long-term\nimprovements for the population.\n", "date": "2022-09-29", "categories": []}
{"id": "2209.13939", "title": "Racial Bias in the Beautyverse", "abstract": "  This short paper proposes a preliminary and yet insightful investigation of\nracial biases in beauty filters techniques currently used on social media. The\nobtained results are a call to action for researchers in Computer Vision: such\nbiases risk being replicated and exaggerated in the Metaverse and, as a\nconsequence, they deserve more attention from the community.\n", "date": "2022-09-28", "categories": []}
{"id": "2209.14237", "title": "Accuracy, Fairness, and Interpretability of Machine Learning Criminal\n  Recidivism Models", "abstract": "  Criminal recidivism models are tools that have gained widespread adoption by\nparole boards across the United States to assist with parole decisions. These\nmodels take in large amounts of data about an individual and then predict\nwhether an individual would commit a crime if released on parole. Although such\nmodels are not the only or primary factor in making the final parole decision,\nquestions have been raised about their accuracy, fairness, and\ninterpretability. In this paper, various machine learning-based criminal\nrecidivism models are created based on a real-world parole decision dataset\nfrom the state of Georgia in the United States. The recidivism models are\ncomparatively evaluated for their accuracy, fairness, and interpretability. It\nis found that there are noted differences and trade-offs between accuracy,\nfairness, and being inherently interpretable. Therefore, choosing the best\nmodel depends on the desired balance between accuracy, fairness, and\ninterpretability, as no model is perfect or consistently the best across\ndifferent criteria.\n", "date": "2022-09-14", "categories": []}
{"id": "2209.13019", "title": "Fast online ranking with fairness of exposure", "abstract": "  As recommender systems become increasingly central for sorting and\nprioritizing the content available online, they have a growing impact on the\nopportunities or revenue of their items producers. For instance, they influence\nwhich recruiter a resume is recommended to, or to whom and how much a music\ntrack, video or news article is being exposed. This calls for recommendation\napproaches that not only maximize (a proxy of) user satisfaction, but also\nconsider some notion of fairness in the exposure of items or groups of items.\nFormally, such recommendations are usually obtained by maximizing a concave\nobjective function in the space of randomized rankings. When the total exposure\nof an item is defined as the sum of its exposure over users, the optimal\nrankings of every users become coupled, which makes the optimization process\nchallenging. Existing approaches to find these rankings either solve the global\noptimization problem in a batch setting, i.e., for all users at once, which\nmakes them inapplicable at scale, or are based on heuristics that have weak\ntheoretical guarantees. In this paper, we propose the first efficient online\nalgorithm to optimize concave objective functions in the space of rankings\nwhich applies to every concave and smooth objective function, such as the ones\nfound for fairness of exposure. Based on online variants of the Frank-Wolfe\nalgorithm, we show that our algorithm is computationally fast, generating\nrankings on-the-fly with computation cost dominated by the sort operation,\nmemory efficient, and has strong theoretical guarantees. Compared to baseline\npolicies that only maximize user-side performance, our algorithm allows to\nincorporate complex fairness of exposure criteria in the recommendations with\nnegligible computational overhead.\n", "date": "2022-09-13", "categories": []}
{"id": "2209.13179", "title": "Explainable Global Fairness Verification of Tree-Based Classifiers", "abstract": "  We present a new approach to the global fairness verification of tree-based\nclassifiers. Given a tree-based classifier and a set of sensitive features\npotentially leading to discrimination, our analysis synthesizes sufficient\nconditions for fairness, expressed as a set of traditional propositional logic\nformulas, which are readily understandable by human experts. The verified\nfairness guarantees are global, in that the formulas predicate over all the\npossible inputs of the classifier, rather than just a few specific test\ninstances. Our analysis is formally proved both sound and complete.\nExperimental results on public datasets show that the analysis is precise,\nexplainable to human experts and efficient enough for practical adoption.\n", "date": "2022-09-27", "categories": []}
{"id": "2209.11762", "title": "Towards Auditing Unsupervised Learning Algorithms and Human Processes\n  For Fairness", "abstract": "  Existing work on fairness typically focuses on making known machine learning\nalgorithms fairer. Fair variants of classification, clustering, outlier\ndetection and other styles of algorithms exist. However, an understudied area\nis the topic of auditing an algorithm's output to determine fairness. Existing\nwork has explored the two group classification problem for binary protected\nstatus variables using standard definitions of statistical parity. Here we\nbuild upon the area of auditing by exploring the multi-group setting under more\ncomplex definitions of fairness.\n", "date": "2022-09-20", "categories": []}
{"id": "2209.11837", "title": "Doubly Fair Dynamic Pricing", "abstract": "  We study the problem of online dynamic pricing with two types of fairness\nconstraints: a \"procedural fairness\" which requires the proposed prices to be\nequal in expectation among different groups, and a \"substantive fairness\" which\nrequires the accepted prices to be equal in expectation among different groups.\nA policy that is simultaneously procedural and substantive fair is referred to\nas \"doubly fair\". We show that a doubly fair policy must be random to have\nhigher revenue than the best trivial policy that assigns the same price to\ndifferent groups. In a two-group setting, we propose an online learning\nalgorithm for the 2-group pricing problems that achieves $\\tilde{O}(\\sqrt{T})$\nregret, zero procedural unfairness and $\\tilde{O}(\\sqrt{T})$ substantive\nunfairness over $T$ rounds of learning. We also prove two lower bounds showing\nthat these results on regret and unfairness are both information-theoretically\noptimal up to iterated logarithmic factors. To the best of our knowledge, this\nis the first dynamic pricing algorithm that learns to price while satisfying\ntwo fairness constraints at the same time.\n", "date": "2022-09-23", "categories": []}
{"id": "2209.10860", "title": "SCALES: From Fairness Principles to Constrained Decision-Making", "abstract": "  This paper proposes SCALES, a general framework that translates\nwell-established fairness principles into a common representation based on the\nConstraint Markov Decision Process (CMDP). With the help of causal language,\nour framework can place constraints on both the procedure of decision making\n(procedural fairness) as well as the outcomes resulting from decisions (outcome\nfairness). Specifically, we show that well-known fairness principles can be\nencoded either as a utility component, a non-causal component, or a causal\ncomponent in a SCALES-CMDP. We illustrate SCALES using a set of case studies\ninvolving a simulated healthcare scenario and the real-world COMPAS dataset.\nExperiments demonstrate that our framework produces fair policies that embody\nalternative fairness principles in single-step and sequential decision-making\nscenarios.\n", "date": "2022-09-22", "categories": []}
{"id": "2209.09592", "title": "Closing the Gender Wage Gap: Adversarial Fairness in Job Recommendation", "abstract": "  The goal of this work is to help mitigate the already existing gender wage\ngap by supplying unbiased job recommendations based on resumes from job\nseekers. We employ a generative adversarial network to remove gender bias from\nword2vec representations of 12M job vacancy texts and 900k resumes. Our results\nshow that representations created from recruitment texts contain algorithmic\nbias and that this bias results in real-world consequences for recommendation\nsystems. Without controlling for bias, women are recommended jobs with\nsignificantly lower salary in our data. With adversarially fair\nrepresentations, this wage gap disappears, meaning that our debiased job\nrecommendations reduce wage discrimination. We conclude that adversarial\ndebiasing of word representations can increase real-world fairness of systems\nand thus may be part of the solution for creating fairness-aware recommendation\nsystems.\n", "date": "2022-09-20", "categories": []}
{"id": "2209.09975", "title": "Power of Explanations: Towards automatic debiasing in hate speech\n  detection", "abstract": "  Hate speech detection is a common downstream application of natural language\nprocessing (NLP) in the real world. In spite of the increasing accuracy,\ncurrent data-driven approaches could easily learn biases from the imbalanced\ndata distributions originating from humans. The deployment of biased models\ncould further enhance the existing social biases. But unlike handling tabular\ndata, defining and mitigating biases in text classifiers, which deal with\nunstructured data, are more challenging. A popular solution for improving\nmachine learning fairness in NLP is to conduct the debiasing process with a\nlist of potentially discriminated words given by human annotators. In addition\nto suffering from the risks of overlooking the biased terms, exhaustively\nidentifying bias with human annotators are unsustainable since discrimination\nis variable among different datasets and may evolve over time. To this end, we\npropose an automatic misuse detector (MiD) relying on an explanation method for\ndetecting potential bias. And built upon that, an end-to-end debiasing\nframework with the proposed staged correction is designed for text classifiers\nwithout any external resources required.\n", "date": "2022-09-07", "categories": []}
{"id": "2208.01755", "title": "Debiasing Gender Bias in Information Retrieval Models", "abstract": "  Biases in culture, gender, ethnicity, etc. have existed for decades and have\naffected many areas of human social interaction. These biases have been shown\nto impact machine learning (ML) models, and for natural language processing\n(NLP), this can have severe consequences for downstream tasks. Mitigating\ngender bias in information retrieval (IR) is important to avoid propagating\nstereotypes. In this work, we employ a dataset consisting of two components:\n(1) relevance of a document to a query and (2) \"gender\" of a document, in which\npronouns are replaced by male, female, and neutral conjugations. We\ndefinitively show that pre-trained models for IR do not perform well in\nzero-shot retrieval tasks when full fine-tuning of a large pre-trained BERT\nencoder is performed and that lightweight fine-tuning performed with adapter\nnetworks improves zero-shot retrieval performance almost by 20% over baseline.\nWe also illustrate that pre-trained models have gender biases that result in\nretrieved articles tending to be more often male than female. We overcome this\nby introducing a debiasing technique that penalizes the model when it prefers\nmales over females, resulting in an effective model that retrieves articles in\na balanced fashion across genders.\n", "date": "2022-08-02", "categories": []}
{"id": "2209.08321", "title": "Enhanced Fairness Testing via Generating Effective Initial Individual\n  Discriminatory Instances", "abstract": "  Fairness testing aims at mitigating unintended discrimination in the\ndecision-making process of data-driven AI systems. Individual discrimination\nmay occur when an AI model makes different decisions for two distinct\nindividuals who are distinguishable solely according to protected attributes,\nsuch as age and race. Such instances reveal biased AI behaviour, and are called\nIndividual Discriminatory Instances (IDIs).\n  In this paper, we propose an approach for the selection of the initial seeds\nto generate IDIs for fairness testing. Previous studies mainly used random\ninitial seeds to this end. However this phase is crucial, as these seeds are\nthe basis of the follow-up IDIs generation. We dubbed our proposed seed\nselection approach I&D. It generates a large number of initial IDIs exhibiting\na great diversity, aiming at improving the overall performance of fairness\ntesting.\n  Our empirical study reveal that I&D is able to produce a larger number of\nIDIs with respect to four state-of-the-art seed generation approaches,\ngenerating 1.68X more IDIs on average. Moreover, we compare the use of I&D to\ntrain machine learning models and find that using I&D reduces the number of\nremaining IDIs by 29% when compared to the state-of-the-art, thus indicating\nthat I&D is effective for improving model fairness\n", "date": "2022-09-17", "categories": []}
{"id": "2209.08648", "title": "Through a fair looking-glass: mitigating bias in image datasets", "abstract": "  With the recent growth in computer vision applications, the question of how\nfair and unbiased they are has yet to be explored. There is abundant evidence\nthat the bias present in training data is reflected in the models, or even\namplified. Many previous methods for image dataset de-biasing, including models\nbased on augmenting datasets, are computationally expensive to implement. In\nthis study, we present a fast and effective model to de-bias an image dataset\nthrough reconstruction and minimizing the statistical dependence between\nintended variables. Our architecture includes a U-net to reconstruct images,\ncombined with a pre-trained classifier which penalizes the statistical\ndependence between target attribute and the protected attribute. We evaluate\nour proposed model on CelebA dataset, compare the results with a\nstate-of-the-art de-biasing method, and show that the model achieves a\npromising fairness-accuracy combination.\n", "date": "2022-09-18", "categories": []}
{"id": "2207.04546", "title": "FairDistillation: Mitigating Stereotyping in Language Models", "abstract": "  Large pre-trained language models are successfully being used in a variety of\ntasks, across many languages. With this ever-increasing usage, the risk of\nharmful side effects also rises, for example by reproducing and reinforcing\nstereotypes. However, detecting and mitigating these harms is difficult to do\nin general and becomes computationally expensive when tackling multiple\nlanguages or when considering different biases. To address this, we present\nFairDistillation: a cross-lingual method based on knowledge distillation to\nconstruct smaller language models while controlling for specific biases. We\nfound that our distillation method does not negatively affect the downstream\nperformance on most tasks and successfully mitigates stereotyping and\nrepresentational harms. We demonstrate that FairDistillation can create fairer\nlanguage models at a considerably lower cost than alternative approaches.\n", "date": "2022-07-10", "categories": []}
{"id": "2209.02836", "title": "Studying Bias in GANs through the Lens of Race", "abstract": "  In this work, we study how the performance and evaluation of generative image\nmodels are impacted by the racial composition of their training datasets. By\nexamining and controlling the racial distributions in various training\ndatasets, we are able to observe the impacts of different training\ndistributions on generated image quality and the racial distributions of the\ngenerated images. Our results show that the racial compositions of generated\nimages successfully preserve that of the training data. However, we observe\nthat truncation, a technique used to generate higher quality images during\ninference, exacerbates racial imbalances in the data. Lastly, when examining\nthe relationship between image quality and race, we find that the highest\nperceived visual quality images of a given race come from a distribution where\nthat race is well-represented, and that annotators consistently prefer\ngenerated images of white people over those of Black people.\n", "date": "2022-09-06", "categories": []}
{"id": "2209.06850", "title": "CAT: Controllable Attribute Translation for Fair Facial Attribute\n  Classification", "abstract": "  As the social impact of visual recognition has been under scrutiny, several\nprotected-attribute balanced datasets emerged to address dataset bias in\nimbalanced datasets. However, in facial attribute classification, dataset bias\nstems from both protected attribute level and facial attribute level, which\nmakes it challenging to construct a multi-attribute-level balanced real\ndataset. To bridge the gap, we propose an effective pipeline to generate\nhigh-quality and sufficient facial images with desired facial attributes and\nsupplement the original dataset to be a balanced dataset at both levels, which\ntheoretically satisfies several fairness criteria. The effectiveness of our\nmethod is verified on sex classification and facial attribute classification by\nyielding comparable task performance as the original dataset and further\nimproving fairness in a comprehensive fairness evaluation with a wide range of\nmetrics. Furthermore, our method outperforms both resampling and balanced\ndataset construction to address dataset bias, and debiasing models to address\ntask bias.\n", "date": "2022-09-14", "categories": []}
{"id": "2209.07044", "title": "Fair Inference for Discrete Latent Variable Models", "abstract": "  It is now well understood that machine learning models, trained on data\nwithout due care, often exhibit unfair and discriminatory behavior against\ncertain populations. Traditional algorithmic fairness research has mainly\nfocused on supervised learning tasks, particularly classification. While\nfairness in unsupervised learning has received some attention, the literature\nhas primarily addressed fair representation learning of continuous embeddings.\nIn this paper, we conversely focus on unsupervised learning using probabilistic\ngraphical models with discrete latent variables. We develop a fair stochastic\nvariational inference technique for the discrete latent variables, which is\naccomplished by including a fairness penalty on the variational distribution\nthat aims to respect the principles of intersectionality, a critical lens on\nfairness from the legal, social science, and humanities literature, and then\noptimizing the variational parameters under this penalty. We first show the\nutility of our method in improving equity and fairness for clustering using\nna\\\"ive Bayes and Gaussian mixture models on benchmark datasets. To demonstrate\nthe generality of our approach and its potential for real-world impact, we then\ndevelop a special-purpose graphical model for criminal justice risk\nassessments, and use our fairness approach to prevent the inferences from\nencoding unfair societal biases.\n", "date": "2022-09-15", "categories": []}
{"id": "2209.07047", "title": "iFlipper: Label Flipping for Individual Fairness", "abstract": "  As machine learning becomes prevalent, mitigating any unfairness present in\nthe training data becomes critical. Among the various notions of fairness, this\npaper focuses on the well-known individual fairness, which states that similar\nindividuals should be treated similarly. While individual fairness can be\nimproved when training a model (in-processing), we contend that fixing the data\nbefore model training (pre-processing) is a more fundamental solution. In\nparticular, we show that label flipping is an effective pre-processing\ntechnique for improving individual fairness. Our system iFlipper solves the\noptimization problem of minimally flipping labels given a limit to the\nindividual fairness violations, where a violation occurs when two similar\nexamples in the training data have different labels. We first prove that the\nproblem is NP-hard. We then propose an approximate linear programming algorithm\nand provide theoretical guarantees on how close its result is to the optimal\nsolution in terms of the number of label flips. We also propose techniques for\nmaking the linear programming solution more optimal without exceeding the\nviolations limit. Experiments on real datasets show that iFlipper significantly\noutperforms other pre-processing baselines in terms of individual fairness and\naccuracy on unseen test sets. In addition, iFlipper can be combined with\nin-processing techniques for even better results.\n", "date": "2022-09-15", "categories": []}
