Datasets:
| {"id": "2606.19594", "title": "Unsupervised Causal Abstractions Discovery", "abstract": "Causal abstractions formalize when a high-level structural causal model (SCM) captures the interventional behavior of a lower-level SCM. Existing applications of this notion largely follow a hypothesis-testing paradigm: an expert proposes a candidate high-level model and then evaluates if the low-level system implements it. We study the complementary problem of learning a high-level model directly from low-level measurements. Our contributions leverage hypotheses from low-rank causal discovery, and can be summarized as follows: (1) we show that observations generated by a low-rank graph induce latents that form a causal abstraction, (2) we provide identifiability results about these latents, and (3) we propose a practical objective to learn this high-level SCM.", "date": "2026-06-17", "categories": []} | |
| {"id": "2606.18509", "title": "Concept Modulation Models: A Unified Framework for Identifiability and Extrapolation", "abstract": "Reliable generalization in conditional latent variable models requires understanding both identifiability and extrapolation: how observed variation across attributes determines latent structure, and how that structure determines distributions at unseen attributes. However, existing identifiability and extrapolation guarantees are largely model-specific, with separate analyses in nonlinear ICA, causal representation learning, perturbation modeling, and related conditional latent variable models. We introduce concept modulation models (CMMs), an attribute-indexed class of conditional generative models with structure $A\\to \\Lambda \\to C\\to X$, where attributes select modulators, modulators induce latent concept laws, and concepts generate observed features. CMMs lift transition-based identifiability to conditional settings by showing that feature agreement on observed attributes induces a latent concept transition constrained by the CMM class. We express these constraints through attribute potentials, log-density ratios between attribute-conditioned concept laws, separating the generic lifting step from model-specific rigidity arguments. The same potentials control extrapolation: agreement at unseen attributes holds exactly when the transported attribute-potential identities extend to those attributes. This yields algebraic extrapolation criteria, identifies the common potential-based proof objects behind several existing identifiability and extrapolation results, and, when combined with the model-specific rigidity arguments in those works, recovers their stated conclusions.", "date": "2026-06-16", "categories": []} | |
| {"id": "2508.04492", "title": "Learning Robust Intervention Representations with Delta Embeddings", "abstract": "Causal representation learning has attracted significant research interest during the past few years, as a means for improving model generalization and robustness. Causal representations of interventional image pairs (also called ``actionable counterfactuals'' in the literature), have the property that only variables corresponding to scene elements affected by the intervention / action are changed between the start state and the end state. While most work in this area has focused on identifying and representing the variables of the scene under a causal model, fewer efforts have focused on representations of the interventions themselves. In this work, we show that an effective strategy for improving out of distribution (OOD) robustness is to focus on the representation of actionable counterfactuals in the latent space. Specifically, we propose that an intervention can be represented by a Causal Delta Embedding that is invariant to the visual scene and sparse in terms of the causal variables it affects. Leveraging this insight, we propose a method for learning causal representations from image pairs, without any additional supervision. Experiments in the Causal Triplet challenge demonstrate that Causal Delta Embeddings are highly effective in OOD settings, significantly exceeding baseline performance in both synthetic and real-world benchmarks.", "date": "2025-08-06", "categories": []} | |
| {"id": "2606.14892", "title": "Relational Structural Causal Models", "abstract": "An artificial intelligence must have a model of its environment that is causal, supporting reasoning about interventions and counterfactuals, and also combinatorial, supporting generalization to unseen combinations of objects. In this work, we formally study when and how such a model can be learned. We develop relational structural causal models, extending structural causal models (Pearl 2009) to settings where objects and their relations vary. First, we show how answers to not only causal but also observational queries about unseen combinations of objects can not be identified without further assumptions. To enable such identification--including in the presence of unobserved confounding--we define relational causal graphs and derive symbolic identification criteria. Finally, we propose relational neural causal models, a provably correct approach that outperforms non-relational baselines on simulated traffic scenes with varying cars, signals, and pedestrians.", "date": "2026-06-12", "categories": []} | |
| {"id": "2603.14483", "title": "Disentangling Dynamical Systems: Causal Representation Learning Meets Local Sparse Attention", "abstract": "Parametric system identification methods estimate the parameters of explicitly defined physical systems from data. Yet, they remain constrained by the need to provide an explicit function space, typically through a predefined library of candidate functions chosen via available domain knowledge. In contrast, deep learning can demonstrably model systems of broad complexity with high fidelity, but black-box function approximation typically fails to yield explicit descriptive or disentangled representations revealing the structure of a system. We develop a novel identifiability theorem, leveraging causal representation learning, to uncover disentangled representations of system parameters without structural assumptions. We derive a graphical criterion specifying when system parameters can be uniquely disentangled from raw trajectory data, up to permutation and diffeomorphism. Crucially, our analysis demonstrates that global causal structures provide a lower bound on the disentanglement guarantees achievable when considering local state-dependent causal structures. We instantiate system parameter identification as a variational inference problem, leveraging a sparsity-regularised transformer to uncover state-dependent causal structures. We empirically validate our approach across four synthetic domains, demonstrating its ability to recover highly disentangled representations that baselines fail to recover. Corroborating our theoretical analysis, our results confirm that enforcing local causal structure is often necessary for full identifiability.", "date": "2026-03-15", "categories": []} | |
| {"id": "2405.17151", "title": "Smoke and Mirrors in Causal Downstream Tasks", "abstract": "Machine Learning and AI have the potential to transform data-driven scientific discovery, enabling accurate predictions for several scientific phenomena. As many scientific questions are inherently causal, this paper looks at the causal inference task of treatment effect estimation, where the outcome of interest is recorded in high-dimensional observations in a Randomized Controlled Trial (RCT). Despite being the simplest possible causal setting and a perfect fit for deep learning, we theoretically find that many common choices in the literature may lead to biased estimates. To test the practical impact of these considerations, we recorded ISTAnt, the first real-world benchmark for causal inference downstream tasks on high-dimensional observations as an RCT studying how garden ants (Lasius neglectus) respond to microparticles applied onto their colony members by hygienic grooming. Comparing 6 480 models fine-tuned from state-of-the-art visual backbones, we find that the sampling and modeling choices significantly affect the accuracy of the causal estimate, and that classification accuracy is not a proxy thereof. We further validated the analysis, repeating it on a synthetically generated visual data set controlling the causal model. Our results suggest that future benchmarks should carefully consider real downstream scientific questions, especially causal ones. Further, we highlight guidelines for representation learning methods to help answer causal questions in the sciences.", "date": "2024-05-27", "categories": []} | |
| {"id": "2509.01916", "title": "Causal Representation Learning from Network Data", "abstract": "Causal disentanglement from soft interventions is identifiable under the assumptions of linear interventional faithfulness and availability of both observational and interventional data. Prior work has focused on unstructured observations without leveraging known relational context among measured entities. In many scientific applications, however, the measured variables come with an observed interaction network that provides structured context, such as protein-protein interactions and pathway-gene membership. We propose GraCE-VAE, a graph-aware causal discrepancy variational autoencoder that treats pathway-level information as an auxiliary view of the latent causal programs. The graph neural network encoder conditions on this auxiliary pathway view and the biological graph to improve amortized inference, while the causal decoder remains a latent SCM with soft interventions. Assuming samples are i.i.d. within each intervention regime, we show that GraCE-VAE inherits the identifiability guarantees of causal discrepancy VAEs and identifies the latent causal graph and intervention targets up to the standard equivalence class. Experiments on three CRISPR perturbation datasets demonstrate that leveraging structured biological context improves prediction of interventional outcomes, including unseen perturbation combinations.", "date": "2025-09-02", "categories": []} | |
| {"id": "2606.06288", "title": "Discrete Causal Representations from Heterogeneous Domains: A Bayesian Approach with Social Survey Applications", "abstract": "Causal representation learning aims to infer the high-level latent causal concepts that give rise to observed low-level measurements. This is particularly relevant for heterogeneous data from different environments or domains since distribution shifts often arise through sparse, localized changes in some of the underlying causal mechanisms, while other parts of the generative process remain unchanged. Whereas identifiability of causal representations has been studied extensively, practical uncertainty-aware methods and real-world use cases remain less explored. In this work, we propose a Bayesian approach to learning causal representations from multi-environment data, focusing on the case of discrete causal concepts and unknown multi-node soft interventions. To this end, we translate causal assumptions and interpretability desiderata into suitable priors and parametric choices within a hierarchical model. We then devise an inference scheme based on sequential Monte Carlo sampling to approximate the resulting multimodal posterior. We showcase our approach through case studies on social survey data, where latent causal concepts correspond to cultural values or political opinions, measurements to survey responses, and environments to different countries or states. Our model infers meaningful high-level concepts and plausible causal relations among them, demonstrating its utility for learning causal representations of complex real-world data.", "date": "2026-06-04", "categories": []} | |