{"id": "2209.07190", "title": "Adaptive Fairness Improvement Based on Causality Analysis", "abstract": "  Given a discriminating neural network, the problem of fairness improvement is\nto systematically reduce discrimination without significantly scarifies its\nperformance (i.e., accuracy). Multiple categories of fairness improving methods\nhave been proposed for neural networks, including pre-processing, in-processing\nand post-processing. Our empirical study however shows that these methods are\nnot always effective (e.g., they may improve fairness by paying the price of\nhuge accuracy drop) or even not helpful (e.g., they may even worsen both\nfairness and accuracy). In this work, we propose an approach which adaptively\nchooses the fairness improving method based on causality analysis. That is, we\nchoose the method based on how the neurons and attributes responsible for\nunfairness are distributed among the input attributes and the hidden neurons.\nOur experimental evaluation shows that our approach is effective (i.e., always\nidentify the best fairness improving method) and efficient (i.e., with an\naverage time overhead of 5 minutes).\n", "date": "2022-09-15", "categories": []}
{"id": "2209.07312", "title": "Multicalibrated Regression for Downstream Fairness", "abstract": "  We show how to take a regression function $\\hat{f}$ that is appropriately\n``multicalibrated'' and efficiently post-process it into an approximately error\nminimizing classifier satisfying a large variety of fairness constraints. The\npost-processing requires no labeled data, and only a modest amount of unlabeled\ndata and computation. The computational and sample complexity requirements of\ncomputing $\\hat f$ are comparable to the requirements for solving a single fair\nlearning task optimally, but it can in fact be used to solve many different\ndownstream fairness-constrained learning problems efficiently. Our\npost-processing method easily handles intersecting groups, generalizing prior\nwork on post-processing regression functions to satisfy fairness constraints\nthat only applied to disjoint groups. Our work extends recent work showing that\nmulticalibrated regression functions are ``omnipredictors'' (i.e. can be\npost-processed to optimally solve unconstrained ERM problems) to constrained\noptimization.\n", "date": "2022-09-15", "categories": []}
{"id": "2209.05602", "title": "It's Not Fairness, and It's Not Fair: The Failure of Distributional\n  Equality and the Promise of Relational Equality in Complete-Information\n  Hiring Games", "abstract": "  Existing efforts to formulate computational definitions of fairness have\nlargely focused on distributional notions of equality, where equality is\ndefined by the resources or decisions given to individuals in the system. Yet\nexisting discrimination and injustice is often the result of unequal social\nrelations, rather than an unequal distribution of resources. Here, we show how\noptimizing for existing computational and economic definitions of fairness and\nequality fail to prevent unequal social relations. To do this, we provide an\nexample of a self-confirming equilibrium in a simple hiring market that is\nrelationally unequal but satisfies existing distributional notions of fairness.\nIn doing so, we introduce a notion of blatant relational unfairness for\ncomplete-information games, and discuss how this definition helps initiate a\nnew approach to incorporating relational equality into computational systems.\n", "date": "2022-09-12", "categories": []}
{"id": "2209.03904", "title": "Analyzing the Effect of Sampling in GNNs on Individual Fairness", "abstract": "  Graph neural network (GNN) based methods have saturated the field of\nrecommender systems. The gains of these systems have been significant, showing\nthe advantages of interpreting data through a network structure. However,\ndespite the noticeable benefits of using graph structures in recommendation\ntasks, this representational form has also bred new challenges which exacerbate\nthe complexity of mitigating algorithmic bias. When GNNs are integrated into\ndownstream tasks, such as recommendation, bias mitigation can become even more\ndifficult. Furthermore, the intractability of applying existing methods of\nfairness promotion to large, real world datasets places even more serious\nconstraints on mitigation attempts. Our work sets out to fill in this gap by\ntaking an existing method for promoting individual fairness on graphs and\nextending it to support mini-batch, or sub-sample based, training of a GNN,\nthus laying the groundwork for applying this method to a downstream\nrecommendation task. We evaluate two popular GNN methods: Graph Convolutional\nNetwork (GCN), which trains on the entire graph, and GraphSAGE, which uses\nprobabilistic random walks to create subgraphs for mini-batch training, and\nassess the effects of sub-sampling on individual fairness. We implement an\nindividual fairness notion called \\textit{REDRESS}, proposed by Dong et al.,\nwhich uses rank optimization to learn individual fair node, or item,\nembeddings. We empirically show on two real world datasets that GraphSAGE is\nable to achieve, not just, comparable accuracy, but also, improved fairness as\ncompared with the GCN model. These finding have consequential ramifications to\nindividual fairness promotion, GNNs, and in downstream form, recommender\nsystems, showing that mini-batch training facilitate individual fairness\npromotion by allowing for local nuance to guide the process of fairness\npromotion in representation learning.\n", "date": "2022-09-08", "categories": []}
{"id": "2209.04043", "title": "Who Pays? Personalization, Bossiness and the Cost of Fairness", "abstract": "  Fairness-aware recommender systems that have a provider-side fairness concern\nseek to ensure that protected group(s) of providers have a fair opportunity to\npromote their items or products. There is a ``cost of fairness'' borne by the\nconsumer side of the interaction when such a solution is implemented. This\nconsumer-side cost raises its own questions of fairness, particularly when\npersonalization is used to control the impact of the fairness constraint. In\nadopting a personalized approach to the fairness objective, researchers may be\nopening their systems up to strategic behavior on the part of users. This type\nof incentive has been studied in the computational social choice literature\nunder the terminology of ``bossiness''. The concern is that a bossy user may be\nable to shift the cost of fairness to others, improving their own outcomes and\nworsening those for others. This position paper introduces the concept of\nbossiness, shows its application in fairness-aware recommendation and discusses\nstrategies for reducing this strategic incentive.\n", "date": "2022-09-08", "categories": []}
{"id": "2207.10077", "title": "Discover and Mitigate Unknown Biases with Debiasing Alternate Networks", "abstract": "  Deep image classifiers have been found to learn biases from datasets. To\nmitigate the biases, most previous methods require labels of protected\nattributes (e.g., age, skin tone) as full-supervision, which has two\nlimitations: 1) it is infeasible when the labels are unavailable; 2) they are\nincapable of mitigating unknown biases -- biases that humans do not\npreconceive. To resolve those problems, we propose Debiasing Alternate Networks\n(DebiAN), which comprises two networks -- a Discoverer and a Classifier. By\ntraining in an alternate manner, the discoverer tries to find multiple unknown\nbiases of the classifier without any annotations of biases, and the classifier\naims at unlearning the biases identified by the discoverer. While previous\nworks evaluate debiasing results in terms of a single bias, we create\nMulti-Color MNIST dataset to better benchmark mitigation of multiple biases in\na multi-bias setting, which not only reveals the problems in previous methods\nbut also demonstrates the advantage of DebiAN in identifying and mitigating\nmultiple biases simultaneously. We further conduct extensive experiments on\nreal-world datasets, showing that the discoverer in DebiAN can identify unknown\nbiases that may be hard to be found by humans. Regarding debiasing, DebiAN\nachieves strong bias mitigation performance.\n", "date": "2022-07-20", "categories": []}
{"id": "2209.03620", "title": "Black-Box Audits for Group Distribution Shifts", "abstract": "  When a model informs decisions about people, distribution shifts can create\nundue disparities. However, it is hard for external entities to check for\ndistribution shift, as the model and its training set are often proprietary. In\nthis paper, we introduce and study a black-box auditing method to detect cases\nof distribution shift that lead to a performance disparity of the model across\ndemographic groups. By extending techniques used in membership and property\ninference attacks -- which are designed to expose private information from\nlearned models -- we demonstrate that an external auditor can gain the\ninformation needed to identify these distribution shifts solely by querying the\nmodel. Our experimental results on real-world datasets show that this approach\nis effective, achieving 80--100% AUC-ROC in detecting shifts involving the\nunderrepresentation of a demographic group in the training set. Researchers and\ninvestigative journalists can use our tools to perform non-collaborative audits\nof proprietary models and expose cases of underrepresentation in the training\ndatasets.\n", "date": "2022-09-08", "categories": []}
{"id": "2209.03805", "title": "FAT Forensics: A Python Toolbox for Implementing and Deploying Fairness,\n  Accountability and Transparency Algorithms in Predictive Systems", "abstract": "  Predictive systems, in particular machine learning algorithms, can take\nimportant, and sometimes legally binding, decisions about our everyday life. In\nmost cases, however, these systems and decisions are neither regulated nor\ncertified. Given the potential harm that these algorithms can cause, their\nqualities such as fairness, accountability and transparency (FAT) are of\nparamount importance. To ensure high-quality, fair, transparent and reliable\npredictive systems, we developed an open source Python package called FAT\nForensics. It can inspect important fairness, accountability and transparency\naspects of predictive algorithms to automatically and objectively report them\nback to engineers and users of such systems. Our toolbox can evaluate all\nelements of a predictive pipeline: data (and their features), models and\npredictions. Published under the BSD 3-Clause open source licence, FAT\nForensics is opened up for personal and commercial usage.\n", "date": "2022-09-08", "categories": []}
{"id": "2209.03942", "title": "Data Feedback Loops: Model-driven Amplification of Dataset Biases", "abstract": "  Datasets scraped from the internet have been critical to the successes of\nlarge-scale machine learning. Yet, this very success puts the utility of future\ninternet-derived datasets at potential risk, as model outputs begin to replace\nhuman annotations as a source of supervision.\n  In this work, we first formalize a system where interactions with one model\nare recorded as history and scraped as training data in the future. We then\nanalyze its stability over time by tracking changes to a test-time bias\nstatistic (e.g. gender bias of model predictions). We find that the degree of\nbias amplification is closely linked to whether the model's outputs behave like\nsamples from the training distribution, a behavior which we characterize and\ndefine as consistent calibration. Experiments in three conditional prediction\nscenarios - image classification, visual role-labeling, and language generation\n- demonstrate that models that exhibit a sampling-like behavior are more\ncalibrated and thus more stable. Based on this insight, we propose an\nintervention to help calibrate and stabilize unstable feedback systems.\n  Code is available at https://github.com/rtaori/data_feedback.\n", "date": "2022-09-08", "categories": []}
{"id": "2209.03089", "title": "Decoding Demographic un-fairness from Indian Names", "abstract": "  Demographic classification is essential in fairness assessment in recommender\nsystems or in measuring unintended bias in online networks and voting systems.\nImportant fields like education and politics, which often lay a foundation for\nthe future of equality in society, need scrutiny to design policies that can\nbetter foster equality in resource distribution constrained by the unbalanced\ndemographic distribution of people in the country.\n  We collect three publicly available datasets to train state-of-the-art\nclassifiers in the domain of gender and caste classification. We train the\nmodels in the Indian context, where the same name can have different styling\nconventions (Jolly Abraham/Kumar Abhishikta in one state may be written as\nAbraham Jolly/Abishikta Kumar in the other). Finally, we also perform\ncross-testing (training and testing on different datasets) to understand the\nefficacy of the above models.\n  We also perform an error analysis of the prediction models. Finally, we\nattempt to assess the bias in the existing Indian system as case studies and\nfind some intriguing patterns manifesting in the complex demographic layout of\nthe sub-continent across the dimensions of gender and caste.\n", "date": "2022-09-07", "categories": []}
{"id": "2209.00858", "title": "A Discussion of Discrimination and Fairness in Insurance Pricing", "abstract": "  Indirect discrimination is an issue of major concern in algorithmic models.\nThis is particularly the case in insurance pricing where protected policyholder\ncharacteristics are not allowed to be used for insurance pricing. Simply\ndisregarding protected policyholder information is not an appropriate solution\nbecause this still allows for the possibility of inferring the protected\ncharacteristics from the non-protected ones. This leads to so-called proxy or\nindirect discrimination. Though proxy discrimination is qualitatively different\nfrom the group fairness concepts in machine learning, these group fairness\nconcepts are proposed to 'smooth out' the impact of protected characteristics\nin the calculation of insurance prices. The purpose of this note is to share\nsome thoughts about group fairness concepts in the light of insurance pricing\nand to discuss their implications. We present a statistical model that is free\nof proxy discrimination, thus, unproblematic from an insurance pricing point of\nview. However, we find that the canonical price in this statistical model does\nnot satisfy any of the three most popular group fairness axioms. This seems\npuzzling and we welcome feedback on our example and on the usefulness of these\ngroup fairness axioms for non-discriminatory insurance pricing.\n", "date": "2022-09-02", "categories": []}
{"id": "2209.00427", "title": "Fair learning with Wasserstein barycenters for non-decomposable\n  performance measures", "abstract": "  This work provides several fundamental characterizations of the optimal\nclassification function under the demographic parity constraint. In the\nawareness framework, akin to the classical unconstrained classification case,\nwe show that maximizing accuracy under this fairness constraint is equivalent\nto solving a corresponding regression problem followed by thresholding at level\n$1/2$. We extend this result to linear-fractional classification measures\n(e.g., ${\\rm F}$-score, AM measure, balanced accuracy, etc.), highlighting the\nfundamental role played by the regression problem in this framework. Our\nresults leverage recently developed connection between the demographic parity\nconstraint and the multi-marginal optimal transport formulation. Informally,\nour result shows that the transition between the unconstrained problems and the\nfair one is achieved by replacing the conditional expectation of the label by\nthe solution of the fair regression problem. Finally, leveraging our analysis,\nwe demonstrate an equivalence between the awareness and the unawareness setups\nin the case of two sensitive groups.\n", "date": "2022-09-01", "categories": []}
{"id": "2206.07247", "title": "Fair Ranking as Fair Division: Impact-Based Individual Fairness in\n  Ranking", "abstract": "  Rankings have become the primary interface in two-sided online markets. Many\nhave noted that the rankings not only affect the satisfaction of the users\n(e.g., customers, listeners, employers, travelers), but that the position in\nthe ranking allocates exposure -- and thus economic opportunity -- to the\nranked items (e.g., articles, products, songs, job seekers, restaurants,\nhotels). This has raised questions of fairness to the items, and most existing\nworks have addressed fairness by explicitly linking item exposure to item\nrelevance. However, we argue that any particular choice of such a link function\nmay be difficult to defend, and we show that the resulting rankings can still\nbe unfair. To avoid these shortcomings, we develop a new axiomatic approach\nthat is rooted in principles of fair division. This not only avoids the need to\nchoose a link function, but also more meaningfully quantifies the impact on the\nitems beyond exposure. Our axioms of envy-freeness and dominance over uniform\nranking postulate that for a fair ranking policy every item should prefer their\nown rank allocation over that of any other item, and that no item should be\nactively disadvantaged by the rankings. To compute ranking policies that are\nfair according to these axioms, we propose a new ranking objective related to\nthe Nash Social Welfare. We show that the solution has guarantees regarding its\nenvy-freeness, its dominance over uniform rankings for every item, and its\nPareto optimality. In contrast, we show that conventional exposure-based\nfairness can produce large amounts of envy and have a highly disparate impact\non the items. Beyond these theoretical results, we illustrate empirically how\nour framework controls the trade-off between impact-based individual item\nfairness and user utility.\n", "date": "2022-06-15", "categories": []}
{"id": "2208.14175", "title": "RAGUEL: Recourse-Aware Group Unfairness Elimination", "abstract": "  While machine learning and ranking-based systems are in widespread use for\nsensitive decision-making processes (e.g., determining job candidates,\nassigning credit scores), they are rife with concerns over unintended biases in\ntheir outcomes, which makes algorithmic fairness (e.g., demographic parity,\nequal opportunity) an objective of interest. 'Algorithmic recourse' offers\nfeasible recovery actions to change unwanted outcomes through the modification\nof attributes. We introduce the notion of ranked group-level recourse fairness,\nand develop a 'recourse-aware ranking' solution that satisfies ranked recourse\nfairness constraints while minimizing the cost of suggested modifications. Our\nsolution suggests interventions that can reorder the ranked list of database\nrecords and mitigate group-level unfairness; specifically, disproportionate\nrepresentation of sub-groups and recourse cost imbalance. This re-ranking\nidentifies the minimum modifications to data points, with these attribute\nmodifications weighted according to their ease of recourse. We then present an\nefficient block-based extension that enables re-ranking at any granularity\n(e.g., multiple brackets of bank loan interest rates, multiple pages of search\nengine results). Evaluation on real datasets shows that, while existing methods\nmay even exacerbate recourse unfairness, our solution -- RAGUEL --\nsignificantly improves recourse-aware fairness. RAGUEL outperforms alternatives\nat improving recourse fairness, through a combined process of counterfactual\ngeneration and re-ranking, whilst remaining efficient for large-scale datasets.\n", "date": "2022-08-30", "categories": []}
{"id": "2208.12606", "title": "Pushing the limits of fairness impossibility: Who's the fairest of them\n  all?", "abstract": "  The impossibility theorem of fairness is a foundational result in the\nalgorithmic fairness literature. It states that outside of special cases, one\ncannot exactly and simultaneously satisfy all three common and intuitive\ndefinitions of fairness - demographic parity, equalized odds, and predictive\nrate parity. This result has driven most works to focus on solutions for one or\ntwo of the metrics. Rather than follow suit, in this paper we present a\nframework that pushes the limits of the impossibility theorem in order to\nsatisfy all three metrics to the best extent possible. We develop an\ninteger-programming based approach that can yield a certifiably optimal\npost-processing method for simultaneously satisfying multiple fairness criteria\nunder small violations. We show experiments demonstrating that our\npost-processor can improve fairness across the different definitions\nsimultaneously with minimal model performance reduction. We also discuss\napplications of our framework for model selection and fairness explainability,\nthereby attempting to answer the question: who's the fairest of them all?\n", "date": "2022-08-24", "categories": []}
{"id": "2208.12786", "title": "LUCID: Exposing Algorithmic Bias through Inverse Design", "abstract": "  AI systems can create, propagate, support, and automate bias in\ndecision-making processes. To mitigate biased decisions, we both need to\nunderstand the origin of the bias and define what it means for an algorithm to\nmake fair decisions. Most group fairness notions assess a model's equality of\noutcome by computing statistical metrics on the outputs. We argue that these\noutput metrics encounter intrinsic obstacles and present a complementary\napproach that aligns with the increasing focus on equality of treatment. By\nLocating Unfairness through Canonical Inverse Design (LUCID), we generate a\ncanonical set that shows the desired inputs for a model given a preferred\noutput. The canonical set reveals the model's internal logic and exposes\npotential unethical biases by repeatedly interrogating the decision-making\nprocess. We evaluate LUCID on the UCI Adult and COMPAS data sets and find that\nsome biases detected by a canonical set differ from those of output metrics.\nThe results show that by shifting the focus towards equality of treatment and\nlooking into the algorithm's internal workings, the canonical sets are a\nvaluable addition to the toolbox of algorithmic fairness evaluation.\n", "date": "2022-08-26", "categories": []}
{"id": "2208.11744", "title": "Enforcing Delayed-Impact Fairness Guarantees", "abstract": "  Recent research has shown that seemingly fair machine learning models, when\nused to inform decisions that have an impact on peoples' lives or well-being\n(e.g., applications involving education, employment, and lending), can\ninadvertently increase social inequality in the long term. This is because\nprior fairness-aware algorithms only consider static fairness constraints, such\nas equal opportunity or demographic parity. However, enforcing constraints of\nthis type may result in models that have negative long-term impact on\ndisadvantaged individuals and communities. We introduce ELF (Enforcing\nLong-term Fairness), the first classification algorithm that provides\nhigh-confidence fairness guarantees in terms of long-term, or delayed, impact.\nWe prove that the probability that ELF returns an unfair solution is less than\na user-specified tolerance and that (under mild assumptions), given sufficient\ntraining data, ELF is able to find and return a fair solution if one exists. We\nshow experimentally that our algorithm can successfully mitigate long-term\nunfairness.\n", "date": "2022-08-24", "categories": []}
{"id": "2208.11321", "title": "TESTSGD: Interpretable Testing of Neural Networks Against Subtle Group\n  Discrimination", "abstract": "  Discrimination has been shown in many machine learning applications, which\ncalls for sufficient fairness testing before their deployment in ethic-relevant\ndomains such as face recognition, medical diagnosis and criminal sentence.\nExisting fairness testing approaches are mostly designed for identifying\nindividual discrimination, i.e., discrimination against individuals. Yet, as\nanother widely concerning type of discrimination, testing against group\ndiscrimination, mostly hidden, is much less studied. To address the gap, in\nthis work, we propose TESTSGD, an interpretable testing approach which\nsystematically identifies and measures hidden (which we call `subtle' group\ndiscrimination} of a neural network characterized by conditions over\ncombinations of the sensitive features. Specifically, given a neural network,\nTESTSGDfirst automatically generates an interpretable rule set which\ncategorizes the input space into two groups exposing the model's group\ndiscrimination. Alongside, TESTSGDalso provides an estimated group fairness\nscore based on sampling the input space to measure the degree of the identified\nsubtle group discrimination, which is guaranteed to be accurate up to an error\nbound. We evaluate TESTSGDon multiple neural network models trained on popular\ndatasets including both structured data and text data. The experiment results\nshow that TESTSGDis effective and efficient in identifying and measuring such\nsubtle group discrimination that has never been revealed before. Furthermore,\nwe show that the testing results of TESTSGDcan guide generation of new samples\nto mitigate such discrimination through retraining with negligible accuracy\ndrop.\n", "date": "2022-08-24", "categories": []}
{"id": "2208.11362", "title": "A novel approach for Fair Principal Component Analysis based on\n  eigendecomposition", "abstract": "  Principal component analysis (PCA), a ubiquitous dimensionality reduction\ntechnique in signal processing, searches for a projection matrix that minimizes\nthe mean squared error between the reduced dataset and the original one. Since\nclassical PCA is not tailored to address concerns related to fairness, its\napplication to actual problems may lead to disparity in the reconstruction\nerrors of different groups (e.g., men and women, whites and blacks, etc.), with\npotentially harmful consequences such as the introduction of bias towards\nsensitive groups. Although several fair versions of PCA have been proposed\nrecently, there still remains a fundamental gap in the search for algorithms\nthat are simple enough to be deployed in real systems. To address this, we\npropose a novel PCA algorithm which tackles fairness issues by means of a\nsimple strategy comprising a one-dimensional search which exploits the\nclosed-form solution of PCA. As attested by numerical experiments, the proposal\ncan significantly improve fairness with a very small loss in the overall\nreconstruction error and without resorting to complex optimization schemes.\nMoreover, our findings are consistent in several real situations as well as in\nscenarios with both unbalanced and balanced datasets.\n", "date": "2022-08-24", "categories": []}
{"id": "2207.01490", "title": "Experts' View on Challenges and Needs for Fairness in Artificial\n  Intelligence for Education", "abstract": "  In recent years, there has been a stimulating discussion on how artificial\nintelligence (AI) can support the science and engineering of intelligent\neducational applications. Many studies in the field are proposing actionable\ndata mining pipelines and machine-learning models driven by learning-related\ndata. The potential of these pipelines and models to amplify unfairness for\ncertain categories of students is however receiving increasing attention. If AI\napplications are to have a positive impact on education, it is crucial that\ntheir design considers fairness at every step. Through anonymous surveys and\ninterviews with experts (researchers and practitioners) who have published\ntheir research at top-tier educational conferences in the last year, we\nconducted the first expert-driven systematic investigation on the challenges\nand needs for addressing fairness throughout the development of educational\nsystems based on AI. We identified common and diverging views about the\nchallenges and the needs faced by educational technologies experts in practice,\nthat lead the community to have a clear understanding on the main questions\nraising doubts in this topic. Based on these findings, we highlighted\ndirections that will facilitate the ongoing research towards fairer AI for\neducation.\n", "date": "2022-06-23", "categories": []}
{"id": "2208.10013", "title": "FairDisCo: Fairer AI in Dermatology via Disentanglement Contrastive\n  Learning", "abstract": "  Deep learning models have achieved great success in automating skin lesion\ndiagnosis. However, the ethnic disparity in these models' predictions, where\nlesions on darker skin types are usually underrepresented and have lower\ndiagnosis accuracy, receives little attention. In this paper, we propose\nFairDisCo, a disentanglement deep learning framework with contrastive learning\nthat utilizes an additional network branch to remove sensitive attributes, i.e.\nskin-type information from representations for fairness and another contrastive\nbranch to enhance feature extraction. We compare FairDisCo to three fairness\nmethods, namely, resampling, reweighting, and attribute-aware, on two newly\nreleased skin lesion datasets with different skin types: Fitzpatrick17k and\nDiverse Dermatology Images (DDI). We adapt two fairness-based metrics DPM and\nEOM for our multiple classes and sensitive attributes task, highlighting the\nskin-type bias in skin lesion classification. Extensive experimental evaluation\ndemonstrates the effectiveness of FairDisCo, with fairer and superior\nperformance on skin lesion classification tasks.\n", "date": "2022-08-22", "categories": []}
{"id": "2208.10095", "title": "Socially Fair Center-based and Linear Subspace Clustering", "abstract": "  Center-based clustering (e.g., $k$-means, $k$-medians) and clustering using\nlinear subspaces are two most popular techniques to partition real-world data\ninto smaller clusters. However, when the data consists of sensitive demographic\ngroups, significantly different clustering cost per point for different\nsensitive groups can lead to fairness-related harms (e.g., different\nquality-of-service). The goal of socially fair clustering is to minimize the\nmaximum cost of clustering per point over all groups. In this work, we propose\na unified framework to solve socially fair center-based clustering and linear\nsubspace clustering, and give practical, efficient approximation algorithms for\nthese problems. We do extensive experiments to show that on multiple benchmark\ndatasets our algorithms either closely match or outperform state-of-the-art\nbaselines.\n", "date": "2022-08-22", "categories": []}
{"id": "2208.07918", "title": "Ex-Ante Assessment of Discrimination in Dataset", "abstract": "  Data owners face increasing liability for how the use of their data could\nharm under-priviliged communities. Stakeholders would like to identify the\ncharacteristics of data that lead to algorithms being biased against any\nparticular demographic groups, for example, defined by their race, gender, age,\nand/or religion. Specifically, we are interested in identifying subsets of the\nfeature space where the ground truth response function from features to\nobserved outcomes differs across demographic groups. To this end, we propose\nFORESEE, a FORESt of decision trEEs algorithm, which generates a score that\ncaptures how likely an individual's response varies with sensitive attributes.\nEmpirically, we find that our approach allows us to identify the individuals\nwho are most likely to be misclassified by several classifiers, including\nRandom Forest, Logistic Regression, Support Vector Machine, and k-Nearest\nNeighbors. The advantage of our approach is that it allows stakeholders to\ncharacterize risky samples that may contribute to discrimination, as well as,\nuse the FORESEE to estimate the risk of upcoming samples.\n", "date": "2022-08-16", "categories": []}
{"id": "2208.08279", "title": "Error Parity Fairness: Testing for Group Fairness in Regression Tasks", "abstract": "  The applications of Artificial Intelligence (AI) surround decisions on\nincreasingly many aspects of human lives. Society responds by imposing legal\nand social expectations for the accountability of such automated decision\nsystems (ADSs). Fairness, a fundamental constituent of AI accountability, is\nconcerned with just treatment of individuals and sensitive groups (e.g., based\non sex, race). While many studies focus on fair learning and fairness testing\nfor the classification tasks, the literature is rather limited on how to\nexamine fairness in regression tasks. This work presents error parity as a\nregression fairness notion and introduces a testing methodology to assess group\nfairness based on a statistical hypothesis testing procedure. The error parity\ntest checks whether prediction errors are distributed similarly across\nsensitive groups to determine if an ADS is fair. It is followed by a suitable\npermutation test to compare groups on several statistics to explore disparities\nand identify impacted groups. The usefulness and applicability of the proposed\nmethodology are demonstrated via a case study on COVID-19 projections in the US\nat the county level, which revealed race-based differences in forecast errors.\nOverall, the proposed regression fairness testing methodology fills a gap in\nthe fair machine learning literature and may serve as a part of larger\naccountability assessments and algorithm audits.\n", "date": "2022-08-16", "categories": []}
{"id": "2208.08382", "title": "Deep Generative Views to Mitigate Gender Classification Bias Across\n  Gender-Race Groups", "abstract": "  Published studies have suggested the bias of automated face-based gender\nclassification algorithms across gender-race groups. Specifically, unequal\naccuracy rates were obtained for women and dark-skinned people. To mitigate the\nbias of gender classifiers, the vision community has developed several\nstrategies. However, the efficacy of these mitigation strategies is\ndemonstrated for a limited number of races mostly, Caucasian and\nAfrican-American. Further, these strategies often offer a trade-off between\nbias and classification accuracy. To further advance the state-of-the-art, we\nleverage the power of generative views, structured learning, and evidential\nlearning towards mitigating gender classification bias. We demonstrate the\nsuperiority of our bias mitigation strategy in improving classification\naccuracy and reducing bias across gender-racial groups through extensive\nexperimental validation, resulting in state-of-the-art performance in intra-\nand cross dataset evaluations.\n", "date": "2022-08-17", "categories": []}
{"id": "2208.07613", "title": "Does lossy image compression affect racial bias within face recognition?", "abstract": "  Yes - This study investigates the impact of commonplace lossy image\ncompression on face recognition algorithms with regard to the racial\ncharacteristics of the subject. We adopt a recently proposed racial\nphenotype-based bias analysis methodology to measure the effect of varying\nlevels of lossy compression across racial phenotype categories. Additionally,\nwe determine the relationship between chroma-subsampling and race-related\nphenotypes for recognition performance. Prior work investigates the impact of\nlossy JPEG compression algorithm on contemporary face recognition performance.\nHowever, there is a gap in how this impact varies with different race-related\ninter-sectional groups and the cause of this impact. Via an extensive\nexperimental setup, we demonstrate that common lossy image compression\napproaches have a more pronounced negative impact on facial recognition\nperformance for specific racial phenotype categories such as darker skin tones\n(by up to 34.55\\%). Furthermore, removing chroma-subsampling during compression\nimproves the false matching rate (up to 15.95\\%) across all phenotype\ncategories affected by the compression, including darker skin tones, wide\nnoses, big lips, and monolid eye categories. In addition, we outline the\ncharacteristics that may be attributable as the underlying cause of such\nphenomenon for lossy compression algorithms such as JPEG.\n", "date": "2022-08-16", "categories": []}
{"id": "2208.06557", "title": "A Novel Regularization Approach to Fair ML", "abstract": "  A number of methods have been introduced for the fair ML issue, most of them\ncomplex and many of them very specific to the underlying ML moethodology. Here\nwe introduce a new approach that is simple, easily explained, and potentially\napplicable to a number of standard ML algorithms. Explicitly Deweighted\nFeatures (EDF) reduces the impact of each feature among the proxies of\nsensitive variables, allowing a different amount of deweighting applied to each\nsuch feature. The user specifies the deweighting hyperparameters, to achieve a\ngiven point in the Utility/Fairness tradeoff spectrum. We also introduce a new,\nsimple criterion for evaluating the degree of protection afforded by any fair\nML method.\n", "date": "2022-08-13", "categories": []}
{"id": "2208.05126", "title": "D-BIAS: A Causality-Based Human-in-the-Loop System for Tackling\n  Algorithmic Bias", "abstract": "  With the rise of AI, algorithms have become better at learning underlying\npatterns from the training data including ingrained social biases based on\ngender, race, etc. Deployment of such algorithms to domains such as hiring,\nhealthcare, law enforcement, etc. has raised serious concerns about fairness,\naccountability, trust and interpretability in machine learning algorithms. To\nalleviate this problem, we propose D-BIAS, a visual interactive tool that\nembodies human-in-the-loop AI approach for auditing and mitigating social\nbiases from tabular datasets. It uses a graphical causal model to represent\ncausal relationships among different features in the dataset and as a medium to\ninject domain knowledge. A user can detect the presence of bias against a\ngroup, say females, or a subgroup, say black females, by identifying unfair\ncausal relationships in the causal network and using an array of fairness\nmetrics. Thereafter, the user can mitigate bias by acting on the unfair causal\nedges. For each interaction, say weakening/deleting a biased causal edge, the\nsystem uses a novel method to simulate a new (debiased) dataset based on the\ncurrent causal model. Users can visually assess the impact of their\ninteractions on different fairness metrics, utility metrics, data distortion,\nand the underlying data distribution. Once satisfied, they can download the\ndebiased dataset and use it for any downstream application for fairer\npredictions. We evaluate D-BIAS by conducting experiments on 3 datasets and\nalso a formal user study. We found that D-BIAS helps reduce bias significantly\ncompared to the baseline debiasing approach across different fairness metrics\nwhile incurring little data distortion and a small loss in utility. Moreover,\nour human-in-the-loop based approach significantly outperforms an automated\napproach on trust, interpretability and accountability.\n", "date": "2022-08-10", "categories": []}