| {"id": "2501.12500", "title": "Learning General Causal Structures with Hidden Dynamic Process for Climate Analysis", "abstract": "Understanding climate dynamics requires going beyond correlations in observational data to uncover the underlying causal process. Latent drivers such as atmospheric processes play a central role in temporal dynamics, while direct causal influences also exist among geographically proximate observed variables. Traditional Causal Representation Learning (CRL) typically focuses on latent factors but overlooks such observable-to-observable causal relations, which limits its applicability to climate analysis. In this paper, we introduce a unified framework that jointly uncovers (i) causal relations among observed variables and (ii) latent driving forces together with their interactions. We establish conditions under which both the hidden dynamic process and the causal structure among observed variables are simultaneously identifiable from time-series data, and our guarantees continue to hold in the nonparametric setting through contextual information that recovers latent variables and causal relations. Building on these insights, we propose CaDRe (Causal Discovery and Representation learning), a time-series generative model with structural constraints that integrates CRL and causal discovery. Experiments on synthetic datasets validate our theoretical results. On real-world climate datasets, CaDRe delivers competitive forecasting accuracy and recovers visualized causal graphs aligned with domain expertise, thereby offering interpretable insights into climate systems. Code is available at https://github.com/MinghaoFu/CaDRe.", "date": "2025-01-21", "categories": []} | |
| {"id": "2509.22553", "title": "Linear Causal Representation Learning by Topological Ordering, Pruning, and Disentanglement", "abstract": "Causal representation learning (CRL) has garnered increasing interest from the causal inference and artificial intelligence communities due to its potential to disentangle complex data-generating mechanism into causally interpretable latent features by leveraging the heterogeneity of modern datasets. In this paper, we further contribute to the CRL literature, by focusing on the stylized linear structural causal model over latent features and assuming a linear mixing function that maps latent features to the observed data or measurements. Existing linear CRL methods often rely on stringent assumptions, such as access to single-node interventional data or restrictive distributional constraints on latent features and/or exogenous measurement noise. However, these prerequisites can be easy to violate in practice. In this work, we propose a novel linear CRL algorithm that, unlike existing methods, operates under weaker assumptions on environment heterogeneity and data-generating distributions while still recovering latent causal features up to an equivalence class. We further validate our new algorithm via synthetic experiments and an interpretability analysis of large language models, demonstrating both its superiority over competing methods in finite samples and its potential in integrating causality into understanding artificial intelligence. The source code is available at https://github.com/utulie/code_for_linear_crl_paper_creator.", "date": "2025-09-26", "categories": []} | |
| {"id": "2605.27043", "title": "Causal Representation Learning for Generalisable Recommendation", "abstract": "Predictive models trained on observational data often fail to generalise to the distributions they encounter when deployed, especially when the training data is a product of the system being optimised. Recommender systems are a canonical example: they are trained on interaction logs confounded by the deployed policy, past user behaviour, and platform filtering. As a result, the training distribution differs substantially from the candidate distribution scored at serving time, a gap that makes offline metrics unreliable predictors of online performance. We address the distribution shift problem with a method motivated by causal representation learning (CRL). We propose an information-theoretic disentanglement criterion and prove that its optimum depends only on the causal components of the input. We then derive a tractable variational lower bound that makes the criterion optimisable from finite observational data alone. The scope of our method is narrower than that of much of the CRL literature, in that we target better generalisation under distribution shift, not full identification of all latent causal factors. This narrower target is what makes the method practical, requiring only the existing confounded logs, applying to any standard supervised model, and adding no inference-time cost. Our headline evaluation is an A/B test with millions of users on Spotify, applied to a production ranker for personalised playlist generation. A capacity-matched CRL variant performed on par offline but delivered substantial online gains in listener engagement. Complementary evidence on the public KuaiRand recommendation dataset and a synthetic benchmark with known causal structure shows the same pattern: offline parity with baseline, gains under distribution shift. Across all three settings, adding our causal disentanglement objective yields meaningfully better out-of-distribution generalisation.", "date": "2026-05-26", "categories": []} | |
| {"id": "2605.20396", "title": "Score-Based Causal Discovery of Latent Variable Causal Models", "abstract": "Identifying latent variables and the causal structure involving them is essential across various scientific fields. While many existing works fall under the category of constraint-based methods (with e.g. conditional independence or rank deficiency tests), they may face empirical challenges such as testing-order dependency, error propagation, and choosing an appropriate significance level. These issues can potentially be mitigated by properly designed score-based methods, such as Greedy Equivalence Search (GES) (Chickering, 2002) in the specific setting without latent variables. Yet, formulating score-based methods with latent variables is highly challenging. In this work, we develop score-based methods that are capable of identifying causal structures containing causally-related latent variables with identifiability guarantees. Specifically, we show that a properly formulated scoring function can achieve score equivalence and consistency for structure learning of latent variable causal models. We further provide a characterization of the degrees of freedom for the marginal over the observed variables under multiple structural assumptions considered in the literature, and accordingly develop both exact and continuous score-based methods. This offers a unified view of several existing constraint-based methods with different structural assumptions. Experimental results validate the effectiveness of the proposed methods.", "date": "2026-05-19", "categories": []} | |
| {"id": "2605.21058", "title": "A Dialogue between Causal and Traditional Representation Learning: Toward Mutual Benefits in a Unified Formulation", "abstract": "Causal representation learning (CRL) and traditional representation learning have largely developed along different trajectories. Traditional representation learning has been driven mainly by applications and empirical objectives, whereas CRL has focused more on theoretical questions, particularly identifiability. This difference in emphasis has created a gap between the two fields in terminology, problem formulation, and evaluation, limiting communication and sometimes leading to disconnected or redundant efforts. In this paper, we argue that these two fields should be brought into dialogue rather than treated as separate paradigms. To this end, we introduce a unified formulation in which the representation learning is characterized by two components: a task component, which specifies what information the learned representation is required to preserve, and a constraint component, which specifies what structure is imposed on the latent space. Under this formulation, the benefits run in both directions. CRL provides theoretical tools for understanding when structured latent constraints are useful or necessary, while traditional representation learning offers practical insights on task design and objective choice that can improve the development of CRL methods. To illustrate this interaction, we experimentally study how different task components affect the behavior of CRL methods under different structured constraints. Results on CausalVerse show that the effectiveness of causal constraints depends strongly on the tasks with which they are paired.", "date": "2026-05-20", "categories": []} | |
| {"id": "2605.19135", "title": "Identifiable Multimodal Causal Representation Learning under Partial Latent Sharing", "abstract": "Causal representation learning (CRL) seeks to uncover meaningful latent variables and their corresponding causal structure from high-dimensional observational data. Although its significance, CRL identifiability remains a crucial property, as it ensures the recovery of the mechanisms behind the data generation process, and hence the interpretability and robustness of the representation. Proving identifiability in CRL is intrinsically difficult, and we address in this work an even more challenging setting: multimodality. We consider multimodal observed data with a latent partially shared structure. Each modality is generated, through non linear mixing functions, from a specific subset of causal latent variables. Under flexible assumptions and without imposing any parametric distribution on the latent variables, we establish component-wise identifiability guarantees for the causal latent representation. Our identifiability results, furthermore, apply to the undercomplete scenario where we have, for each modality, more observed than latent variables. To instantiate our theoretical analysis, we introduce a Wasserstein-based module to recover the partially shared latent structure. Due to its differentiability, the latter can be easily integrated into all types of architecture, only requiring minimal changes. Extensive experiments on synthetic and realistic datasets validate the superiority of our approach over SOTA methods.", "date": "2026-05-18", "categories": []} | |