{"id": "2208.03621", "title": "Bias Reducing Multitask Learning on Mental Health Prediction", "abstract": "  There has been an increase in research in developing machine learning models\nfor mental health detection or prediction in recent years due to increased\nmental health issues in society. Effective use of mental health prediction or\ndetection models can help mental health practitioners re-define mental\nillnesses more objectively than currently done, and identify illnesses at an\nearlier stage when interventions may be more effective. However, there is still\na lack of standard in evaluating bias in such machine learning models in the\nfield, which leads to challenges in providing reliable predictions and in\naddressing disparities. This lack of standards persists due to factors such as\ntechnical difficulties, complexities of high dimensional clinical health data,\netc., which are especially true for physiological signals. This along with\nprior evidence of relations between some physiological signals with certain\ndemographic identities restates the importance of exploring bias in mental\nhealth prediction models that utilize physiological signals. In this work, we\naim to perform a fairness analysis and implement a multi-task learning based\nbias mitigation method on anxiety prediction models using ECG data. Our method\nis based on the idea of epistemic uncertainty and its relationship with model\nweights and feature space representation. Our analysis showed that our anxiety\nprediction base model introduced some bias with regards to age, income,\nethnicity, and whether a participant is born in the U.S. or not, and our bias\nmitigation method performed better at reducing the bias in the model, when\ncompared to the reweighting mitigation technique. Our analysis on feature\nimportance also helped identify relationships between heart rate variability\nand multiple demographic groupings.\n", "date": "2022-08-07", "categories": []}
{"id": "2208.03209", "title": "Bias and Fairness in Computer Vision Applications of the Criminal\n  Justice System", "abstract": "  Discriminatory practices involving AI-driven police work have been the\nsubject of much controversies in the past few years, with algorithms such as\nCOMPAS, PredPol and ShotSpotter being accused of unfairly impacting minority\ngroups. At the same time, the issues of fairness in machine learning, and in\nparticular in computer vision, have been the subject of a growing number of\nacademic works. In this paper, we examine how these area intersect. We provide\ninformation on how these practices have come to exist and the difficulties in\nalleviating them. We then examine three applications currently in development\nto understand what risks they pose to fairness and how those risks can be\nmitigated.\n", "date": "2022-08-05", "categories": []}
{"id": "2208.02192", "title": "RealPatch: A Statistical Matching Framework for Model Patching with Real\n  Samples", "abstract": "  Machine learning classifiers are typically trained to minimise the average\nerror across a dataset. Unfortunately, in practice, this process often exploits\nspurious correlations caused by subgroup imbalance within the training data,\nresulting in high average performance but highly variable performance across\nsubgroups. Recent work to address this problem proposes model patching with\nCAMEL. This previous approach uses generative adversarial networks to perform\nintra-class inter-subgroup data augmentations, requiring (a) the training of a\nnumber of computationally expensive models and (b) sufficient quality of\nmodel's synthetic outputs for the given domain. In this work, we propose\nRealPatch, a framework for simpler, faster, and more data-efficient data\naugmentation based on statistical matching. Our framework performs model\npatching by augmenting a dataset with real samples, mitigating the need to\ntrain generative models for the target task. We demonstrate the effectiveness\nof RealPatch on three benchmark datasets, CelebA, Waterbirds and a subset of\niWildCam, showing improvements in worst-case subgroup performance and in\nsubgroup performance gap in binary classification. Furthermore, we conduct\nexperiments with the imSitu dataset with 211 classes, a setting where\ngenerative model-based patching such as CAMEL is impractical. We show that\nRealPatch can successfully eliminate dataset leakage while reducing model\nleakage and maintaining high utility. The code for RealPatch can be found at\nhttps://github.com/wearepal/RealPatch.\n", "date": "2022-08-03", "categories": []}
{"id": "2206.01410", "title": "Fair Classification via Transformer Neural Networks: Case Study of an\n  Educational Domain", "abstract": "  Educational technologies nowadays increasingly use data and Machine Learning\n(ML) models. This gives the students, instructors, and administrators support\nand insights for the optimum policy. However, it is well acknowledged that ML\nmodels are subject to bias, which raises concerns about the fairness, bias, and\ndiscrimination of using these automated ML algorithms in education and its\nunintended and unforeseen negative consequences. The contribution of bias\nduring the decision-making comes from datasets used for training ML models and\nthe model architecture. This paper presents a preliminary investigation of the\nfairness of transformer neural networks on the two tabular datasets: Law School\nand Student-Mathematics. In contrast to classical ML models, the\ntransformer-based models transform these tabular datasets into a richer\nrepresentation while solving the classification task. We use different fairness\nmetrics for evaluation and check the trade-off between fairness and accuracy of\nthe transformer-based models over the tabular datasets. Empirically, our\napproach shows impressive results regarding the trade-off between fairness and\nperformance on the Law School dataset.\n", "date": "2022-06-03", "categories": []}
{"id": "2206.06279", "title": "A Machine Learning Model for Predicting, Diagnosing, and Mitigating\n  Health Disparities in Hospital Readmission", "abstract": "  The management of hyperglycemia in hospitalized patients has a significant\nimpact on both morbidity and mortality. Therefore, it is important to predict\nthe need for diabetic patients to be hospitalized. However, using standard\nmachine learning approaches to make these predictions may result in health\ndisparities caused by biases in the data related to social determinants (such\nas race, age, and gender). These biases must be removed early in the data\ncollection process, before they enter the system and are reinforced by model\npredictions, resulting in biases in the model's decisions. In this paper, we\npropose a machine learning pipeline capable of making predictions as well as\ndetecting and mitigating biases in the data and model predictions. This\npipeline analyses the clinical data and determines whether biases exist in the\ndata, if so, it removes those biases before making predictions. We evaluate the\nperformance of the proposed method on a clinical dataset using accuracy and\nfairness measures. The findings of the results show that when we mitigate\nbiases early during the data ingestion, we get fairer predictions.\n", "date": "2022-06-13", "categories": []}
{"id": "2208.00781", "title": "Debiasing Deep Chest X-Ray Classifiers using Intra- and Post-processing\n  Methods", "abstract": "  Deep neural networks for image-based screening and computer-aided diagnosis\nhave achieved expert-level performance on various medical imaging modalities,\nincluding chest radiographs. Recently, several works have indicated that these\nstate-of-the-art classifiers can be biased with respect to sensitive patient\nattributes, such as race or gender, leading to growing concerns about\ndemographic disparities and discrimination resulting from algorithmic and\nmodel-based decision-making in healthcare. Fair machine learning has focused on\nmitigating such biases against disadvantaged or marginalised groups, mainly\nconcentrating on tabular data or natural images. This work presents two novel\nintra-processing techniques based on fine-tuning and pruning an already-trained\nneural network. These methods are simple yet effective and can be readily\napplied post hoc in a setting where the protected attribute is unknown during\nthe model development and test time. In addition, we compare several intra- and\npost-processing approaches applied to debiasing deep chest X-ray classifiers.\nTo the best of our knowledge, this is one of the first efforts studying\ndebiasing methods on chest radiographs. Our results suggest that the considered\napproaches successfully mitigate biases in fully connected and convolutional\nneural networks offering stable performance under various settings. The\ndiscussed methods can help achieve group fairness of deep medical image\nclassifiers when deploying them in domains with different fairness\nconsiderations and constraints.\n", "date": "2022-07-26", "categories": []}
{"id": "2207.14355", "title": "Multiple Attribute Fairness: Application to Fraud Detection", "abstract": "  We propose a fairness measure relaxing the equality conditions in the popular\nequal odds fairness regime for classification. We design an iterative,\nmodel-agnostic, grid-based heuristic that calibrates the outcomes per sensitive\nattribute value to conform to the measure. The heuristic is designed to handle\nhigh arity attribute values and performs a per attribute sanitization of\noutcomes across different protected attribute values. We also extend our\nheuristic for multiple attributes. Highlighting our motivating application,\nfraud detection, we show that the proposed heuristic is able to achieve\nfairness across multiple values of a single protected attribute, multiple\nprotected attributes. When compared to current fairness techniques, that focus\non two groups, we achieve comparable performance across several public data\nsets.\n", "date": "2022-07-28", "categories": []}
{"id": "2206.03883", "title": "Efficient Resource Allocation with Fairness Constraints in Restless\n  Multi-Armed Bandits", "abstract": "  Restless Multi-Armed Bandits (RMAB) is an apt model to represent\ndecision-making problems in public health interventions (e.g., tuberculosis,\nmaternal, and child care), anti-poaching planning, sensor monitoring,\npersonalized recommendations and many more. Existing research in RMAB has\ncontributed mechanisms and theoretical results to a wide variety of settings,\nwhere the focus is on maximizing expected value. In this paper, we are\ninterested in ensuring that RMAB decision making is also fair to different arms\nwhile maximizing expected value. In the context of public health settings, this\nwould ensure that different people and/or communities are fairly represented\nwhile making public health intervention decisions. To achieve this goal, we\nformally define the fairness constraints in RMAB and provide planning and\nlearning methods to solve RMAB in a fair manner. We demonstrate key theoretical\nproperties of fair RMAB and experimentally demonstrate that our proposed\nmethods handle fairness constraints without sacrificing significantly on\nsolution quality.\n", "date": "2022-06-08", "categories": []}
{"id": "2207.13343", "title": "Towards Soft Fairness in Restless Multi-Armed Bandits", "abstract": "  Restless multi-armed bandits (RMAB) is a framework for allocating limited\nresources under uncertainty. It is an extremely useful model for monitoring\nbeneficiaries and executing timely interventions to ensure maximum benefit in\npublic health settings (e.g., ensuring patients take medicines in tuberculosis\nsettings, ensuring pregnant mothers listen to automated calls about good\npregnancy practices). Due to the limited resources, typically certain\ncommunities or regions are starved of interventions that can have follow-on\neffects. To avoid starvation in the executed interventions across\nindividuals/regions/communities, we first provide a soft fairness constraint\nand then provide an approach to enforce the soft fairness constraint in RMABs.\nThe soft fairness constraint requires that an algorithm never probabilistically\nfavor one arm over another if the long-term cumulative reward of choosing the\nlatter arm is higher. Our approach incorporates softmax based value iteration\nmethod in the RMAB setting to design selection algorithms that manage to\nsatisfy the proposed fairness constraint. Our method, referred to as SoftFair,\nalso provides theoretical performance guarantees and is asymptotically optimal.\nFinally, we demonstrate the utility of our approaches on simulated benchmarks\nand show that the soft fairness constraint can be handled without a significant\nsacrifice on value.\n", "date": "2022-07-27", "categories": []}
{"id": "2207.11385", "title": "Causal Fairness Analysis", "abstract": "  Decision-making systems based on AI and machine learning have been used\nthroughout a wide range of real-world scenarios, including healthcare, law\nenforcement, education, and finance. It is no longer far-fetched to envision a\nfuture where autonomous systems will be driving entire business decisions and,\nmore broadly, supporting large-scale decision-making infrastructure to solve\nsociety's most challenging problems. Issues of unfairness and discrimination\nare pervasive when decisions are being made by humans, and remain (or are\npotentially amplified) when decisions are made using machines with little\ntransparency, accountability, and fairness. In this paper, we introduce a\nframework for \\textit{causal fairness analysis} with the intent of filling in\nthis gap, i.e., understanding, modeling, and possibly solving issues of\nfairness in decision-making settings. The main insight of our approach will be\nto link the quantification of the disparities present on the observed data with\nthe underlying, and often unobserved, collection of causal mechanisms that\ngenerate the disparity in the first place, challenge we call the Fundamental\nProblem of Causal Fairness Analysis (FPCFA). In order to solve the FPCFA, we\nstudy the problem of decomposing variations and empirical measures of fairness\nthat attribute such variations to structural mechanisms and different units of\nthe population. Our effort culminates in the Fairness Map, which is the first\nsystematic attempt to organize and explain the relationship between different\ncriteria found in the literature. Finally, we study which causal assumptions\nare minimally needed for performing causal fairness analysis and propose a\nFairness Cookbook, which allows data scientists to assess the existence of\ndisparate impact and disparate treatment.\n", "date": "2022-07-23", "categories": []}
{"id": "2207.12138", "title": "Towards Fairness-Aware Multi-Objective Optimization", "abstract": "  Recent years have seen the rapid development of fairness-aware machine\nlearning in mitigating unfairness or discrimination in decision-making in a\nwide range of applications. However, much less attention has been paid to the\nfairness-aware multi-objective optimization, which is indeed commonly seen in\nreal life, such as fair resource allocation problems and data driven\nmulti-objective optimization problems. This paper aims to illuminate and\nbroaden our understanding of multi-objective optimization from the perspective\nof fairness. To this end, we start with a discussion of user preferences in\nmulti-objective optimization and then explore its relationship to fairness in\nmachine learning and multi-objective optimization. Following the above\ndiscussions, representative cases of fairness-aware multiobjective optimization\nare presented, further elaborating the importance of fairness in traditional\nmulti-objective optimization, data-driven optimization and federated\noptimization. Finally, challenges and opportunities in fairness-aware\nmulti-objective optimization are addressed. We hope that this article makes a\nsmall step forward towards understanding fairness in the context of\noptimization and promote research interest in fairness-aware multi-objective\noptimization.\n", "date": "2022-07-22", "categories": []}
{"id": "2207.10888", "title": "FairGRAPE: Fairness-aware GRAdient Pruning mEthod for Face Attribute\n  Classification", "abstract": "  Existing pruning techniques preserve deep neural networks' overall ability to\nmake correct predictions but may also amplify hidden biases during the\ncompression process. We propose a novel pruning method, Fairness-aware GRAdient\nPruning mEthod (FairGRAPE), that minimizes the disproportionate impacts of\npruning on different sub-groups. Our method calculates the per-group importance\nof each model weight and selects a subset of weights that maintain the relative\nbetween-group total importance in pruning. The proposed method then prunes\nnetwork edges with small importance values and repeats the procedure by\nupdating importance values. We demonstrate the effectiveness of our method on\nfour different datasets, FairFace, UTKFace, CelebA, and ImageNet, for the tasks\nof face attribute classification where our method reduces the disparity in\nperformance degradation by up to 90% compared to the state-of-the-art pruning\nalgorithms. Our method is substantially more effective in a setting with a high\npruning rate (99%). The code and dataset used in the experiments are available\nat https://github.com/Bernardo1998/FairGRAPE\n", "date": "2022-07-22", "categories": []}
{"id": "2207.10991", "title": "Algorithmic Fairness in Business Analytics: Directions for Research and\n  Practice", "abstract": "  The extensive adoption of business analytics (BA) has brought financial gains\nand increased efficiencies. However, these advances have simultaneously drawn\nattention to rising legal and ethical challenges when BA inform decisions with\nfairness implications. As a response to these concerns, the emerging study of\nalgorithmic fairness deals with algorithmic outputs that may result in\ndisparate outcomes or other forms of injustices for subgroups of the\npopulation, especially those who have been historically marginalized. Fairness\nis relevant on the basis of legal compliance, social responsibility, and\nutility; if not adequately and systematically addressed, unfair BA systems may\nlead to societal harms and may also threaten an organization's own survival,\nits competitiveness, and overall performance. This paper offers a\nforward-looking, BA-focused review of algorithmic fairness. We first review the\nstate-of-the-art research on sources and measures of bias, as well as bias\nmitigation algorithms. We then provide a detailed discussion of the\nutility-fairness relationship, emphasizing that the frequent assumption of a\ntrade-off between these two constructs is often mistaken or short-sighted.\nFinally, we chart a path forward by identifying opportunities for business\nscholars to address impactful, open challenges that are key to the effective\nand responsible deployment of BA.\n", "date": "2022-07-22", "categories": []}
{"id": "2207.11173", "title": "Verifying Fairness in Quantum Machine Learning", "abstract": "  Due to the beyond-classical capability of quantum computing, quantum machine\nlearning is applied independently or embedded in classical models for decision\nmaking, especially in the field of finance. Fairness and other ethical issues\nare often one of the main concerns in decision making. In this work, we define\na formal framework for the fairness verification and analysis of quantum\nmachine learning decision models, where we adopt one of the most popular\nnotions of fairness in the literature based on the intuition -- any two similar\nindividuals must be treated similarly and are thus unbiased. We show that\nquantum noise can improve fairness and develop an algorithm to check whether a\n(noisy) quantum machine learning model is fair. In particular, this algorithm\ncan find bias kernels of quantum data (encoding individuals) during checking.\nThese bias kernels generate infinitely many bias pairs for investigating the\nunfairness of the model. Our algorithm is designed based on a highly efficient\ndata structure -- Tensor Networks -- and implemented on Google's TensorFlow\nQuantum. The utility and effectiveness of our algorithm are confirmed by the\nexperimental results, including income prediction and credit scoring on\nreal-world data, for a class of random (noisy) quantum decision models with 27\nqubits ($2^{27}$-dimensional state space) tripling ($2^{18}$ times more than)\nthat of the state-of-the-art algorithms for verifying quantum machine learning\nmodels.\n", "date": "2022-07-22", "categories": []}