| {"id": "2601.21350", "title": "Factored Causal Representation Learning for Robust Reward Modeling in RLHF", "abstract": "A reliable reward model is essential for aligning large language models with human preferences through reinforcement learning from human feedback. However, standard reward models are susceptible to spurious features that are not causally related to human labels. This can lead to reward hacking, where high predicted reward does not translate into better behavior. In this work, we address this problem from a causal perspective by proposing a factored representation learning framework that decomposes the model's contextual embedding into (1) causal factors that are sufficient for reward prediction and (2) non-causal factors that capture reward-irrelevant attributes such as length or sycophantic bias. The reward head is then constrained to depend only on the causal component. In addition, we introduce an adversarial head trained to predict reward from the non-causal factors, while applying gradient reversal to discourage them from encoding reward-relevant information. Experiments on both mathematical and dialogue tasks demonstrate that our method learns more robust reward models and consistently improves downstream RLHF performance over state-of-the-art baselines. Analyses on length and sycophantic bias further validate the effectiveness of our method in mitigating reward hacking behaviors.", "date": "2026-01-29", "categories": []} | |
| {"id": "2605.17827", "title": "Content-Style Identification via Differential Independence", "abstract": "Generative analysis often models multi-domain observations as nonlinear mixtures of domain-invariant content variables and domain-specific style variables. Identifying both factors from unpaired domains enables tasks such as domain transfer and counterfactual data generation. Prior work establishes identifiability under (block-wise) statistical independence between content and style, or via sparse Jacobian assumptions on the nonlinear mixing function, but such conditions can be restrictive in practice. In this work, we introduce content-style differential independence (CSDI), an alternative structural condition requiring that infinitesimal variations in content and style induce orthogonal directions on the data manifold, thereby enabling identifiability even when content and style are dependent and the Jacobian is dense. We operationalize this condition through a blockwise orthogonality constraint on the Jacobian subspaces associated with content and style. To support high-dimensional generative models, we design a stochastic regularizer based on numerical Jacobian approximation, enabling scalable training in settings such as high-resolution image generation. Experiments across multiple datasets corroborate the identifiability analysis and demonstrate practical benefits on counterfactual generation and domain translation.", "date": "2026-05-18", "categories": []} | |
| {"id": "2510.13583", "title": "On the Identifiability of Causal Graphs with the Invariance Principle", "abstract": "Causal discovery from i.i.d. observational data is known to be generally ill-posed. We demonstrate that if we have access to the distribution {induced} by a structural causal model, and additional data from (in the best case) \\textit{only two} environments that sufficiently differ in the noise statistics, the unique causal graph is identifiable. Notably, this is the first result in the literature that guarantees the entire causal graph recovery with a constant number of environments and arbitrary nonlinear mechanisms. Our only constraint is the Gaussianity of the noise terms; however, we propose potential ways to relax this requirement. Of interest on its own, we expand on the well-known duality between independent component analysis (ICA) and causal discovery; recent advancements have shown that nonlinear ICA can be solved from multiple environments, at least as many as the number of sources: we show that the same can be achieved for causal discovery while having access to much less auxiliary information.", "date": "2025-10-15", "categories": []} | |
| {"id": "2410.14375", "title": "Causal Fine-Tuning under Latent Confounded Shift", "abstract": "Adapting to latent confounded shift remains a core challenge in modern AI. This setting is driven by hidden variables that induce spurious correlations between inputs and outputs during training, leading models to rely on non-causal shortcuts. For example, a model may learn to treat metadata (e.g., data source like \"Amazon\") as a proxy for positive sentiment, causing failure when the source becomes predominantly negative during deployment. To address this latent confounded shift, we introduce Causal Fine-Tuning(CFT). Using a structural causal model as an inductive bias, we derive sufficient identification conditions that motivate a fine-tuning objective for decomposing representations into high-level stable and low-level shift-sensitive components. Instantiating this framework in BERT, we show that learning such causal/spurious representations and adjusting them accordingly yield a more robust predictor. Experiments on spurious correlation injection attacks in text demonstrate that our method outperforms black-box domain generalization baselines, highlighting the benefits of explicitly modeling causal structure.", "date": "2024-10-18", "categories": []} | |
| {"id": "2509.22196", "title": "Mechanistic Independence: A Principle for Identifiable Disentangled Representations", "abstract": "Disentangled representations seek to recover latent factors of variation underlying observed data, yet their identifiability is still not fully understood. We introduce a unified framework in which disentanglement is achieved through mechanistic independence, which characterizes latent factors by how they act on observed variables rather than by their latent distribution. This perspective is invariant to changes of the latent density, even when such changes induce statistical dependencies among factors. Within this framework, we propose several related independence criteria -- ranging from support-based and sparsity-based to higher-order conditions -- and show that each yields identifiability of latent subspaces, even under nonlinear, non-invertible mixing. We further establish a hierarchy among these criteria and provide a graph-theoretic characterization of latent subspaces as connected components. Together, these results clarify the conditions under which disentangled representations can be identified without relying on statistical assumptions.", "date": "2025-09-26", "categories": []} | |
| {"id": "2605.05524", "title": "MOSAIC: Module Discovery via Sparse Additive Identifiable Causal Learning for Scientific Time Series", "abstract": "Causal representation learning (CRL) seeks to recover latent variables with identifiability guarantees, typically up to permutation and component-wise reparameterization under appropriate assumptions. However, identifiability does not imply interpretability: latent semantics are typically assigned post hoc by alignment with known ground-truth factors. This limitation is particularly acute in scientific time series, where underlying mechanisms are unknown and discovering interpretable structure is a primary goal. In contrast, scientific observations (such as residue-pair distances, climate indices, or process sensors) are inherently semantic, as they correspond to named physical quantities. This raises a key question: can the interpretability of observations be transferred to the identifiable latent space? We propose MOSAIC (Module discovery via Sparse Additive Identifiable Causal learning), a sparse temporal VAE that integrates temporal CRL identifiability with support recovery over observed variables. MOSAIC identifies latent variables via regime-conditioned temporal variation, and recovers for each latent a sparse set of associated observations through an additive decoder, yielding module-level interpretability. We show that ANOVA main-effect supports are identifiable under general smooth mixing functions, and provide finite-sample recovery guarantees for a tractable sparse-additive variant. Empirically, MOSAIC recovers domain-consistent variable groups across RNA molecular dynamics, solar wind, ENSO climate, the Tennessee Eastman process, and a synthetic tokamak benchmark, enabling interpretable discovery of latent mechanisms in scientific time series.", "date": "2026-05-06", "categories": []} | |
| {"id": "2605.05706", "title": "Resolving the bias-precision paradox with stochastic causal representation learning for personalized medicine", "abstract": "Estimating individualized treatment effects from longitudinal observational data is central to data-driven medicine, yet existing methods face a fundamental limitation: reducing confounding bias often suppresses clinically informative heterogeneity, degrading patient-specific predictions. Here, we identify this tension as a bias-precision paradox in causal representation learning and introduce sampling-based maximum mean discrepancy (sMMD), a stochastic alignment strategy that replaces global adversarial balancing with subset-level matching. We instantiate this approach in a framework for counterfactual outcome prediction with attribution-grounded interpretability. Across two large-scale ICU cohorts (n = 27,783), our framework improves accuracy under distribution shift, reducing error by up to 11.5% and substantially increasing recall in high-risk tasks. Mechanistic analyses show that sMMD selectively preserves clinically decisive variables. In human-AI evaluation, our method outperforms clinicians-in-training and large language models, and improves clinician accuracy by 14.7% while reducing decision time, enabling interpretable, real-time clinical decision support.", "date": "2026-05-07", "categories": []} | |
| {"id": "2605.05029", "title": "The Predictive-Causal Gap: An Impossibility Theorem and Large-Scale Neural Evidence", "abstract": "We report a systematic failure mode in predictive representation learning. Across 2695 neural network configurations trained to predict linear-Gaussian dynamics, the optimal encoder tracks the environment rather than the system it is meant to model. The mean causal fidelity -- the fraction of encoder sensitivity allocated to system degrees of freedom -- is 0.49, and only 2.5% of configurations exceed 0.70. The failure intensifies with dimension: at N=100, the optimal encoder becomes causally blind (fidelity ~10^{-8}) while achieving 92% lower prediction error than the causal representation. We prove this is not an optimization artifact but a structural property of the predictive objective: when environment modes are slower or less noisy than system modes, every minimizer of the population risk encodes the former. The set of dynamics exhibiting this predictive-causal gap is open and of positive measure in parameter space. In a nonlinear Duffing-GRU sweep, unconstrained predictors learn environment-dominant representations in 55% of tasks (95% CI 41--68%) versus 24% under operational grounding (p=2.3e-3); the median out-of-distribution MSE inflation under environment shift is 1.82x versus 1.00x. Operational grounding -- restricting the loss to system observables -- partially suppresses the gap, but causal fidelity is never recovered without an explicit system-environment boundary. The results identify the predictive-causal gap as a structural limit of learning, with implications for self-supervised representation learning, world models, and the scaling paradigm.", "date": "2026-05-06", "categories": []} | |