{"id": "2207.10246", "title": "GBDF: Gender Balanced DeepFake Dataset Towards Fair DeepFake Detection", "abstract": "  Facial forgery by deepfakes has raised severe societal concerns. Several\nsolutions have been proposed by the vision community to effectively combat the\nmisinformation on the internet via automated deepfake detection systems. Recent\nstudies have demonstrated that facial analysis-based deep learning models can\ndiscriminate based on protected attributes. For the commercial adoption and\nmassive roll-out of the deepfake detection technology, it is vital to evaluate\nand understand the fairness (the absence of any prejudice or favoritism) of\ndeepfake detectors across demographic variations such as gender and race. As\nthe performance differential of deepfake detectors between demographic\nsubgroups would impact millions of people of the deprived sub-group. This paper\naims to evaluate the fairness of the deepfake detectors across males and\nfemales. However, existing deepfake datasets are not annotated with demographic\nlabels to facilitate fairness analysis. To this aim, we manually annotated\nexisting popular deepfake datasets with gender labels and evaluated the\nperformance differential of current deepfake detectors across gender. Our\nanalysis on the gender-labeled version of the datasets suggests (a) current\ndeepfake datasets have skewed distribution across gender, and (b) commonly\nadopted deepfake detectors obtain unequal performance across gender with mostly\nmales outperforming females. Finally, we contributed a gender-balanced and\nannotated deepfake dataset, GBDF, to mitigate the performance differential and\nto promote research and development towards fairness-aware deep fake detectors.\nThe GBDF dataset is publicly available at: https://github.com/aakash4305/GBDF\n", "date": "2022-07-21", "categories": []}
{"id": "2207.09833", "title": "AI Fairness: from Principles to Practice", "abstract": "  This paper summarizes and evaluates various approaches, methods, and\ntechniques for pursuing fairness in artificial intelligence (AI) systems. It\nexamines the merits and shortcomings of these measures and proposes practical\nguidelines for defining, measuring, and preventing bias in AI. In particular,\nit cautions against some of the simplistic, yet common, methods for evaluating\nbias in AI systems, and offers more sophisticated and effective alternatives.\nThe paper also addresses widespread controversies and confusions in the field\nby providing a common language among different stakeholders of high-impact AI\nsystems. It describes various trade-offs involving AI fairness, and provides\npractical recommendations for balancing them. It offers techniques for\nevaluating the costs and benefits of fairness targets, and defines the role of\nhuman judgment in setting these targets. This paper provides discussions and\nguidelines for AI practitioners, organization leaders, and policymakers, as\nwell as various links to additional materials for a more technical audience.\nNumerous real-world examples are provided to clarify the concepts, challenges,\nand recommendations from a practical perspective.\n", "date": "2022-07-20", "categories": []}
{"id": "2207.09960", "title": "Measuring and signing fairness as performance under multiple stakeholder\n  distributions", "abstract": "  As learning machines increase their influence on decisions concerning human\nlives, analyzing their fairness properties becomes a subject of central\nimportance. Yet, our best tools for measuring the fairness of learning systems\nare rigid fairness metrics encapsulated as mathematical one-liners, offer\nlimited power to the stakeholders involved in the prediction task, and are easy\nto manipulate when we exhort excessive pressure to optimize them. To advance\nthese issues, we propose to shift focus from shaping fairness metrics to\ncurating the distributions of examples under which these are computed. In\nparticular, we posit that every claim about fairness should be immediately\nfollowed by the tagline \"Fair under what examples, and collected by whom?\". By\nhighlighting connections to the literature in domain generalization, we propose\nto measure fairness as the ability of the system to generalize under multiple\nstress tests -- distributions of examples with social relevance. We encourage\neach stakeholder to curate one or multiple stress tests containing examples\nreflecting their (possibly conflicting) interests. The machine passes or fails\neach stress test by falling short of or exceeding a pre-defined metric value.\nThe test results involve all stakeholders in a discussion about how to improve\nthe learning system, and provide flexible assessments of fairness dependent on\ncontext and based on interpretable data. We provide full implementation\nguidelines for stress testing, illustrate both the benefits and shortcomings of\nthis framework, and introduce a cryptographic scheme to enable a degree of\nprediction accountability from system providers.\n", "date": "2022-07-20", "categories": []}
{"id": "2207.10020", "title": "MANI-Rank: Multiple Attribute and Intersectional Group Fairness for\n  Consensus Ranking", "abstract": "  Combining the preferences of many rankers into one single consensus ranking\nis critical for consequential applications from hiring and admissions to\nlending. While group fairness has been extensively studied for classification,\ngroup fairness in rankings and in particular rank aggregation remains in its\ninfancy. Recent work introduced the concept of fair rank aggregation for\ncombining rankings but restricted to the case when candidates have a single\nbinary protected attribute, i.e., they fall into two groups only. Yet it\nremains an open problem how to create a consensus ranking that represents the\npreferences of all rankers while ensuring fair treatment for candidates with\nmultiple protected attributes such as gender, race, and nationality. In this\nwork, we are the first to define and solve this open Multi-attribute Fair\nConsensus Ranking (MFCR) problem. As a foundation, we design novel group\nfairness criteria for rankings, called MANI-RANK, ensuring fair treatment of\ngroups defined by individual protected attributes and their intersection.\nLeveraging the MANI-RANK criteria, we develop a series of algorithms that for\nthe first time tackle the MFCR problem. Our experimental study with a rich\nvariety of consensus scenarios demonstrates our MFCR methodology is the only\napproach to achieve both intersectional and protected attribute fairness while\nalso representing the preferences expressed through many base rankings. Our\nreal-world case study on merit scholarships illustrates the effectiveness of\nour MFCR methods to mitigate bias across multiple protected attributes and\ntheir intersections. This is an extended version of \"MANI-Rank: Multiple\nAttribute and Intersectional Group Fairness for Consensus Ranking\", to appear\nin ICDE 2022.\n", "date": "2022-07-20", "categories": []}
{"id": "2207.07072", "title": "A Query-Optimal Algorithm for Finding Counterfactuals", "abstract": "  We design an algorithm for finding counterfactuals with strong theoretical\nguarantees on its performance. For any monotone model $f : X^d \\to \\{0,1\\}$ and\ninstance $x^\\star$, our algorithm makes \\[ {S(f)^{O(\\Delta_f(x^\\star))}\\cdot\n\\log d}\\] queries to $f$ and returns {an {\\sl optimal}} counterfactual for\n$x^\\star$: a nearest instance $x'$ to $x^\\star$ for which $f(x')\\ne\nf(x^\\star)$. Here $S(f)$ is the sensitivity of $f$, a discrete analogue of the\nLipschitz constant, and $\\Delta_f(x^\\star)$ is the distance from $x^\\star$ to\nits nearest counterfactuals. The previous best known query complexity was\n$d^{\\,O(\\Delta_f(x^\\star))}$, achievable by brute-force local search. We\nfurther prove a lower bound of $S(f)^{\\Omega(\\Delta_f(x^\\star))} + \\Omega(\\log\nd)$ on the query complexity of any algorithm, thereby showing that the\nguarantees of our algorithm are essentially optimal.\n", "date": "2022-07-14", "categories": []}
{"id": "2207.05777", "title": "Long Term Fairness for Minority Groups via Performative Distributionally\n  Robust Optimization", "abstract": "  Fairness researchers in machine learning (ML) have coalesced around several\nfairness criteria which provide formal definitions of what it means for an ML\nmodel to be fair. However, these criteria have some serious limitations. We\nidentify four key shortcomings of these formal fairness criteria, and aim to\nhelp to address them by extending performative prediction to include a\ndistributionally robust objective.\n", "date": "2022-07-12", "categories": []}
{"id": "2207.05811", "title": "Revealing Unfair Models by Mining Interpretable Evidence", "abstract": "  The popularity of machine learning has increased the risk of unfair models\ngetting deployed in high-stake applications, such as justice system,\ndrug/vaccination design, and medical diagnosis. Although there are effective\nmethods to train fair models from scratch, how to automatically reveal and\nexplain the unfairness of a trained model remains a challenging task. Revealing\nunfairness of machine learning models in interpretable fashion is a critical\nstep towards fair and trustworthy AI. In this paper, we systematically tackle\nthe novel task of revealing unfair models by mining interpretable evidence\n(RUMIE). The key idea is to find solid evidence in the form of a group of data\ninstances discriminated most by the model. To make the evidence interpretable,\nwe also find a set of human-understandable key attributes and decision rules\nthat characterize the discriminated data instances and distinguish them from\nthe other non-discriminated data. As demonstrated by extensive experiments on\nmany real-world data sets, our method finds highly interpretable and solid\nevidence to effectively reveal the unfairness of trained models. Moreover, it\nis much more scalable than all of the baseline methods.\n", "date": "2022-07-12", "categories": []}
{"id": "2207.06084", "title": "Towards A Holistic View of Bias in Machine Learning: Bridging\n  Algorithmic Fairness and Imbalanced Learning", "abstract": "  Machine learning (ML) is playing an increasingly important role in rendering\ndecisions that affect a broad range of groups in society. ML models inform\ndecisions in criminal justice, the extension of credit in banking, and the\nhiring practices of corporations. This posits the requirement of model\nfairness, which holds that automated decisions should be equitable with respect\nto protected features (e.g., gender, race, or age) that are often\nunder-represented in the data. We postulate that this problem of\nunder-representation has a corollary to the problem of imbalanced data\nlearning. This class imbalance is often reflected in both classes and protected\nfeatures. For example, one class (those receiving credit) may be\nover-represented with respect to another class (those not receiving credit) and\na particular group (females) may be under-represented with respect to another\ngroup (males). A key element in achieving algorithmic fairness with respect to\nprotected groups is the simultaneous reduction of class and protected group\nimbalance in the underlying training data, which facilitates increases in both\nmodel accuracy and fairness. We discuss the importance of bridging imbalanced\nlearning and group fairness by showing how key concepts in these fields overlap\nand complement each other; and propose a novel oversampling algorithm, Fair\nOversampling, that addresses both skewed class distributions and protected\nfeatures. Our method: (i) can be used as an efficient pre-processing algorithm\nfor standard ML algorithms to jointly address imbalance and group equity; and\n(ii) can be combined with fairness-aware learning algorithms to improve their\nrobustness to varying levels of class imbalance. Additionally, we take a step\ntoward bridging the gap between fairness and imbalanced learning with a new\nmetric, Fair Utility, that combines balanced accuracy with fairness.\n", "date": "2022-07-13", "categories": []}
{"id": "2207.06273", "title": "Understanding Unfairness in Fraud Detection through Model and Data Bias\n  Interactions", "abstract": "  In recent years, machine learning algorithms have become ubiquitous in a\nmultitude of high-stakes decision-making applications. The unparalleled ability\nof machine learning algorithms to learn patterns from data also enables them to\nincorporate biases embedded within. A biased model can then make decisions that\ndisproportionately harm certain groups in society -- limiting their access to\nfinancial services, for example. The awareness of this problem has given rise\nto the field of Fair ML, which focuses on studying, measuring, and mitigating\nunfairness in algorithmic prediction, with respect to a set of protected groups\n(e.g., race or gender). However, the underlying causes for algorithmic\nunfairness still remain elusive, with researchers divided between blaming\neither the ML algorithms or the data they are trained on. In this work, we\nmaintain that algorithmic unfairness stems from interactions between models and\nbiases in the data, rather than from isolated contributions of either of them.\nTo this end, we propose a taxonomy to characterize data bias and we study a set\nof hypotheses regarding the fairness-accuracy trade-offs that fairness-blind ML\nalgorithms exhibit under different data bias settings. On our real-world\naccount-opening fraud use case, we find that each setting entails specific\ntrade-offs, affecting fairness in expected value and variance -- the latter\noften going unnoticed. Moreover, we show how algorithms compare differently in\nterms of accuracy and fairness, depending on the biases affecting the data.\nFinally, we note that under specific data bias conditions, simple\npre-processing interventions can successfully balance group-wise error rates,\nwhile the same techniques fail in more complex settings.\n", "date": "2022-07-13", "categories": []}
{"id": "2207.02912", "title": "Towards Substantive Conceptions of Algorithmic Fairness: Normative\n  Guidance from Equal Opportunity Doctrines", "abstract": "  In this work we use Equal Oppportunity (EO) doctrines from political\nphilosophy to make explicit the normative judgements embedded in different\nconceptions of algorithmic fairness. We contrast formal EO approaches that\nnarrowly focus on fair contests at discrete decision points, with substantive\nEO doctrines that look at people's fair life chances more holistically over the\ncourse of a lifetime. We use this taxonomy to provide a moral interpretation of\nthe impossibility results as the incompatibility between different conceptions\nof a fair contest -- foward-looking versus backward-looking -- when people do\nnot have fair life chances. We use this result to motivate substantive\nconceptions of algorithmic fairness and outline two plausible procedures based\non the luck-egalitarian doctrine of EO, and Rawls's principle of fair equality\nof opportunity.\n", "date": "2022-07-06", "categories": []}
{"id": "2207.02799", "title": "A multi-task network approach for calculating discrimination-free\n  insurance prices", "abstract": "  In applications of predictive modeling, such as insurance pricing, indirect\nor proxy discrimination is an issue of major concern. Namely, there exists the\npossibility that protected policyholder characteristics are implicitly inferred\nfrom non-protected ones by predictive models, and are thus having an\nundesirable (or illegal) impact on prices. A technical solution to this problem\nrelies on building a best-estimate model using all policyholder characteristics\n(including protected ones) and then averaging out the protected characteristics\nfor calculating individual prices. However, such approaches require full\nknowledge of policyholders' protected characteristics, which may in itself be\nproblematic. Here, we address this issue by using a multi-task neural network\narchitecture for claim predictions, which can be trained using only partial\ninformation on protected characteristics, and it produces prices that are free\nfrom proxy discrimination. We demonstrate the use of the proposed model and we\nfind that its predictive accuracy is comparable to a conventional feedforward\nneural network (on full information). However, this multi-task network has\nclearly superior performance in the case of partially missing policyholder\ninformation.\n", "date": "2022-07-06", "categories": []}
{"id": "2207.02842", "title": "When does Bias Transfer in Transfer Learning?", "abstract": "  Using transfer learning to adapt a pre-trained \"source model\" to a downstream\n\"target task\" can dramatically increase performance with seemingly no downside.\nIn this work, we demonstrate that there can exist a downside after all: bias\ntransfer, or the tendency for biases of the source model to persist even after\nadapting the model to the target class. Through a combination of synthetic and\nnatural experiments, we show that bias transfer both (a) arises in realistic\nsettings (such as when pre-training on ImageNet or other standard datasets) and\n(b) can occur even when the target dataset is explicitly de-biased. As\ntransfer-learned models are increasingly deployed in the real world, our work\nhighlights the importance of understanding the limitations of pre-trained\nsource models. Code is available at https://github.com/MadryLab/bias-transfer\n", "date": "2022-07-06", "categories": []}
{"id": "2207.00691", "title": "American == White in Multimodal Language-and-Image AI", "abstract": "  Three state-of-the-art language-and-image AI models, CLIP, SLIP, and BLIP,\nare evaluated for evidence of a bias previously observed in social and\nexperimental psychology: equating American identity with being White. Embedding\nassociation tests (EATs) using standardized images of self-identified Asian,\nBlack, Latina/o, and White individuals from the Chicago Face Database (CFD)\nreveal that White individuals are more associated with collective in-group\nwords than are Asian, Black, or Latina/o individuals. In assessments of three\ncore aspects of American identity reported by social psychologists,\nsingle-category EATs reveal that images of White individuals are more\nassociated with patriotism and with being born in America, but that, consistent\nwith prior findings in psychology, White individuals are associated with being\nless likely to treat people of all races and backgrounds equally. Three\ndownstream machine learning tasks demonstrate biases associating American with\nWhite. In a visual question answering task using BLIP, 97% of White individuals\nare identified as American, compared to only 3% of Asian individuals. When\nasked in what state the individual depicted lives in, the model responds China\n53% of the time for Asian individuals, but always with an American state for\nWhite individuals. In an image captioning task, BLIP remarks upon the race of\nAsian individuals as much as 36% of the time, but never remarks upon race for\nWhite individuals. Finally, provided with an initialization image from the CFD\nand the text \"an American person,\" a synthetic image generator (VQGAN) using\nthe text-based guidance of CLIP lightens the skin tone of individuals of all\nraces (by 35% for Black individuals, based on pixel brightness). The results\nindicate that biases equating American identity with being White are learned by\nlanguage-and-image AI, and propagate to downstream applications of such models.\n", "date": "2022-07-01", "categories": []}
{"id": "2207.01168", "title": "How Robust is Your Fairness? Evaluating and Sustaining Fairness under\n  Unseen Distribution Shifts", "abstract": "  Increasing concerns have been raised on deep learning fairness in recent\nyears. Existing fairness-aware machine learning methods mainly focus on the\nfairness of in-distribution data. However, in real-world applications, it is\ncommon to have distribution shift between the training and test data. In this\npaper, we first show that the fairness achieved by existing methods can be\neasily broken by slight distribution shifts. To solve this problem, we propose\na novel fairness learning method termed CUrvature MAtching (CUMA), which can\nachieve robust fairness generalizable to unseen domains with unknown\ndistributional shifts. Specifically, CUMA enforces the model to have similar\ngeneralization ability on the majority and minority groups, by matching the\nloss curvature distributions of the two groups. We evaluate our method on three\npopular fairness datasets. Compared with existing methods, CUMA achieves\nsuperior fairness under unseen distribution shifts, without sacrificing either\nthe overall accuracy or the in-distribution fairness.\n", "date": "2022-07-04", "categories": []}