| {"id": "2605.03268", "title": "Partially Observed Structural Causal Models", "abstract": "Here we introduce Partially Observed Structural Causal Models (POSCMs) that formalize causal systems where latent contexts co-determine both the interaction structure and downstream mechanisms on observed variables. POSCMs provide an extension of structural causal models (SCMs), as a self-contained causal modeling framework for endogenous graphs, allowing for an intervention hierarchy spanning node- and edge-level context and endogenous variable interventions. To enable surgical edge interventions, we adopt a Kolmogorov-Arnold-Sprecher edge-functional decomposition, an existence theorem for representing each node mechanism as a sum of univariate functions of its parents, yielding an explicit parametrization of dyadic functional contributions. We provide an identifiability theory that clarifies which intervention families would suffice to disentangle structure formation from mechanisms. We empirically validate these predictions in a biophysically detailed virtual human retina simulator, constructing intervention protocols that (i) reproduce the non-identifiability predicted when context is latent and no context-level interventions are available, (ii) exhibit structure-mechanism confounding under latent edges when only node interventions are observed, and (iii) recover synaptic input-output relationships via targeted node interventions, consistent with our positive kernel identifiability result. Our work generalizes SCMs in a way that allows it to work in a world closer to the one we live in.", "date": "2026-05-05", "categories": []} | |
| {"id": "2603.11907", "title": "Causal Representation Learning with Optimal Compression under Complex Treatments", "abstract": "Estimating Individual Treatment Effects (ITE) in multi-treatment scenarios faces two critical challenges: the Hyperparameter Selection Dilemma for balancing weights and the Curse of Dimensionality in computational scalability. This paper derives a novel multi-treatment generalization bound and proposes a theoretical estimator for the optimal balancing weight $\\alpha$, eliminating expensive heuristic tuning. We investigate three balancing strategies: Pairwise, One-vs-All (OVA), and Treatment Aggregation. While OVA achieves superior precision in low-dimensional settings, our proposed Treatment Aggregation ensures both accuracy and O(1) scalability as the treatment space expands. Furthermore, we extend our framework to a generative architecture, Multi-Treatment CausalEGM, which preserves the Wasserstein geodesic structure of the treatment manifold. Experiments on semi-synthetic and image datasets demonstrate that our approach significantly outperforms traditional models in estimation accuracy and efficiency, particularly in large-scale intervention scenarios.", "date": "2026-03-12", "categories": []} | |
| {"id": "2604.26919", "title": "Causal Learning with Neural Assemblies", "abstract": "Can Neural Assemblies -- groups of neurons that fire together and strengthen through co-activation -- learn the direction of causal influence between variables? While established as a computationally general substrate for classification, parsing, and planning, neural assemblies have not yet been shown to internalize causal directionality. We demonstrate that the inherent operations of neural assemblies -- projection, local plasticity control, and sparse winner selection -- are sufficient for directional learning. We introduce DIRECT (DIRectional Edge Coupling/Training), a mechanism that co-activates source and target assemblies under an adaptive gain schedule to internalize directed relations. Unlike backpropagation-based methods, DIRECT relies solely on local plasticity, making the resulting causal claims auditable at the mechanism level. Our findings are verified through a dual-readout validation strategy: (i) synaptic-strength asymmetry, measuring the emergent weight gap between forward and reverse links, and (ii) functional propagation overlap, quantifying the reliability of directional signal flow. Across multiple domains, the framework achieves perfect structural recovery under a supervised, known-structure setting. These results establish neural assemblies as an auditable bridge between biologically plausible dynamics and formal causal models, offering an \"explainable by design\" framework where causal claims are traceable to specific neural winners and synaptic asymmetries.", "date": "2026-04-29", "categories": []} | |
| {"id": "2604.23800", "title": "Causal Representation Learning from General Environments under Nonparametric Mixing", "abstract": "Causal representation learning aims to recover the latent causal variables and their causal relations, typically represented by directed acyclic graphs (DAGs), from low-level observations such as image pixels. A prevailing line of research exploits multiple environments, which assume how data distributions change, including single-node interventions, coupled interventions, or hard interventions, or parametric constraints on the mixing function or the latent causal model, such as linearity. Despite the novelty and elegance of the results, they are often violated in real problems. Accordingly, we formalize a set of desiderata for causal representation learning that applies to a broader class of environments, referred to as general environments. Interestingly, we show that one can fully recover the latent DAG and identify the latent variables up to minor indeterminacies under a nonparametric mixing function and nonlinear latent causal models, such as additive (Gaussian) noise models or heteroscedastic noise models, by properly leveraging sufficient change conditions on the causal mechanisms up to third-order derivatives. These represent, to our knowledge, the first results to fully recover the latent DAG from general environments under nonparametric mixing. Notably, our results match or improve upon many existing works, but require less restrictive assumptions about changing environments.", "date": "2026-04-26", "categories": []} | |
| {"id": "2604.17568", "title": "Diverse Dictionary Learning", "abstract": "Given only observational data $X = g(Z)$, where both the latent variables $Z$ and the generating process $g$ are unknown, recovering $Z$ is ill-posed without additional assumptions. Existing methods often assume linearity or rely on auxiliary supervision and functional constraints. However, such assumptions are rarely verifiable in practice, and most theoretical guarantees break down under even mild violations, leaving uncertainty about how to reliably understand the hidden world. To make identifiability actionable in the real-world scenarios, we take a complementary view: in the general settings where full identifiability is unattainable, what can still be recovered with guarantees, and what biases could be universally adopted? We introduce the problem of diverse dictionary learning to formalize this view. Specifically, we show that intersections, complements, and symmetric differences of latent variables linked to arbitrary observations, along with the latent-to-observed dependency structure, are still identifiable up to appropriate indeterminacies even without strong assumptions. These set-theoretic results can be composed using set algebra to construct structured and essential views of the hidden world, such as genus-differentia definitions. When sufficient structural diversity is present, they further imply full identifiability of all latent variables. Notably, all identifiability benefits follow from a simple inductive bias during estimation that can be readily integrated into most models. We validate the theory and demonstrate the benefits of the bias on both synthetic and real-world data.", "date": "2026-04-19", "categories": []} | |
| {"id": "2604.13218", "title": "Identifiability of Potentially Degenerate Gaussian Mixture Models With Piecewise Affine Mixing", "abstract": "Causal representation learning (CRL) aims to identify the underlying latent variables from high-dimensional observations, even when variables are dependent with each other. We study this problem for latent variables that follow a potentially degenerate Gaussian mixture distribution and that are only observed through the transformation via a piecewise affine mixing function. We provide a series of progressively stronger identifiability results for this challenging setting in which the probability density functions are ill-defined because of the potential degeneracy. For identifiability up to permutation and scaling, we leverage a sparsity regularization on the learned representation. Based on our theoretical results, we propose a two-stage method to estimate the latent variables by enforcing sparsity and Gaussianity in the learned representations. Experiments on synthetic and image data highlight our method's effectiveness in recovering the ground-truth latent variables.", "date": "2026-04-14", "categories": []} | |
| {"id": "2507.04754", "title": "Intervening to Learn and Compose Causally Disentangled Representations", "abstract": "In designing generative models, it is commonly believed that in order to learn useful latent structure, we face a fundamental tension between expressivity and structure. In this paper we challenge this view by proposing a new approach to training arbitrarily expressive generative models that simultaneously learn causally disentangled concepts. This is accomplished by adding a simple context module to an arbitrarily complex black-box model, which learns to process concept information by implicitly inverting linear representations from the model's encoder. Inspired by the notion of intervention in a causal model, our module selectively modifies its architecture during training, allowing it to learn a compact joint model over different contexts. We show how adding this module leads to causally disentangled representations that can be composed for out-of-distribution generation on both real and simulated data. The resulting models can be trained end-to-end or fine-tuned from pre-trained models. To further validate our proposed approach, we prove a new identifiability result that extends existing work on identifying structured representations.", "date": "2025-07-07", "categories": []} | |
| {"id": "2603.25796", "title": "Beyond identifiability: Learning causal representations with few environments and finite samples", "abstract": "We provide explicit, finite-sample guarantees for learning causal representations from data with a sublinear number of environments. Causal representation learning seeks to provide a rigourous foundation for the general representation learning problem by bridging causal models with latent factor models in order to learn interpretable representations with causal semantics. Despite a blossoming theory of identifiability in causal representation learning, estimation and finite-sample bounds are less well understood. We show that causal representations can be learned with only a logarithmic number of unknown, multi-node interventions, and that the intervention targets need not be carefully designed in advance. Through a careful perturbation analysis, we provide a new analysis of this problem that guarantees consistent recovery of (a) the latent causal graph, (b) the mixing matrix and representations, and (c) \\emph{unknown} intervention targets.", "date": "2026-03-26", "categories": []} | |