{"id": "2206.14853", "title": "Fairness via In-Processing in the Over-parameterized Regime: A\n  Cautionary Tale", "abstract": "  The success of DNNs is driven by the counter-intuitive ability of\nover-parameterized networks to generalize, even when they perfectly fit the\ntraining data. In practice, test error often continues to decrease with\nincreasing over-parameterization, referred to as double descent. This allows\npractitioners to instantiate large models without having to worry about\nover-fitting. Despite its benefits, however, prior work has shown that\nover-parameterization can exacerbate bias against minority subgroups. Several\nfairness-constrained DNN training methods have been proposed to address this\nconcern. Here, we critically examine MinDiff, a fairness-constrained training\nprocedure implemented within TensorFlow's Responsible AI Toolkit, that aims to\nachieve Equality of Opportunity. We show that although MinDiff improves\nfairness for under-parameterized models, it is likely to be ineffective in the\nover-parameterized regime. This is because an overfit model with zero training\nloss is trivially group-wise fair on training data, creating an \"illusion of\nfairness,\" thus turning off the MinDiff optimization (this will apply to any\ndisparity-based measures which care about errors or accuracy. It won't apply to\ndemographic parity). Within specified fairness constraints, under-parameterized\nMinDiff models can even have lower error compared to their over-parameterized\ncounterparts (despite baseline over-parameterized models having lower error).\nWe further show that MinDiff optimization is very sensitive to choice of batch\nsize in the under-parameterized regime. Thus, fair model training using MinDiff\nrequires time-consuming hyper-parameter searches. Finally, we suggest using\npreviously proposed regularization techniques, viz. L2, early stopping and\nflooding in conjunction with MinDiff to train fair over-parameterized models.\n", "date": "2022-06-29", "categories": []}
{"id": "2206.15437", "title": "Understanding Instance-Level Impact of Fairness Constraints", "abstract": "  A variety of fairness constraints have been proposed in the literature to\nmitigate group-level statistical bias. Their impacts have been largely\nevaluated for different groups of populations corresponding to a set of\nsensitive attributes, such as race or gender. Nonetheless, the community has\nnot observed sufficient explorations for how imposing fairness constraints fare\nat an instance level. Building on the concept of influence function, a measure\nthat characterizes the impact of a training example on the target model and its\npredictive performance, this work studies the influence of training examples\nwhen fairness constraints are imposed. We find out that under certain\nassumptions, the influence function with respect to fairness constraints can be\ndecomposed into a kernelized combination of training examples. One promising\napplication of the proposed fairness influence function is to identify\nsuspicious training examples that may cause model discrimination by ranking\ntheir influence scores. We demonstrate with extensive experiments that training\non a subset of weighty data examples leads to lower fairness violations with a\ntrade-off of accuracy.\n", "date": "2022-06-30", "categories": []}
{"id": "2206.13757", "title": "Flexible text generation for counterfactual fairness probing", "abstract": "  A common approach for testing fairness issues in text-based classifiers is\nthrough the use of counterfactuals: does the classifier output change if a\nsensitive attribute in the input is changed? Existing counterfactual generation\nmethods typically rely on wordlists or templates, producing simple\ncounterfactuals that don't take into account grammar, context, or subtle\nsensitive attribute references, and could miss issues that the wordlist\ncreators had not considered. In this paper, we introduce a task for generating\ncounterfactuals that overcomes these shortcomings, and demonstrate how large\nlanguage models (LLMs) can be leveraged to make progress on this task. We show\nthat this LLM-based method can produce complex counterfactuals that existing\nmethods cannot, comparing the performance of various counterfactual generation\nmethods on the Civil Comments dataset and showing their value in evaluating a\ntoxicity classifier.\n", "date": "2022-06-28", "categories": []}
{"id": "2206.11436", "title": "Context matters for fairness -- a case study on the effect of spatial\n  distribution shifts", "abstract": "  With the ever growing involvement of data-driven AI-based decision making\ntechnologies in our daily social lives, the fairness of these systems is\nbecoming a crucial phenomenon. However, an important and often challenging\naspect in utilizing such systems is to distinguish validity for the range of\ntheir application especially under distribution shifts, i.e., when a model is\ndeployed on data with different distribution than the training set. In this\npaper, we present a case study on the newly released American Census datasets,\na reconstruction of the popular Adult dataset, to illustrate the importance of\ncontext for fairness and show how remarkably can spatial distribution shifts\naffect predictive- and fairness-related performance of a model. The problem\npersists for fairness-aware learning models with the effects of\ncontext-specific fairness interventions differing across the states and\ndifferent population groups. Our study suggests that robustness to distribution\nshifts is necessary before deploying a model to another context.\n", "date": "2022-06-23", "categories": []}
{"id": "2206.12739", "title": "On how to avoid exacerbating spurious correlations when models are\n  overparameterized", "abstract": "  Overparameterized models fail to generalize well in the presence of data\nimbalance even when combined with traditional techniques for mitigating\nimbalances. This paper focuses on imbalanced classification datasets, in which\na small subset of the population -- a minority -- may contain features that\ncorrelate spuriously with the class label. For a parametric family of\ncross-entropy loss modifications and a representative Gaussian mixture model,\nwe derive non-asymptotic generalization bounds on the worst-group error that\nshed light on the role of different hyper-parameters. Specifically, we prove\nthat, when appropriately tuned, the recently proposed VS-loss learns a model\nthat is fair towards minorities even when spurious features are strong. On the\nother hand, alternative heuristics, such as the weighted CE and the LA-loss,\ncan fail dramatically. Compared to previous works, our bounds hold for more\ngeneral models, they are non-asymptotic, and, they apply even at scenarios of\nextreme imbalance.\n", "date": "2022-06-25", "categories": []}
{"id": "2206.13183", "title": "Prisoners of Their Own Devices: How Models Induce Data Bias in\n  Performative Prediction", "abstract": "  The unparalleled ability of machine learning algorithms to learn patterns\nfrom data also enables them to incorporate biases embedded within. A biased\nmodel can then make decisions that disproportionately harm certain groups in\nsociety. Much work has been devoted to measuring unfairness in static ML\nenvironments, but not in dynamic, performative prediction ones, in which most\nreal-world use cases operate. In the latter, the predictive model itself plays\na pivotal role in shaping the distribution of the data. However, little\nattention has been heeded to relating unfairness to these interactions. Thus,\nto further the understanding of unfairness in these settings, we propose a\ntaxonomy to characterize bias in the data, and study cases where it is shaped\nby model behaviour. Using a real-world account opening fraud detection case\nstudy as an example, we study the dangers to both performance and fairness of\ntwo typical biases in performative prediction: distribution shifts, and the\nproblem of selective labels.\n", "date": "2022-06-27", "categories": []}
{"id": "2206.11993", "title": "A Disability Lens towards Biases in GPT-3 Generated Open-Ended Languages", "abstract": "  Language models (LM) are becoming prevalent in many language-based\napplication spaces globally. Although these LMs are improving our day-to-day\ninteractions with digital products, concerns remain whether open-ended\nlanguages or text generated from these models reveal any biases toward a\nspecific group of people, thereby risking the usability of a certain product.\nThere is a need to identify whether these models possess bias to improve the\nfairness in these models. This gap motivates our ongoing work, where we\nmeasured the two aspects of bias in GPT-3 generated text through a disability\nlens.\n", "date": "2022-06-23", "categories": []}
{"id": "2206.12333", "title": "Achievement and Fragility of Long-term Equitability", "abstract": "  Equipping current decision-making tools with notions of fairness,\nequitability, or other ethically motivated outcomes, is one of the top\npriorities in recent research efforts in machine learning, AI, and\noptimization. In this paper, we investigate how to allocate limited resources\nto {locally interacting} communities in a way to maximize a pertinent notion of\nequitability. In particular, we look at the dynamic setting where the\nallocation is repeated across multiple periods (e.g., yearly), the local\ncommunities evolve in the meantime (driven by the provided allocation), and the\nallocations are modulated by feedback coming from the communities themselves.\nWe employ recent mathematical tools stemming from data-driven feedback online\noptimization, by which communities can learn their (possibly unknown)\nevolution, satisfaction, as well as they can share information with the\ndeciding bodies. We design dynamic policies that converge to an allocation that\nmaximize equitability in the long term. We further demonstrate our model and\nmethodology with realistic examples of healthcare and education subsidies\ndesign in Sub-Saharian countries. One of the key empirical takeaways from our\nsetting is that long-term equitability is fragile, in the sense that it can be\neasily lost when deciding bodies weigh in other factors (e.g., equality in\nallocation) in the allocation strategy. Moreover, a naive compromise, while not\nproviding significant advantage to the communities, can promote inequality in\nsocial outcomes.\n", "date": "2022-06-24", "categories": []}
{"id": "2206.11423", "title": "Input-agnostic Certified Group Fairness via Gaussian Parameter Smoothing", "abstract": "  Only recently, researchers attempt to provide classification algorithms with\nprovable group fairness guarantees. Most of these algorithms suffer from\nharassment caused by the requirement that the training and deployment data\nfollow the same distribution. This paper proposes an input-agnostic certified\ngroup fairness algorithm, FairSmooth, for improving the fairness of\nclassification models while maintaining the remarkable prediction accuracy. A\nGaussian parameter smoothing method is developed to transform base classifiers\ninto their smooth versions. An optimal individual smooth classifier is learnt\nfor each group with only the data regarding the group and an overall smooth\nclassifier for all groups is generated by averaging the parameters of all the\nindividual smooth ones. By leveraging the theory of nonlinear functional\nanalysis, the smooth classifiers are reformulated as output functions of a\nNemytskii operator. Theoretical analysis is conducted to derive that the\nNemytskii operator is smooth and induces a Frechet differentiable smooth\nmanifold. We theoretically demonstrate that the smooth manifold has a global\nLipschitz constant that is independent of the domain of the input data, which\nderives the input-agnostic certified group fairness.\n", "date": "2022-06-22", "categories": []}
{"id": "2206.09076", "title": "Fair Generalized Linear Models with a Convex Penalty", "abstract": "  Despite recent advances in algorithmic fairness, methodologies for achieving\nfairness with generalized linear models (GLMs) have yet to be explored in\ngeneral, despite GLMs being widely used in practice. In this paper we introduce\ntwo fairness criteria for GLMs based on equalizing expected outcomes or\nlog-likelihoods. We prove that for GLMs both criteria can be achieved via a\nconvex penalty term based solely on the linear components of the GLM, thus\npermitting efficient optimization. We also derive theoretical properties for\nthe resulting fair GLM estimator. To empirically demonstrate the efficacy of\nthe proposed fair GLM, we compare it with other well-known fair prediction\nmethods on an extensive set of benchmark datasets for binary classification and\nregression. In addition, we demonstrate that the fair GLM can generate fair\npredictions for a range of response variables, other than binary and continuous\noutcomes.\n", "date": "2022-06-18", "categories": []}
{"id": "2206.09346", "title": "Fairness-aware Model-agnostic Positive and Unlabeled Learning", "abstract": "  With the increasing application of machine learning in high-stake\ndecision-making problems, potential algorithmic bias towards people from\ncertain social groups poses negative impacts on individuals and our society at\nlarge. In the real-world scenario, many such problems involve positive and\nunlabeled data such as medical diagnosis, criminal risk assessment and\nrecommender systems. For instance, in medical diagnosis, only the diagnosed\ndiseases will be recorded (positive) while others will not (unlabeled). Despite\nthe large amount of existing work on fairness-aware machine learning in the\n(semi-)supervised and unsupervised settings, the fairness issue is largely\nunder-explored in the aforementioned Positive and Unlabeled Learning (PUL)\ncontext, where it is usually more severe. In this paper, to alleviate this\ntension, we propose a fairness-aware PUL method named FairPUL. In particular,\nfor binary classification over individuals from two populations, we aim to\nachieve similar true positive rates and false positive rates in both\npopulations as our fairness metric. Based on the analysis of the optimal fair\nclassifier for PUL, we design a model-agnostic post-processing framework,\nleveraging both the positive examples and unlabeled ones. Our framework is\nproven to be statistically consistent in terms of both the classification error\nand the fairness metric. Experiments on the synthetic and real-world data sets\ndemonstrate that our framework outperforms state-of-the-art in both PUL and\nfair classification.\n", "date": "2022-06-19", "categories": []}
{"id": "2206.09860", "title": "Fewer Errors, but More Stereotypes? The Effect of Model Size on Gender\n  Bias", "abstract": "  The size of pretrained models is increasing, and so is their performance on a\nvariety of NLP tasks. However, as their memorization capacity grows, they might\npick up more social biases. In this work, we examine the connection between\nmodel size and its gender bias (specifically, occupational gender bias). We\nmeasure bias in three masked language model families (RoBERTa, DeBERTa, and T5)\nin two setups: directly using prompt based method, and using a downstream task\n(Winogender). We find on the one hand that larger models receive higher bias\nscores on the former task, but when evaluated on the latter, they make fewer\ngender errors. To examine these potentially conflicting results, we carefully\ninvestigate the behavior of the different models on Winogender. We find that\nwhile larger models outperform smaller ones, the probability that their\nmistakes are caused by gender bias is higher. Moreover, we find that the\nproportion of stereotypical errors compared to anti-stereotypical ones grows\nwith the model size. Our findings highlight the potential risks that can arise\nfrom increasing model size.\n", "date": "2022-06-20", "categories": []}
{"id": "2206.09875", "title": "Algorithmic Fairness and Vertical Equity: Income Fairness with IRS Tax\n  Audit Models", "abstract": "  This study examines issues of algorithmic fairness in the context of systems\nthat inform tax audit selection by the United States Internal Revenue Service\n(IRS). While the field of algorithmic fairness has developed primarily around\nnotions of treating like individuals alike, we instead explore the concept of\nvertical equity -- appropriately accounting for relevant differences across\nindividuals -- which is a central component of fairness in many public policy\nsettings. Applied to the design of the U.S. individual income tax system,\nvertical equity relates to the fair allocation of tax and enforcement burdens\nacross taxpayers of different income levels. Through a unique collaboration\nwith the Treasury Department and IRS, we use access to anonymized individual\ntaxpayer microdata, risk-selected audits, and random audits from 2010-14 to\nstudy vertical equity in tax administration. In particular, we assess how the\nuse of modern machine learning methods for selecting audits may affect vertical\nequity. First, we show how the use of more flexible machine learning\n(classification) methods -- as opposed to simpler models -- shifts audit\nburdens from high to middle-income taxpayers. Second, we show that while\nexisting algorithmic fairness techniques can mitigate some disparities across\nincome, they can incur a steep cost to performance. Third, we show that the\nchoice of whether to treat risk of underreporting as a classification or\nregression problem is highly consequential. Moving from classification to\nregression models to predict underreporting shifts audit burden substantially\ntoward high income individuals, while increasing revenue. Last, we explore the\nrole of differential audit cost in shaping the audit distribution. We show that\na narrow focus on return-on-investment can undermine vertical equity. Our\nresults have implications for the design of algorithmic tools across the public\nsector.\n", "date": "2022-06-20", "categories": []}
{"id": "2206.07826", "title": "Metric-Fair Classifier Derandomization", "abstract": "  We study the problem of classifier derandomization in machine learning: given\na stochastic binary classifier $f: X \\to [0,1]$, sample a deterministic\nclassifier $\\hat{f}: X \\to \\{0,1\\}$ that approximates the output of $f$ in\naggregate over any data distribution. Recent work revealed how to efficiently\nderandomize a stochastic classifier with strong output approximation\nguarantees, but at the cost of individual fairness -- that is, if $f$ treated\nsimilar inputs similarly, $\\hat{f}$ did not. In this paper, we initiate a\nsystematic study of classifier derandomization with metric fairness guarantees.\nWe show that the prior derandomization approach is almost maximally\nmetric-unfair, and that a simple ``random threshold'' derandomization achieves\noptimal fairness preservation but with weaker output approximation. We then\ndevise a derandomization procedure that provides an appealing tradeoff between\nthese two: if $f$ is $\\alpha$-metric fair according to a metric $d$ with a\nlocality-sensitive hash (LSH) family, then our derandomized $\\hat{f}$ is, with\nhigh probability, $O(\\alpha)$-metric fair and a close approximation of $f$. We\nalso prove generic results applicable to all (fair and unfair) classifier\nderandomization procedures, including a bias-variance decomposition and\nreductions between various notions of metric fairness.\n", "date": "2022-06-15", "categories": []}