| {"id": "2603.18404", "title": "Multi-Domain Empirical Bayes for Linearly-Mixed Causal Representations", "abstract": "Causal representation learning (CRL) aims to learn low-dimensional causal latent variables from high-dimensional observations. While identifiability has been extensively studied for CRL, estimation has been less explored. In this paper, we explore the use of empirical Bayes (EB) to estimate causal representations. In particular, we consider the problem of learning from data from multiple domains, where differences between domains are modeled by interventions in a shared underlying causal model. Multi-domain CRL naturally poses a simultaneous inference problem that EB is designed to tackle. Here, we propose an EB $f$-modeling algorithm that improves the quality of learned causal variables by exploiting invariant structure within and across domains. Specifically, we consider a linear measurement model and interventional priors arising from a shared acyclic SCM. When the graph and intervention targets are known, we develop an EM-style algorithm based on causally structured score matching. We further discuss EB $g$-modeling in the context of existing CRL approaches. In experiments on synthetic data, our proposed method achieves more accurate estimation than other methods for CRL.", "date": "2026-03-19", "categories": []} | |
| {"id": "2602.19414", "title": "Federated Causal Representation Learning in State-Space Systems for Decentralized Counterfactual Reasoning", "abstract": "Networks of interdependent industrial assets (clients) are tightly coupled through physical processes and control inputs, raising a key question: how would the output of one client change if another client were operated differently? This is difficult to answer because client-specific data are high-dimensional and private, making centralization of raw data infeasible. Each client also maintains proprietary local models that cannot be modified. We propose a federated framework for causal representation learning in state-space systems that captures interdependencies among clients under these constraints. Each client maps high-dimensional observations into low-dimensional latent states that disentangle intrinsic dynamics from control-driven influences. A central server estimates the global state-transition and control structure. This enables decentralized counterfactual reasoning where clients predict how outputs would change under alternative control inputs at others while only exchanging compact latent states. We prove convergence to a centralized oracle and provide privacy guarantees. Our experiments demonstrate scalability, and accurate cross-client counterfactual inference on synthetic and real-world industrial control system datasets.", "date": "2026-02-23", "categories": []} | |
| {"id": "2603.17405", "title": "Causal Representation Learning on High-Dimensional Data: Benchmarks, Reproducibility, and Evaluation Metrics", "abstract": "Causal representation learning (CRL) models aim to transform high-dimensional data into a latent space, enabling interventions to generate counterfactual samples or modify existing data based on the causal relationships among latent variables. To facilitate the development and evaluation of these models, a variety of synthetic and real-world datasets have been proposed, each with distinct advantages and limitations. For practical applications, CRL models must perform robustly across multiple evaluation directions, including reconstruction, disentanglement, causal discovery, and counterfactual reasoning, using appropriate metrics for each direction. However, this multi-directional evaluation can complicate model comparison, as a model may excel in some direction while under-performing in others. Another significant challenge in this field is reproducibility: the source code corresponding to published results must be publicly available, and repeated runs should yield performance consistent with the original reports. In this study, we critically analyzed the synthetic and real-world datasets currently employed in the literature, highlighting their limitations and proposing a set of essential characteristics for suitable datasets in CRL model development. We also introduce a single aggregate metric that consolidates performance across all evaluation directions, providing a comprehensive score for each model. Finally, we reviewed existing implementations from the literature and assessed them in terms of reproducibility, identifying gaps and best practices in the field.", "date": "2026-03-18", "categories": []} | |
| {"id": "2603.08682", "title": "Structural Causal Bottleneck Models", "abstract": "We introduce structural causal bottleneck models (SCBMs), a novel class of structural causal models. At the core of SCBMs lies the assumption that causal effects between high-dimensional variables only depend on low-dimensional summary statistics, or bottlenecks, of the causes. SCBMs provide a flexible framework for task-specific dimension reduction while being estimable via standard, simple learning algorithms in practice. We analyse identifiability in SCBMs, connect them to information bottlenecks in the sense of Tishby & Zaslavsky (2015), and illustrate how to estimate them experimentally. We also demonstrate the benefit of bottlenecks for effect estimation in low-sample transfer learning settings. We argue that SCBMs provide an alternative to existing causal dimension reduction frameworks like causal representation learning or causal abstraction learning.", "date": "2026-03-09", "categories": []} | |
| {"id": "2603.12305", "title": "HCP-DCNet: A Hierarchical Causal Primitive Dynamic Composition Network for Self-Improving Causal Understanding", "abstract": "The ability to understand and reason about cause and effect -- encompassing interventions, counterfactuals, and underlying mechanisms -- is a cornerstone of robust artificial intelligence. While deep learning excels at pattern recognition, it fundamentally lacks a model of causality, making systems brittle under distribution shifts and unable to answer ``what-if'' questions. This paper introduces the \\emph{Hierarchical Causal Primitive Dynamic Composition Network (HCP-DCNet)}, a unified framework that bridges continuous physical dynamics with discrete symbolic causal inference. Departing from monolithic representations, HCP-DCNet decomposes causal scenes into reusable, typed \\emph{causal primitives} organized into four abstraction layers: physical, functional, event, and rule. A dual-channel routing network dynamically composes these primitives into task-specific, fully differentiable \\emph{Causal Execution Graphs (CEGs)}. Crucially, the system employs a \\emph{causal-intervention-driven meta-evolution} strategy, enabling autonomous self-improvement through a constrained Markov decision process. We establish rigorous theoretical guarantees, including type-safe composition, routing convergence, and universal approximation of causal dynamics. Extensive experiments across simulated physical and social environments demonstrate that HCP-DCNet significantly outperforms state-of-the-art baselines in causal discovery, counterfactual reasoning, and compositional generalization. This work provides a principled, scalable, and interpretable architecture for building AI systems with human-like causal abstraction and continual self-refinement capabilities.", "date": "2026-03-12", "categories": []} | |
| {"id": "2603.11970", "title": "Statistical and structural identifiability in representation learning", "abstract": "Representation learning models exhibit a surprising stability in their internal representations. Whereas most prior work treats this stability as a single property, we formalize it as two distinct concepts: statistical identifiability (consistency of representations across runs) and structural identifiability (alignment of representations with some unobserved ground truth). Recognizing that perfect pointwise identifiability is generally unrealistic for modern representation learning models, we propose new model-agnostic definitions of statistical and structural near-identifiability of representations up to some error tolerance $\\epsilon$. Leveraging these definitions, we prove a statistical $\\epsilon$-near-identifiability result for the representations of models with nonlinear decoders, generalizing existing identifiability theory beyond last-layer representations in e.g. generative pre-trained transformers (GPTs) to near-identifiability of the intermediate representations of a broad class of models including (masked) autoencoders (MAEs) and supervised learners. Although these weaker assumptions confer weaker identifiability, we show that independent components analysis (ICA) can resolve much of the remaining linear ambiguity for this class of models, and validate and measure our near-identifiability claims empirically. With additional assumptions on the data-generating process, statistical identifiability extends to structural identifiability, yielding a simple and practical recipe for disentanglement: ICA post-processing of latent representations. On synthetic benchmarks, this approach achieves state-of-the-art disentanglement using a vanilla autoencoder. With a foundation model-scale MAE for cell microscopy, it disentangles biological variation from technical batch effects, substantially improving downstream generalization.", "date": "2026-03-12", "categories": []} | |
| {"id": "2402.06223", "title": "Beyond DAGs: A Latent Partial Causal Model for Multimodal Learning", "abstract": "Directed Acyclic Graphs (DAGs) are a standard tool in causal modeling, but their suitability for capturing the complexity of large-scale multimodal data is questionable. In practice, real-world multimodal datasets are often collected from heterogeneous generative processes that do not conform to a single DAG. Instead, they may involve multiple, and even opposing, DAG structures with inverse causal directions. To address this gap, in this work, we first propose a novel latent partial causal model tailored for multimodal data representation learning, featuring two latent coupled variables parts connected by an undirected edge, to represent the transfer of knowledge across modalities. Under specific statistical assumptions, we establish an identifiability result, demonstrating that representations learned by MultiModal Contrastive Learning (MMCL) correspond to the latent coupled variables up to a trivial transformation. This result deepens our understanding of the why MMCL works, highlights its potential for representation disentanglement, and expands the utility of pre-trained models like CLIP. Synthetic experiments confirm the robustness of our findings, even when the assumptions are partially violated. Most importantly, experiments on a pre-trained CLIP model embodies disentangled representations, enabling few-shot learning and improving domain generalization across diverse real-world datasets. Together, these contributions push the boundaries of MMCL, both in theory and in practical applications.", "date": "2024-02-09", "categories": []} | |
| {"id": "2602.24266", "title": "Efficient Discovery of Approximate Causal Abstractions via Neural Mechanism Sparsification", "abstract": "Neural networks are hypothesized to implement interpretable causal mechanisms, yet verifying this requires finding a causal abstraction -- a simpler, high-level Structural Causal Model (SCM) faithful to the network under interventions. Discovering such abstractions is hard: it typically demands brute-force interchange interventions or retraining. We reframe the problem by viewing structured pruning as a search over approximate abstractions. Treating a trained network as a deterministic SCM, we derive an Interventional Risk objective whose second-order expansion yields closed-form criteria for replacing units with constants or folding them into neighbors. Under uniform curvature, our score reduces to activation variance, recovering variance-based pruning as a special case while clarifying when it fails. The resulting procedure efficiently extracts sparse, intervention-faithful abstractions from pretrained networks, which we validate via interchange interventions.", "date": "2026-02-27", "categories": []} | |