{"id": "2206.08403", "title": "Learning to Teach Fairness-aware Deep Multi-task Learning", "abstract": "  Fairness-aware learning mainly focuses on single task learning (STL). The\nfairness implications of multi-task learning (MTL) have only recently been\nconsidered and a seminal approach has been proposed that considers the\nfairness-accuracy trade-off for each task and the performance trade-off among\ndifferent tasks. Instead of a rigid fairness-accuracy trade-off formulation, we\npropose a flexible approach that learns how to be fair in a MTL setting by\nselecting which objective (accuracy or fairness) to optimize at each step. We\nintroduce the L2T-FMT algorithm that is a teacher-student network trained\ncollaboratively; the student learns to solve the fair MTL problem while the\nteacher instructs the student to learn from either accuracy or fairness,\ndepending on what is harder to learn for each task. Moreover, this dynamic\nselection of which objective to use at each step for each task reduces the\nnumber of trade-off weights from 2T to T, where T is the number of tasks. Our\nexperiments on three real datasets show that L2T-FMT improves on both fairness\n(12-19%) and accuracy (up to 2%) over state-of-the-art approaches.\n", "date": "2022-06-16", "categories": []}
{"id": "2206.08450", "title": "Active Fairness Auditing", "abstract": "  The fast spreading adoption of machine learning (ML) by companies across\nindustries poses significant regulatory challenges. One such challenge is\nscalability: how can regulatory bodies efficiently audit these ML models,\nensuring that they are fair? In this paper, we initiate the study of\nquery-based auditing algorithms that can estimate the demographic parity of ML\nmodels in a query-efficient manner. We propose an optimal deterministic\nalgorithm, as well as a practical randomized, oracle-efficient algorithm with\ncomparable guarantees. Furthermore, we make inroads into understanding the\noptimal query complexity of randomized active fairness estimation algorithms.\nOur first exploration of active fairness estimation aims to put AI governance\non firmer theoretical foundations.\n", "date": "2022-06-16", "categories": []}
{"id": "2206.08454", "title": "Quantifying Feature Contributions to Overall Disparity Using Information\n  Theory", "abstract": "  When a machine-learning algorithm makes biased decisions, it can be helpful\nto understand the sources of disparity to explain why the bias exists. Towards\nthis, we examine the problem of quantifying the contribution of each individual\nfeature to the observed disparity. If we have access to the decision-making\nmodel, one potential approach (inspired from intervention-based approaches in\nexplainability literature) is to vary each individual feature (while keeping\nthe others fixed) and use the resulting change in disparity to quantify its\ncontribution. However, we may not have access to the model or be able to\ntest/audit its outputs for individually varying features. Furthermore, the\ndecision may not always be a deterministic function of the input features\n(e.g., with human-in-the-loop). For these situations, we might need to explain\ncontributions using purely distributional (i.e., observational) techniques,\nrather than interventional. We ask the question: what is the \"potential\"\ncontribution of each individual feature to the observed disparity in the\ndecisions when the exact decision-making mechanism is not accessible? We first\nprovide canonical examples (thought experiments) that help illustrate the\ndifference between distributional and interventional approaches to explaining\ncontributions, and when either is better suited. When unable to intervene on\nthe inputs, we quantify the \"redundant\" statistical dependency about the\nprotected attribute that is present in both the final decision and an\nindividual feature, by leveraging a body of work in information theory called\nPartial Information Decomposition. We also perform a simple case study to show\nhow this technique could be applied to quantify contributions.\n", "date": "2022-06-16", "categories": []}
{"id": "2206.07801", "title": "Beyond Adult and COMPAS: Fairness in Multi-Class Prediction", "abstract": "  We consider the problem of producing fair probabilistic classifiers for\nmulti-class classification tasks. We formulate this problem in terms of\n\"projecting\" a pre-trained (and potentially unfair) classifier onto the set of\nmodels that satisfy target group-fairness requirements. The new, projected\nmodel is given by post-processing the outputs of the pre-trained classifier by\na multiplicative factor. We provide a parallelizable iterative algorithm for\ncomputing the projected classifier and derive both sample complexity and\nconvergence guarantees. Comprehensive numerical comparisons with\nstate-of-the-art benchmarks demonstrate that our approach maintains competitive\nperformance in terms of accuracy-fairness trade-off curves, while achieving\nfavorable runtime on large datasets. We also evaluate our method at scale on an\nopen dataset with multiple classes, multiple intersectional protected groups,\nand over 1M samples.\n", "date": "2022-06-15", "categories": []}
{"id": "2206.07173", "title": "Measuring Representational Harms in Image Captioning", "abstract": "  Previous work has largely considered the fairness of image captioning systems\nthrough the underspecified lens of \"bias.\" In contrast, we present a set of\ntechniques for measuring five types of representational harms, as well as the\nresulting measurements obtained for two of the most popular image captioning\ndatasets using a state-of-the-art image captioning system. Our goal was not to\naudit this image captioning system, but rather to develop normatively grounded\nmeasurement techniques, in turn providing an opportunity to reflect on the many\nchallenges involved. We propose multiple measurement techniques for each type\nof harm. We argue that by doing so, we are better able to capture the\nmulti-faceted nature of each type of harm, in turn improving the (collective)\nvalidity of the resulting measurements. Throughout, we discuss the assumptions\nunderlying our measurement approach and point out when they do not hold.\n", "date": "2022-06-14", "categories": []}
{"id": "2206.06685", "title": "Causal Discovery for Fairness", "abstract": "  It is crucial to consider the social and ethical consequences of AI and ML\nbased decisions for the safe and acceptable use of these emerging technologies.\nFairness, in particular, guarantees that the ML decisions do not result in\ndiscrimination against individuals or minorities. Identifying and measuring\nreliably fairness/discrimination is better achieved using causality which\nconsiders the causal relation, beyond mere association, between the sensitive\nattribute (e.g. gender, race, religion, etc.) and the decision (e.g. job\nhiring, loan granting, etc.). The big impediment to the use of causality to\naddress fairness, however, is the unavailability of the causal model (typically\nrepresented as a causal graph). Existing causal approaches to fairness in the\nliterature do not address this problem and assume that the causal model is\navailable. In this paper, we do not make such assumption and we review the\nmajor algorithms to discover causal relations from observable data. This study\nfocuses on causal discovery and its impact on fairness. In particular, we show\nhow different causal discovery approaches may result in different causal models\nand, most importantly, how even slight differences between causal models can\nhave significant impact on fairness/discrimination conclusions. These results\nare consolidated by empirical analysis using synthetic and standard fairness\nbenchmark datasets. The main goal of this study is to highlight the importance\nof the causal discovery step to appropriately address fairness using causality.\n", "date": "2022-06-14", "categories": []}
{"id": "2206.06960", "title": "ABCinML: Anticipatory Bias Correction in Machine Learning Applications", "abstract": "  The idealization of a static machine-learned model, trained once and deployed\nforever, is not practical. As input distributions change over time, the model\nwill not only lose accuracy, any constraints to reduce bias against a protected\nclass may fail to work as intended. Thus, researchers have begun to explore\nways to maintain algorithmic fairness over time. One line of work focuses on\ndynamic learning: retraining after each batch, and the other on robust learning\nwhich tries to make algorithms robust against all possible future changes.\nDynamic learning seeks to reduce biases soon after they have occurred and\nrobust learning often yields (overly) conservative models. We propose an\nanticipatory dynamic learning approach for correcting the algorithm to mitigate\nbias before it occurs. Specifically, we make use of anticipations regarding the\nrelative distributions of population subgroups (e.g., relative ratios of male\nand female applicants) in the next cycle to identify the right parameters for\nan importance weighing fairness approach. Results from experiments over\nmultiple real-world datasets suggest that this approach has promise for\nanticipatory bias correction.\n", "date": "2022-06-14", "categories": []}
{"id": "2206.03426", "title": "Improving Fairness in Graph Neural Networks via Mitigating Sensitive\n  Attribute Leakage", "abstract": "  Graph Neural Networks (GNNs) have shown great power in learning node\nrepresentations on graphs. However, they may inherit historical prejudices from\ntraining data, leading to discriminatory bias in predictions. Although some\nwork has developed fair GNNs, most of them directly borrow fair representation\nlearning techniques from non-graph domains without considering the potential\nproblem of sensitive attribute leakage caused by feature propagation in GNNs.\nHowever, we empirically observe that feature propagation could vary the\ncorrelation of previously innocuous non-sensitive features to the sensitive\nones. This can be viewed as a leakage of sensitive information which could\nfurther exacerbate discrimination in predictions. Thus, we design two feature\nmasking strategies according to feature correlations to highlight the\nimportance of considering feature propagation and correlation variation in\nalleviating discrimination. Motivated by our analysis, we propose Fair View\nGraph Neural Network (FairVGNN) to generate fair views of features by\nautomatically identifying and masking sensitive-correlated features considering\ncorrelation variation after feature propagation. Given the learned fair views,\nwe adaptively clamp weights of the encoder to avoid using sensitive-related\nfeatures. Experiments on real-world datasets demonstrate that FairVGNN enjoys a\nbetter trade-off between model utility and fairness. Our code is publicly\navailable at https://github.com/YuWVandy/FairVGNN.\n", "date": "2022-06-07", "categories": []}
{"id": "2206.04789", "title": "Comprehensive Fair Meta-learned Recommender System", "abstract": "  In recommender systems, one common challenge is the cold-start problem, where\ninteractions are very limited for fresh users in the systems. To address this\nchallenge, recently, many works introduce the meta-optimization idea into the\nrecommendation scenarios, i.e. learning to learn the user preference by only a\nfew past interaction items. The core idea is to learn global shared\nmeta-initialization parameters for all users and rapidly adapt them into local\nparameters for each user respectively. They aim at deriving general knowledge\nacross preference learning of various users, so as to rapidly adapt to the\nfuture new user with the learned prior and a small amount of training data.\nHowever, previous works have shown that recommender systems are generally\nvulnerable to bias and unfairness. Despite the success of meta-learning at\nimproving the recommendation performance with cold-start, the fairness issues\nare largely overlooked. In this paper, we propose a comprehensive fair\nmeta-learning framework, named CLOVER, for ensuring the fairness of\nmeta-learned recommendation models. We systematically study three kinds of\nfairness - individual fairness, counterfactual fairness, and group fairness in\nthe recommender systems, and propose to satisfy all three kinds via a\nmulti-task adversarial learning scheme. Our framework offers a generic training\nparadigm that is applicable to different meta-learned recommender systems. We\ndemonstrate the effectiveness of CLOVER on the representative meta-learned user\npreference estimator on three real-world data sets. Empirical results show that\nCLOVER achieves comprehensive fairness without deteriorating the overall\ncold-start recommendation performance.\n", "date": "2022-06-09", "categories": []}
{"id": "2206.05050", "title": "Improved Approximation for Fair Correlation Clustering", "abstract": "  Correlation clustering is a ubiquitous paradigm in unsupervised machine\nlearning where addressing unfairness is a major challenge. Motivated by this,\nwe study Fair Correlation Clustering where the data points may belong to\ndifferent protected groups and the goal is to ensure fair representation of all\ngroups across clusters. Our paper significantly generalizes and improves on the\nquality guarantees of previous work of Ahmadi et al. and Ahmadian et al. as\nfollows.\n  - We allow the user to specify an arbitrary upper bound on the representation\nof each group in a cluster.\n  - Our algorithm allows individuals to have multiple protected features and\nensure fairness simultaneously across them all.\n  - We prove guarantees for clustering quality and fairness in this general\nsetting. Furthermore, this improves on the results for the special cases\nstudied in previous work. Our experiments on real-world data demonstrate that\nour clustering quality compared to the optimal solution is much better than\nwhat our theoretical result suggests.\n", "date": "2022-06-09", "categories": []}
{"id": "2206.04475", "title": "Individually Fair Learning with One-Sided Feedback", "abstract": "  We consider an online learning problem with one-sided feedback, in which the\nlearner is able to observe the true label only for positively predicted\ninstances. On each round, $k$ instances arrive and receive classification\noutcomes according to a randomized policy deployed by the learner, whose goal\nis to maximize accuracy while deploying individually fair policies. We first\nextend the framework of Bechavod et al. (2020), which relies on the existence\nof a human fairness auditor for detecting fairness violations, to instead\nincorporate feedback from dynamically-selected panels of multiple, possibly\ninconsistent, auditors. We then construct an efficient reduction from our\nproblem of online learning with one-sided feedback and a panel reporting\nfairness violations to the contextual combinatorial semi-bandit problem\n(Cesa-Bianchi & Lugosi, 2009, Gy\\\"{o}rgy et al., 2007). Finally, we show how to\nleverage the guarantees of two algorithms in the contextual combinatorial\nsemi-bandit setting: Exp2 (Bubeck et al., 2012) and the oracle-efficient\nContext-Semi-Bandit-FTPL (Syrgkanis et al., 2016), to provide multi-criteria no\nregret guarantees simultaneously for accuracy and fairness. Our results\neliminate two potential sources of bias from prior work: the \"hidden outcomes\"\nthat are not available to an algorithm operating in the full information\nsetting, and human biases that might be present in any single human auditor,\nbut can be mitigated by selecting a well chosen panel.\n", "date": "2022-06-09", "categories": []}
{"id": "2206.03515", "title": "How does overparametrization affect performance on minority groups?", "abstract": "  The benefits of overparameterization for the overall performance of modern\nmachine learning (ML) models are well known. However, the effect of\noverparameterization at a more granular level of data subgroups is less\nunderstood. Recent empirical studies demonstrate encouraging results: (i) when\ngroups are not known, overparameterized models trained with empirical risk\nminimization (ERM) perform better on minority groups; (ii) when groups are\nknown, ERM on data subsampled to equalize group sizes yields state-of-the-art\nworst-group-accuracy in the overparameterized regime. In this paper, we\ncomplement these empirical studies with a theoretical investigation of the risk\nof overparameterized random feature models on minority groups. In a setting in\nwhich the regression functions for the majority and minority groups are\ndifferent, we show that overparameterization always improves minority group\nperformance.\n", "date": "2022-06-07", "categories": []}
{"id": "2206.02792", "title": "FIFA: Making Fairness More Generalizable in Classifiers Trained on\n  Imbalanced Data", "abstract": "  Algorithmic fairness plays an important role in machine learning and imposing\nfairness constraints during learning is a common approach. However, many\ndatasets are imbalanced in certain label classes (e.g. \"healthy\") and sensitive\nsubgroups (e.g. \"older patients\"). Empirically, this imbalance leads to a lack\nof generalizability not only of classification, but also of fairness\nproperties, especially in over-parameterized models. For example,\nfairness-aware training may ensure equalized odds (EO) on the training data,\nbut EO is far from being satisfied on new users. In this paper, we propose a\ntheoretically-principled, yet Flexible approach that is\nImbalance-Fairness-Aware (FIFA). Specifically, FIFA encourages both\nclassification and fairness generalization and can be flexibly combined with\nmany existing fair learning methods with logits-based losses. While our main\nfocus is on EO, FIFA can be directly applied to achieve equalized opportunity\n(EqOpt); and under certain conditions, it can also be applied to other fairness\nnotions. We demonstrate the power of FIFA by combining it with a popular fair\nclassification algorithm, and the resulting algorithm achieves significantly\nbetter fairness generalization on several real-world datasets.\n", "date": "2022-06-06", "categories": []}
{"id": "2206.02237", "title": "Enforcing Group Fairness in Algorithmic Decision Making: Utility\n  Maximization Under Sufficiency", "abstract": "  Binary decision making classifiers are not fair by default. Fairness\nrequirements are an additional element to the decision making rationale, which\nis typically driven by maximizing some utility function. In that sense,\nalgorithmic fairness can be formulated as a constrained optimization problem.\nThis paper contributes to the discussion on how to implement fairness, focusing\non the fairness concepts of positive predictive value (PPV) parity, false\nomission rate (FOR) parity, and sufficiency (which combines the former two). We\nshow that group-specific threshold rules are optimal for PPV parity and FOR\nparity, similar to well-known results for other group fairness criteria.\nHowever, depending on the underlying population distributions and the utility\nfunction, we find that sometimes an upper-bound threshold rule for one group is\noptimal: utility maximization under PPV parity (or FOR parity) might thus lead\nto selecting the individuals with the smallest utility for one group, instead\nof selecting the most promising individuals. This result is counter-intuitive\nand in contrast to the analogous solutions for statistical parity and equality\nof opportunity. We also provide a solution for the optimal decision rules\nsatisfying the fairness constraint sufficiency. We show that more complex\ndecision rules are required and that this leads to within-group unfairness for\nall but one of the groups. We illustrate our findings based on simulated and\nreal data.\n", "date": "2022-06-05", "categories": []}
{"id": "2206.01326", "title": "Improving Fairness in Large-Scale Object Recognition by CrowdSourced\n  Demographic Information", "abstract": "  There has been increasing awareness of ethical issues in machine learning,\nand fairness has become an important research topic. Most fairness efforts in\ncomputer vision have been focused on human sensing applications and preventing\ndiscrimination by people's physical attributes such as race, skin color or age\nby increasing visual representation for particular demographic groups. We argue\nthat ML fairness efforts should extend to object recognition as well.\nBuildings, artwork, food and clothing are examples of the objects that define\nhuman culture. Representing these objects fairly in machine learning datasets\nwill lead to models that are less biased towards a particular culture and more\ninclusive of different traditions and values. There exist many research\ndatasets for object recognition, but they have not carefully considered which\nclasses should be included, or how much training data should be collected per\nclass. To address this, we propose a simple and general approach, based on\ncrowdsourcing the demographic composition of the contributors: we define fair\nrelevance scores, estimate them, and assign them to each class. We showcase its\napplication to the landmark recognition domain, presenting a detailed analysis\nand the final fairer landmark rankings. We present analysis which leads to a\nmuch fairer coverage of the world compared to existing datasets. The evaluation\ndataset was used for the 2021 Google Landmark Challenges, which was the first\nof a kind with an emphasis on fairness in generic object recognition.\n", "date": "2022-06-02", "categories": []}
{"id": "2206.01691", "title": "Measuring Gender Bias in Word Embeddings of Gendered Languages Requires\n  Disentangling Grammatical Gender Signals", "abstract": "  Does the grammatical gender of a language interfere when measuring the\nsemantic gender information captured by its word embeddings? A number of\nanomalous gender bias measurements in the embeddings of gendered languages\nsuggest this possibility. We demonstrate that word embeddings learn the\nassociation between a noun and its grammatical gender in grammatically gendered\nlanguages, which can skew social gender bias measurements. Consequently, word\nembedding post-processing methods are introduced to quantify, disentangle, and\nevaluate grammatical gender signals. The evaluation is performed on five\ngendered languages from the Germanic, Romance, and Slavic branches of the\nIndo-European language family. Our method reduces the strength of grammatical\ngender signals, which is measured in terms of effect size (Cohen's d), by a\nsignificant average of d = 1.3 for French, German, and Italian, and d = 0.56\nfor Polish and Spanish. Once grammatical gender is disentangled, the\nassociation between over 90% of 10,000 inanimate nouns and their assigned\ngrammatical gender weakens, and cross-lingual bias results from the Word\nEmbedding Association Test (WEAT) become more congruent with country-level\nimplicit bias measurements. The results further suggest that disentangling\ngrammatical gender signals from word embeddings may lead to improvement in\nsemantic machine learning tasks.\n", "date": "2022-06-03", "categories": []}
{"id": "2206.00701", "title": "What Changed? Investigating Debiasing Methods using Causal Mediation\n  Analysis", "abstract": "  Previous work has examined how debiasing language models affect downstream\ntasks, specifically, how debiasing techniques influence task performance and\nwhether debiased models also make impartial predictions in downstream tasks or\nnot. However, what we don't understand well yet is why debiasing methods have\nvarying impacts on downstream tasks and how debiasing techniques affect\ninternal components of language models, i.e., neurons, layers, and attentions.\nIn this paper, we decompose the internal mechanisms of debiasing language\nmodels with respect to gender by applying causal mediation analysis to\nunderstand the influence of debiasing methods on toxicity detection as a\ndownstream task. Our findings suggest a need to test the effectiveness of\ndebiasing methods with different bias metrics, and to focus on changes in the\nbehavior of certain components of the models, e.g.,first two layers of language\nmodels, and attention heads.\n", "date": "2022-06-01", "categories": []}
{"id": "2606.22858", "title": "The Unseen Hand: Manipulating Model Fairness and SHAP with Targeted Identity Re-Association Attacks", "abstract": "As machine learning models grow more influential and opaque, algorithmic fairness and explainability are critical for ensuring accountability. However, we demonstrate that these auditing mechanisms are themselves vulnerable to subtle manipulation, camouflaging the influence of protected features. While prior work on data-agnostic attacks has exposed this vulnerability, they leave behind detectable artifacts that compromise their stealth. We introduce Targeted Identity Re-Association (TIRA) attacks, a novel family of attacks that iteratively and probabilistically manipulate a model's outputs without requiring access to the model's internals or feature representations. We formalize two algorithms: Probabilistic Micro-Shuffling (PMiS), which applies localized adjacent swaps, and Probabilistic Rank-Shift Micro-Perturbation (PRSMP), which introduces small, randomized rank shifts. We empirically demonstrate that TIRA attacks are highly effective at pushing fairness metrics towards ideal values. Crucially, TIRA attacks successfully confound SHAP-based explanations, leaving effectively zero residual attribution for protected features, a major improvement over prior work.", "date": "2026-06-22", "categories": []}
{"id": "2606.23515", "title": "FairBED: A Bayesian Experimental Design Approach to Gathering Fairer Data", "abstract": "Frameworks for ensuring fairness in machine learning typically focus on learning fair models from existing data. But this endeavor is often undermined by biases already present in that data. We therefore look to modify the data acquisition process itself to help gather fairer data that is inherently more suitable for training fair predictors. To this end, we introduce FairBED, which provides novel formulations for quantifying the fairness of datasets themselves based on the idea that fair datasets should be uninformative about sensitive attributes. We then use this to construct practical fairness-aware Bayesian experimental design (BED) objectives that maximize expected information gain about the target quantity of interest while minimizing expected information gain about sensitive attributes. We further derive a theoretical link between FairBED and demographic parity, and show empirically that models trained on data gathered using FairBED provide improved fairness-accuracy trade-offs compared to randomly acquired data and conventional BED.", "date": "2026-06-22", "categories": []}
{"id": "2606.24161", "title": "Dual-Branch Cross-Projection Debiasing through Diffusion-based Disentanglement", "abstract": "Foundation models trained on biased datasets often rely on spurious correlations between target labels and non-causal attributes, resulting in poor generalization on minority groups. Bias mitigation remains challenging due to two fundamental issues. First, when group labels are unavailable, existing group-unsupervised methods typically infer spurious attributes implicitly from model behavior, making it difficult to identify spurious factors that are semantically aligned with real-world biases. Second, even with pseudo spurious supervision, most existing debiasing methods follow a single-branch design that operates within a single shared feature space, where target and spurious attributes are intrinsically entangled. To address the first challenge, we introduce Confidence-guided Bias Concept Mining (CBCM), which leverages diffusion-disentangled, semantically grounded concept representations to identify reliable spurious attributes without attribute annotations. To address the second challenge, we propose Dual-branch Cross-projection Debiasing (DCD), a prompt-tuning framework that separates target and spurious representations into two branches and explicitly removes spurious information through cross null-space projection while preserving target-relevant semantics. Extensive experiments on four benchmark datasets show that our method achieves state-of-the-art worst group accuracy among group-unsupervised approaches, while tuning at most 0.22% of the model parameters. The source code is available in the supplementary materials.", "date": "2026-06-23", "categories": []}
{"id": "2606.26200", "title": "Statistical and Structural Approaches to Algorithmic Fairness", "abstract": "Modern machine learning systems have outgrown their origins as isolated predictive constructs, evolving into complex socio-technical architectures that actively mediate human opportunity. As algorithms increasingly determine access to economic and social opportunities, it has become widely recognized that these systems are deeply embedded with the structural inequalities and prejudices of their environments. The field of algorithmic fairness emerged in response to the growing recognition that models optimized for predictive accuracy can systematically disadvantage marginalized groups. Early mitigation strategies, however, rested on fragile simplifications that limited their effectiveness in complex socio-technical environments. This thesis identifies and addresses two fundamental limitations of contemporary fairness paradigms: the reliance on deterministic point estimates for auditing and the treatment of individuals as isolated entities devoid of structural context.", "date": "2026-06-24", "categories": []}
{"id": "2606.26369", "title": "Scoring Is Not Enough: Addressing Gaps in Utility-fairness Trade-offs for Ranking", "abstract": "Scoring functions are used to represent the relevance of individual documents. In modern information retrieval or recommendation systems, they are often learned from data and play a pivotal role in ranking sets of documents or items in a way that maximizes utility to a query or user. With the recent interest in algorithmic fairness, the success of scoring has naturally led to methods that learn scores that simultaneously trade off fairness and utility. In this work, we show that in stark contrast with utility-centric objectives, scoring is sub-optimal in achieving all utility-fairness trade-offs. We establish this with a series of counter-examples with a generic fairness formulation. We show that the issue persists whether we have a deterministic scoring function or a randomized one, or whether we measure fairness at the scope of a single query or across multiple queries. On the positive side, we empirically demonstrate that semi-greedy post-processing has the potential to achieve much better trade-offs, often approaching the ideal of exhaustive post-processing in a tractable way.", "date": "2026-06-24", "categories": []}
{"id": "2606.27558", "title": "Productionized Fairness Measurement Under Privacy Constraints", "abstract": "Fairness measurements in the form of disaggregated evaluations often rely on demographic signals that are legally constrained or culturally sensitive. Race and ethnicity signals are among the more difficult signals to curate and use for this task. This paper presents Privacy-Preserving Probabilistic Race/Ethnicity Estimation (PPRE) as a method for enabling fairness measurements with respect to race/ethnicity for U.S.\\ LinkedIn members in a privacy-preserving manner. PPRE applies privacy technologies (specifically: secure two-party computation, differential privacy, and additive homomorphic encryption) on top of two race/ethnicity demographic signal sources (the Bayesian Improved Surname Geocoding estimator and a sparse golden survey set of self-reported demographics) to power a fairness measurement solution with respect to US-based race/ethnicity demographics. We detail its privacy guarantees and demonstrate its application on candidate- and viewer-side fairness measurements. We close with a transferable framework for institutions seeking to implement similar privacy-preserving measurement infrastructure.", "date": "2026-06-25", "categories": []}
{"id": "2606.28097", "title": "Fair Classification with Efficient and Post-hoc Controllable Fairness-Accuracy Trade-off", "abstract": "Post-hoc controllability of fair machine learning models, the ability to control the trade-off between fairness and accuracy after training, is valuable for practical deployment. Existing post-processing methods provide such post-hoc controllability but often suffer from significant accuracy degradation, whereas in-processing methods achieve efficient trade-offs but require computationally expensive retraining for each change in trade-off ratio. To achieve both post-hoc controllability and efficient trade-offs, we propose a novel fair classification algorithm that learns effective feature representations to improve the trade-off efficiency of post-processing fair classifiers, by a gradient-based optimization approach. Experimental results on real-world datasets demonstrate that our method achieves trade-off efficiency comparable to, or even surpassing, in-processing methods, without requiring any retraining.", "date": "2026-06-26", "categories": []}
{"id": "2606.28620", "title": "Reproducing FACTER: Fairness via Conformal Thresholding and Prompt Repair", "abstract": "Fayyazi et al. (2025) recently proposed FACTER, a model-agnostic framework designed to jointly enforce fairness and statistical coverage in LLM-based recommendation through conformal thresholding and iterative prompt repair. In this work, we conduct a reproducibility study of the FACTER framework across diverse architectures and dataset sparsity levels, evaluating both the original open-ended generation task and a constrained re-ranking extension. Under the strict reproduction, we observe a divergence in recommendation utility, which we trace to underspecified target-set evaluation in the original study. We then use the constrained re-ranking setting to evaluate FACTER when the candidate set is fixed, and introduce a static Fair Zero-Shot baseline to isolate the contribution of the iterative prompt repair loop. Our analysis shows that FACTER consistently reduces adaptive-threshold violation counts, but that these reductions are not consistently reflected under the fixed threshold or in global fairness metrics. In the constrained ranking setting, static fairness instructions achieve comparable semantic-parity outcomes to FACTER's dynamic repair loop, suggesting that the additional online repair mechanism provides limited benefit in this formulation. All code and reproduction artifacts are available at https://github.com/oscar-omlf/facter-repr.", "date": "2026-06-26", "categories": []}
{"id": "2606.28978", "title": "Can LLMs Hire Fairly? Racial Bias in Resume Screening", "abstract": "We audit fourteen mainstream large language models (LLMs) for hiring discrimination using the paired-resume methodology of Kline, Rose, and Walters (2022). The sole 2023-vintage model reproduces the pro-White callback gap documented in field experiments on labor market discrimination ($+2.12$ pp, significant at the 1\\% level). Every model released in 2024 or after shows either a null gap or a significant pro-Black reversal (up to $-3.01$ pp). The same pattern holds on the gender axis. Based on 24,024 paired postings per model across 14 models, our results document a reversal in the direction of algorithmic hiring bias across model generations.", "date": "2026-06-27", "categories": []}
{"id": "2606.30338", "title": "Sequential Fairness Auditing with Limited Output Access", "abstract": "External evaluations are becoming increasingly central to the governance of AI systems. In practice, however, independent auditors often have limited access to deployed models and must rely on query-based interactions. Most existing fairness evaluation methods assume static datasets and fixed-sample statistical tests, making them poorly suited to real-world auditing scenarios in which evidence must be collected sequentially under query constraints. In this work, we formulate fairness auditing as a tolerance-aware sequential hypothesis-testing problem under limited model output access. We develop a sequential generalized likelihood-ratio framework that allows auditors to accumulate evidence from a finite audit pool and stop once sufficient support for compliance or violation has been obtained. The framework is instantiated for decision-based Statistical Parity and Equal Opportunity audits, and extended to score- and logit-based proxy audits when richer observables are available. Our results show that both the fairness metric and the level of model access significantly affect audit efficiency, and that the benefits of richer output information are not uniform across auditing settings. In particular, richer outputs can substantially reduce the number of queries required for some fairness metrics and operating regimes, while offering limited gains in near-threshold cases. This work provides a practical statistical framework for sequential fairness auditing under realistic deployment constraints.", "date": "2026-06-29", "categories": []}
{"id": "2606.30989", "title": "Wait, am I Being Fair? Characterizing Deductive Stereotyping and Mitigating It with Fair-GCG", "abstract": "Warning: This paper contains several toxic and offensive statements. While reasoning generally improves fairness in recent large language models (LLMs), failures persist. In this work, we identify a failure mode, deductive stereotyping, in which models apply population-level statistical regularities to individual cases, producing logically coherent yet socially biased inferences. We provide a statistical interpretation of this phenomenon. To steer models toward fairness-aware reasoning, we propose a reasoning-time injection framework. We further introduce Fair-GCG to systematically discover effective injection phrases. Injection phrases discovered by Fair-GCG improve performance across multiple fairness benchmarks, generalize from smaller to larger LLMs, improves reasoning-level fairness, reduces bias in open-ended generation, and transfer to real-world fairness-sensitive tasks.", "date": "2026-06-30", "categories": []}
{"id": "2606.31644", "title": "Moral Safety in LLMs: Exposing Performative Compliance with Puzzled Cues", "abstract": "As large language models take on morally consequential roles in healthcare, legal, and hiring contexts, we need to examine whether their ethical behaviors are genuine or superficial. We show that current fairness evaluations substantially overestimate moral safety. Models appear fair when demographic identity is stated as an explicit label, yet become measurably less fair when the same identity must be inferred. We term this failure performative compliance, where a model is fair when the presentation resembles a fairness evaluation and less fair as that cue weakens. We introduce a cue-variation methodology that holds the moral dilemma and the demographic identity fixed and varies only how that identity is conveyed. Hiding the explicit label raises harmful decisions by $+4.4$~pp, changes model safety rankings, and the shift persists when models correctly infer the demographic, ruling out attribution error. We propose the Cue Visibility Gap, a model-agnostic robustness metric that can be added to any existing fairness benchmark to separate genuine from performative moral safety. Fairness evaluations that omit cue variation measure surface compliance, not moral robustness, and should not ground deployment decisions in high-stakes settings.", "date": "2026-06-30", "categories": []}
{"id": "2606.31704", "title": "WIDER-FAIR: An Annotated Version of the WIDER-FACE Dataset for Fairness Evaluation", "abstract": "The deployment of face detection models in real-world applications raises important fairness concerns, as these systems may showcase performance disparities across demographic groups. A key obstacle to studying and mitigating such biases is the lack of face detection datasets with sensitive feature annotations. To address this gap, we introduce WIDER-FAIR, a new dataset built on the widely used WIDER-FACE benchmark, manually annotated with the perceived ethnicity and sex of each face. The dataset contains 16,256 images annotated across four ethnic groups: Asian, Black, Indian, and White, and two sex categories. We assess the quality and coherence of the annotations using face embeddings, a K-Nearest Neighbors classifier, and a t-SNE visualization, all of which support the consistency of the labeling process. As a demonstration of the dataset's potential, we train a YOLOv5 model and perform ablation studies on each sensitive feature. Among other findings, our experiments show that detection performance is notably lower for faces of Black individuals, and that excluding this group from training increases fairness disparity more than excluding any other ethnic group. These observations illustrate the value of demographically annotated datasets for understanding and evaluating bias in face detection models.", "date": "2026-06-30", "categories": []}