| {"id": "2602.24278", "title": "Who Guards the Guardians? The Challenges of Evaluating Identifiability of Learned Representations", "abstract": "Identifiability in representation learning is commonly evaluated using standard metrics (e.g., MCC, DCI, R^2) on synthetic benchmarks with known ground-truth factors. These metrics are assumed to reflect recovery up to the equivalence class guaranteed by identifiability theory. We show that this assumption holds only under specific structural conditions: each metric implicitly encodes assumptions about both the data-generating process (DGP) and the encoder. When these assumptions are violated, metrics become misspecified and can produce systematic false positives and false negatives. Such failures occur both within classical identifiability regimes and in post-hoc settings where identifiability is most needed. We introduce a taxonomy separating DGP assumptions from encoder geometry, use it to characterise the validity domains of existing metrics, and release an evaluation suite for reproducible stress testing and comparison.", "date": "2026-02-27", "categories": []} | |
| {"id": "2602.19782", "title": "Addressing Instrument-Outcome Confounding in Mendelian Randomization through Representation Learning", "abstract": "Mendelian Randomization (MR) is a prominent observational epidemiological research method designed to address unobserved confounding when estimating causal effects. However, core assumptions -- particularly the independence between instruments and unobserved confounders -- are often violated due to population stratification or assortative mating. Leveraging the increasing availability of multi-environment data, we propose a representation learning framework that exploits cross-environment invariance to recover latent exogenous components of genetic instruments. We provide theoretical guarantees for identifying these latent instruments under various mixing mechanisms and demonstrate the effectiveness of our approach through simulations and semi-synthetic experiments using data from the All of Us Research Hub.", "date": "2026-02-23", "categories": []} | |
| {"id": "2208.14153", "title": "Identifying Weight-Variant Latent Causal Models", "abstract": "The task of causal representation learning aims to uncover latent higher-level causal variables that affect lower-level observations. Identifying the true latent causal variables from observed data, while allowing instantaneous causal relations among latent variables, remains a challenge, however. To this end, we start with the analysis of three intrinsic indeterminacies in identifying latent variables from observations: transitivity, permutation indeterminacy, and scaling indeterminacy. We find that transitivity acts as a key role in impeding the identifiability of latent causal variables. To address the unidentifiable issue due to transitivity, we introduce a novel identifiability condition where the underlying latent causal model satisfies a linear-Gaussian model, in which the causal coefficients and the distribution of Gaussian noise are modulated by an additional observed variable. Under certain assumptions, including the existence of a reference condition under which latent causal influences vanish, we can show that the latent causal variables can be identified up to trivial permutation and scaling, and that partial identifiability results can still be obtained when this reference condition is violated for a subset of latent variables. Furthermore, based on these theoretical results, we propose a novel method, termed Structural caUsAl Variational autoEncoder (SuaVE), which directly learns causal representations and causal relationships among them, together with the mapping from the latent causal variables to the observed ones. Experimental results on synthetic and real data demonstrate the identifiability and consistency results and the efficacy of SuaVE in learning causal representations.", "date": "2022-08-30", "categories": []} | |
| {"id": "2602.05033", "title": "Causal Representation Meets Stochastic Modeling under Generic Geometry", "abstract": "Learning meaningful causal representations from observations has emerged as a crucial task for facilitating machine learning applications and driving scientific discoveries in fields such as climate science, biology, and physics. This process involves disentangling high-level latent variables and their causal relationships from low-level observations. Previous work in this area that achieves identifiability typically focuses on cases where the observations are either i.i.d. or follow a latent discrete-time process. Nevertheless, many real-world settings require identifying latent variables that are continuous-time stochastic processes (e.g., multivariate point processes). To this end, we develop identifiable causal representation learning for continuous-time latent stochastic point processes. We study its identifiability by analyzing the geometry of the parameter space. Furthermore, we develop MUTATE, an identifiable variational autoencoder framework with a time-adaptive transition module to infer stochastic dynamics. Across simulated and empirical studies, we find that MUTATE can effectively answer scientific questions, such as the accumulation of mutations in genomics and the mechanisms driving neuron spike triggers in response to time-varying dynamics.", "date": "2026-02-04", "categories": []} | |
| {"id": "2602.02623", "title": "Learning Consistent Causal Abstraction Networks", "abstract": "Causal artificial intelligence aims to enhance explainability, trustworthiness, and robustness in AI by leveraging structural causal models (SCMs). In this pursuit, recent advances formalize network sheaves and cosheaves of causal knowledge. Pushing in the same direction, we tackle the learning of consistent causal abstraction network (CAN), a sheaf-theoretic framework where (i) SCMs are Gaussian, (ii) restriction maps are transposes of constructive linear causal abstractions (CAs) adhering to the semantic embedding principle, and (iii) edge stalks correspond--up to permutation--to the node stalks of more detailed SCMs. Our problem formulation separates into edge-specific local Riemannian problems and avoids nonconvex objectives. We propose an efficient search procedure, solving the local problems with SPECTRAL, our iterative method with closed-form updates and suitable for positive definite and semidefinite covariance matrices. Experiments on synthetic data show competitive performance in the CA learning task, and successful recovery of diverse CAN structures.", "date": "2026-02-02", "categories": []} | |
| {"id": "2505.18410", "title": "On Theoretical Identifiability of Discrete Latent Causal Graphical Models", "abstract": "This paper considers a challenging problem of identifying a causal graphical model under the presence of latent variables. While various identifiability conditions have been proposed in the literature, they often require multiple pure children per latent variable or restrictions on the latent causal graph. Furthermore, it is common for all observed variables to exhibit the same modality. Consequently, the existing identifiability conditions are often too stringent for complex real-world data. We consider a general nonparametric measurement model with arbitrary observed variable types and binary latent variables, and propose a double triangular graphical condition that guarantees identifiability of the entire causal graphical model. The proposed condition significantly relaxes the popular pure children condition. We also establish necessary conditions for identifiability and provide valuable insights into fundamental limits of identifiability. Simulation studies verify that latent structures satisfying our conditions can be accurately estimated from data.", "date": "2025-05-23", "categories": []} | |
| {"id": "2403.15711", "title": "Towards Identifiable Latent Additive Noise Models", "abstract": "Causal representation learning (CRL) offers the promise of uncovering the underlying causal model by which observed data was generated, but the practical applicability of existing methods remains limited by the strong assumptions required for identifiability and by challenges in applying them to real-world settings. Most current approaches are applicable only to relatively restrictive model classes, such as linear or polynomial models, which limits their flexibility and robustness in practice. One promising approach to this problem seeks to address these issues by leveraging changes in causal influences among latent variables. In this vein we propose a more general and relaxed framework than typically applied, formulated by imposing constraints on the function classes applied. Within this framework, we establish partial identifiability results under weaker conditions, including scenarios where only a subset of causal influences change. We then extend our analysis to a broader class of latent post-nonlinear models. Building on these theoretical insights, we develop a flexible method for learning latent causal representations. We demonstrate the effectiveness of our approach on synthetic and semi-synthetic datasets, and further showcase its applicability in a case study on human motion analysis, a complex real-world domain that also highlights the potential to broaden the practical reach of identifiable CRL models.", "date": "2024-03-23", "categories": []} | |
| {"id": "2601.21135", "title": "TRACE: Trajectory Recovery for Continuous Mechanism Evolution in Causal Representation Learning", "abstract": "Temporal causal representation learning methods assume that causal mechanisms switch instantaneously between discrete domains, yet real-world systems often exhibit continuous mechanism transitions. For example, a vehicle's dynamics evolve gradually through a turning maneuver, and human gait shifts smoothly from walking to running. We formalize this setting by modeling transitional mechanisms as convex combinations of finitely many atomic mechanisms, governed by time-varying mixing coefficients. Our theoretical contributions establish that both the latent causal variables and the continuous mixing trajectory are jointly identifiable. We further propose TRACE, a Mixture-of-Experts framework where each expert learns one atomic mechanism during training, enabling recovery of mechanism trajectories at test time. This formulation generalizes to intermediate mechanism states never observed during training. Experiments on synthetic and real-world data demonstrate that TRACE recovers mixing trajectories with up to 0.99 correlation, substantially outperforming discrete-switching baselines.", "date": "2026-01-29", "categories": []} | |
| {"id": "2601.21567", "title": "FlexCausal: Flexible Causal Disentanglement via Structural Flow Priors and Manifold-Aware Interventions", "abstract": "Causal Disentangled Representation Learning(CDRL) aims to learn and disentangle low dimensional representations and their underlying causal structure from observations. However, existing disentanglement methods rely on a standard mean-field approximation with a diagonal posterior covariance, which decorrelates all latent dimensions. Additionally, these methods often assume isotropic Gaussian priors for exogenous noise, failing to capture the complex, non-Gaussian statistical properties prevalent in real-world causal factors. Therefore, we propose FlexCausal, a novel CDRL framework based on a block-diagonal covariance VAE. FlexCausal utilizes a Factorized Flow-based Prior to realistically model the complex densities of exogenous noise, effectively decoupling the learning of causal mechanisms from distributional statistics. By integrating supervised alignment objectives with counterfactual consistency constraints, our framework ensures a precise structural correspondence between the learned latent subspaces and the ground-truth causal relations. Finally, we introduce a manifold-aware relative intervention strategy to ensure high-fidelity generation. Experimental results on both synthetic and real-world datasets demonstrate that FlexCausal significantly outperforms other methods.", "date": "2026-01-29", "categories": []} | |
| {"id": "2504.11609", "title": "Towards Interpretable Deep Generative Models via Causal Representation Learning", "abstract": "Recent developments in generative artificial intelligence (AI) rely on machine learning techniques such as deep learning and generative modeling to achieve state-of-the-art performance across wide-ranging domains. These methods' surprising performance is due in part to their ability to learn implicit \"representations\" of complex, multi-modal data. Unfortunately, deep neural networks are notoriously black boxes that obscure these representations, making them difficult to interpret or analyze. To resolve these difficulties, one approach is to build new interpretable neural network models from the ground up. This is the goal of the emerging field of causal representation learning (CRL) that uses causality as a vector for building flexible, interpretable, and transferable generative AI. CRL can be seen as a synthesis of three intrinsically statistical ideas: (i) latent variable models such as factor analysis; (ii) causal graphical models with latent variables; and (iii) nonparametric statistics and deep learning. This paper introduces CRL from a statistical perspective, focusing on connections to classical models as well as statistical and causal identifiability results. We also highlights key application areas, implementation strategies, and open statistical questions.", "date": "2025-04-15", "categories": []} | |
| {"id": "2601.18118", "title": "LungCRCT: Causal Representation based Lung CT Processing for Lung Cancer Treatment", "abstract": "Due to silence in early stages, lung cancer has been one of the most leading causes of mortality in cancer patients world-wide. Moreover, major symptoms of lung cancer are hard to differentiate with other respiratory disease symptoms such as COPD, further leading patients to overlook cancer progression in early stages. Thus, to enhance survival rates in lung cancer, early detection from consistent proactive respiratory system monitoring becomes crucial. One of the most prevalent and effective methods for lung cancer monitoring would be low-dose computed tomography(LDCT) chest scans, which led to remarkable enhancements in lung cancer detection or tumor classification tasks under rapid advancements and applications of computer vision based AI models such as EfficientNet or ResNet in image processing. However, though advanced CNN models under transfer learning or ViT based models led to high performing lung cancer detections, due to its intrinsic limitations in terms of correlation dependence and low interpretability due to complexity, expansions of deep learning models to lung cancer treatment analysis or causal intervention analysis simulations are still limited. Therefore, this research introduced LungCRCT: a latent causal representation learning based lung cancer analysis framework that retrieves causal representations of factors within the physical causal mechanism of lung cancer progression. With the use of advanced graph autoencoder based causal discovery algorithms with distance Correlation disentanglement and entropy-based image reconstruction refinement, LungCRCT not only enables causal intervention analysis for lung cancer treatments, but also leads to robust, yet extremely light downstream models in malignant tumor classification tasks with an AUC score of 93.91%.", "date": "2026-01-26", "categories": []} | |
| {"id": "2405.15325", "title": "On the Identification of Temporally Causal Representation with Instantaneous Dependence", "abstract": "Temporally causal representation learning aims to identify the latent causal process from time series observations, but most methods require the assumption that the latent causal processes do not have instantaneous relations. Although some recent methods achieve identifiability in the instantaneous causality case, they require either interventions on the latent variables or grouping of the observations, which are in general difficult to obtain in real-world scenarios. To fill this gap, we propose an \\textbf{ID}entification framework for instantane\\textbf{O}us \\textbf{L}atent dynamics (\\textbf{IDOL}) by imposing a sparse influence constraint that the latent causal processes have sparse time-delayed and instantaneous relations. Specifically, we establish identifiability results of the latent causal process based on sufficient variability and the sparse influence constraint by employing contextual information of time series data. Based on these theories, we incorporate a temporally variational inference architecture to estimate the latent variables and a gradient-based sparsity regularization to identify the latent causal process. Experimental results on simulation datasets illustrate that our method can identify the latent causal process. Furthermore, evaluations on multiple human motion forecasting benchmarks with instantaneous dependencies indicate the effectiveness of our method in real-world settings.", "date": "2024-05-24", "categories": []} | |
| {"id": "2512.22150", "title": "Towards Unsupervised Causal Representation Learning via Latent Additive Noise Model Causal Autoencoders", "abstract": "Unsupervised representation learning seeks to recover latent generative factors, yet standard methods relying on statistical independence often fail to capture causal dependencies. A central challenge is identifiability: as established in disentangled representation learning and nonlinear ICA literature, disentangling causal variables from observational data is impossible without supervision, auxiliary signals, or strong inductive biases. In this work, we propose the Latent Additive Noise Model Causal Autoencoder (LANCA) to operationalize the Additive Noise Model (ANM) as a strong inductive bias for unsupervised discovery. Theoretically, we prove that while the ANM constraint does not guarantee unique identifiability in the general mixing case, it resolves component-wise indeterminacy by restricting the admissible transformations from arbitrary diffeomorphisms to the affine class. Methodologically, arguing that the stochastic encoding inherent to VAEs obscures the structural residuals required for latent causal discovery, LANCA employs a deterministic Wasserstein Auto-Encoder (WAE) coupled with a differentiable ANM Layer. This architecture transforms residual independence from a passive assumption into an explicit optimization objective. Empirically, LANCA outperforms state-of-the-art baselines on synthetic physics benchmarks (Pendulum, Flow), and on photorealistic environments (CANDLE), where it demonstrates superior robustness to spurious correlations arising from complex background scenes.", "date": "2025-12-15", "categories": []} | |
| {"id": "2512.18135", "title": "Unifying Causal Reinforcement Learning: Survey, Taxonomy, Algorithms and Applications", "abstract": "Integrating causal inference (CI) with reinforcement learning (RL) has emerged as a powerful paradigm to address critical limitations in classical RL, including low explainability, lack of robustness and generalization failures. Traditional RL techniques, which typically rely on correlation-driven decision-making, struggle when faced with distribution shifts, confounding variables, and dynamic environments. Causal reinforcement learning (CRL), leveraging the foundational principles of causal inference, offers promising solutions to these challenges by explicitly modeling cause-and-effect relationships. In this survey, we systematically review recent advancements at the intersection of causal inference and RL. We categorize existing approaches into causal representation learning, counterfactual policy optimization, offline causal RL, causal transfer learning, and causal explainability. Through this structured analysis, we identify prevailing challenges, highlight empirical successes in practical applications, and discuss open problems. Finally, we provide future research directions, underscoring the potential of CRL for developing robust, generalizable, and interpretable artificial intelligence systems.", "date": "2025-12-19", "categories": []} | |
| {"id": "2506.17182", "title": "Variational Learning of Disentangled Representations", "abstract": "Disentangled representations enable models to separate factors of variation that are shared across experimental conditions from those that are condition-specific. This separation is essential in domains such as biomedical data analysis, where generalization to new treatments, patients, or species depends on isolating stable biological signals from context-dependent effects. While extensions of the variational autoencoder (VAE) framework have been proposed to address this problem, they frequently suffer from leakage between latent representations, limiting their ability to generalize to unseen conditions. Here, we introduce DISCoVeR, a new variational framework that explicitly separates condition-invariant and condition-specific factors. DISCoVeR integrates three key components: (i) a dual-latent architecture that models shared and specific factors separately; (ii) two parallel reconstructions that ensure both representations remain informative; and (iii) a novel max-min objective that encourages clean separation without relying on handcrafted priors, while making only minimal assumptions. Theoretically, we show that this objective maximizes data likelihood while promoting disentanglement, and that it admits a unique equilibrium. Empirically, we demonstrate that DISCoVeR achieves improved disentanglement on synthetic datasets, natural images, and single-cell RNA-seq data. Together, these results establish DISCoVeR as a principled approach for learning disentangled representations in multi-condition settings.", "date": "2025-06-20", "categories": []} | |
| {"id": "2507.05412", "title": "Incorporating Interventional Independence Improves Robustness against Interventional Distribution Shift", "abstract": "We study the problem of learning robust discriminative representations of causally related latent variables given the underlying causal graph and a training set comprising passively collected observational data and interventional data obtained through targeted interventions on some of these latent variables. We desire to learn representations that are robust against the resulting interventional distribution shifts. Existing approaches treat observational and interventional data alike, ignoring the independence relations arising from these interventions, even with known underlying causal models. As a result, their representations lead to large predictive performance disparities between observational and interventional data. This performance disparity worsens when interventional training data is scarce. In this paper, (1) we first identify a strong correlation between this performance disparity and the representations' violation of statistical independence induced during interventions. (2) For linear models, we derive sufficient conditions on the proportion of interventional training data, for which enforcing statistical independence between representations of the intervened node and its non-descendants during interventions lowers the test-time error on interventional data. Combining these insights, (3) we propose RepLIn, a training algorithm that explicitly enforces this statistical independence between interventional representations. We demonstrate the utility of RepLIn on a synthetic dataset, and on real image and text datasets on facial attribute classification and toxicity detection, respectively, with semi-synthetic causal structures. Our experiments show that RepLIn is scalable with the number of nodes in the causal graph and is suitable to improve robustness against interventional distribution shifts of both continuous and discrete latent variables compared to the ERM baselines.", "date": "2025-07-07", "categories": []} | |
| {"id": "2510.04564", "title": "Conditional Representation Learning for Customized Tasks", "abstract": "Conventional representation learning methods learn a universal representation that primarily captures dominant semantics, which may not always align with customized downstream tasks. For instance, in animal habitat analysis, researchers prioritize scene-related features, whereas universal embeddings emphasize categorical semantics, leading to suboptimal results. As a solution, existing approaches resort to supervised fine-tuning, which however incurs high computational and annotation costs. In this paper, we propose Conditional Representation Learning (CRL), aiming to extract representations tailored to arbitrary user-specified criteria. Specifically, we reveal that the semantics of a space are determined by its basis, thereby enabling a set of descriptive words to approximate the basis for a customized feature space. Building upon this insight, given a user-specified criterion, CRL first employs a large language model (LLM) to generate descriptive texts to construct the semantic basis, then projects the image representation into this conditional feature space leveraging a vision-language model (VLM). The conditional representation better captures semantics for the specific criterion, which could be utilized for multiple customized tasks. Extensive experiments on classification and retrieval tasks demonstrate the superiority and generality of the proposed CRL. The code is available at https://github.com/XLearning-SCU/2025-NeurIPS-CRL.", "date": "2025-10-06", "categories": []} | |
| {"id": "2511.06634", "title": "CaberNet: Causal Representation Learning for Cross-Domain HVAC Energy Prediction", "abstract": "Cross-domain HVAC energy prediction is essential for scalable building energy management, particularly because collecting extensive labeled data for every new building is both costly and impractical. Yet, this task remains highly challenging due to the scarcity and heterogeneity of data across different buildings, climate zones, and seasonal patterns. In particular, buildings situated in distinct climatic regions introduce variability that often leads existing methods to overfit to spurious correlations, rely heavily on expert intervention, or compromise on data diversity. To address these limitations, we propose CaberNet, a causal and interpretable deep sequence model that learns invariant (Markov blanket) representations for robust cross-domain prediction. In a purely data-driven fashion and without requiring any prior knowledge, CaberNet integrates i) a global feature gate trained with a self-supervised Bernoulli regularization to distinguish superior causal features from inferior ones, and ii) a domain-wise training scheme that balances domain contributions, minimizes cross-domain loss variance, and promotes latent factor independence. We evaluate CaberNet on real-world datasets collected from three buildings located in three climatically diverse cities, and it consistently outperforms all baselines, achieving a 22.9% reduction in normalized mean squared error (NMSE) compared to the best benchmark. Our code is available at https://github.com/SusCom-Lab/CaberNet-CRL.", "date": "2025-11-10", "categories": []} | |
| {"id": "2512.04980", "title": "Learning Causality for Longitudinal Data", "abstract": "This thesis develops methods for causal inference and causal representation learning (CRL) in high-dimensional, time-varying data.\n The first contribution introduces the Causal Dynamic Variational Autoencoder (CDVAE), a model for estimating Individual Treatment Effects (ITEs) by capturing unobserved heterogeneity in treatment response driven by latent risk factors that affect only outcomes. CDVAE comes with theoretical guarantees on valid latent adjustment and generalization bounds for ITE error. Experiments on synthetic and real datasets show that CDVAE outperforms baselines, and that state-of-the-art models greatly improve when augmented with its latent substitutes, approaching oracle performance without access to true adjustment variables.\n The second contribution proposes an efficient framework for long-term counterfactual regression based on RNNs enhanced with Contrastive Predictive Coding (CPC) and InfoMax. It captures long-range dependencies under time-varying confounding while avoiding the computational cost of transformers, achieving state-of-the-art results and introducing CPC into causal inference.\n The third contribution advances CRL by addressing how latent causes manifest in observed variables. We introduce a model-agnostic interpretability layer based on the geometry of the decoder Jacobian. A sparse self-expression prior induces modular, possibly overlapping groups of observed features aligned with shared latent influences. We provide recovery guarantees in both disjoint and overlapping settings and show that meaningful latent-to-observed structure can be recovered without anchor features or single-parent assumptions. Scalable Jacobian-based regularization techniques are also developed.", "date": "2025-12-04", "categories": []} | |
| {"id": "2512.00048", "title": "Causal Reinforcement Learning based Agent-Patient Interaction with Clinical Domain Knowledge", "abstract": "Reinforcement Learning (RL) faces significant challenges in adaptive healthcare interventions, such as dementia care, where data is scarce, decisions require interpretability, and underlying patient-state dynamic are complex and causal in nature. In this work, we present a novel framework called Causal structure-aware Reinforcement Learning (CRL) that explicitly integrates causal discovery and reasoning into policy optimization. This method enables an agent to learn and exploit a directed acyclic graph (DAG) that describes the causal dependencies between human behavioral states and robot actions, facilitating more efficient, interpretable, and robust decision-making. We validate our approach in a simulated robot-assisted cognitive care scenario, where the agent interacts with a virtual patient exhibiting dynamic emotional, cognitive, and engagement states. The experimental results show that CRL agents outperform conventional model-free RL baselines by achieving higher cumulative rewards, maintaining desirable patient states more consistently, and exhibiting interpretable, clinically-aligned behavior. We further demonstrate that CRL's performance advantage remains robust across different weighting strategies and hyperparameter settings. In addition, we demonstrate a lightweight LLM-based deployment: a fixed policy is embedded into a system prompt that maps inferred states to actions, producing consistent, supportive dialogue without LLM finetuning. Our work illustrates the promise of causal reinforcement learning for human-robot interaction applications, where interpretability, adaptiveness, and data efficiency are paramount.", "date": "2025-11-17", "categories": []} | |
| {"id": "2405.16219", "title": "Structural Disentanglement of Causal and Correlated Concepts", "abstract": "Controllable data generation aims to synthesize data by specifying values for target concepts. Achieving this reliably requires modeling the underlying generative factors and their relationships. In real-world scenarios, these factors exhibit both causal and correlational dependencies, yet most existing methods model only part of this structure. We propose the Causal-Correlation Variational Autoencoder (C2VAE), a unified framework that jointly captures causal and correlational relationships among latent factors. C2VAE organizes the latent space into a structured graph, identifying a set of root causes that govern the generative processes. By optimizing only the root factors relevant to target concepts, the model enables efficient and faithful control. Experiments on synthetic and real-world datasets demonstrate that C2VAE improves generation quality, disentanglement, and intervention fidelity over existing baselines.", "date": "2024-05-25", "categories": []} | |
| {"id": "2506.20582", "title": "Causal Representation Learning with Observational Grouping for CXR Classification", "abstract": "Identifiable causal representation learning seeks to uncover the true causal relationships underlying a data generation process. In medical imaging, this presents opportunities to improve the generalisability and robustness of task-specific latent features. This work introduces the concept of grouping observations to learn identifiable representations for disease classification in chest X-rays via an end-to-end framework. Our experiments demonstrate that these causal representations improve generalisability and robustness across multiple classification tasks when grouping is used to enforce invariance w.r.t race, sex, and imaging views.", "date": "2025-06-25", "categories": []} | |
| {"id": "2405.04715", "title": "Causality Pursuit from Heterogeneous Environments via Neural Adversarial Invariance Learning", "abstract": "Pursuing causality from data is a fundamental problem in scientific discovery, treatment intervention, and transfer learning. This paper introduces a novel algorithmic method for addressing nonparametric invariance and causality learning in regression models across multiple environments, where the joint distribution of response variables and covariates varies, but the conditional expectations of outcome given an unknown set of quasi-causal variables are invariant. The challenge of finding such an unknown set of quasi-causal or invariant variables is compounded by the presence of endogenous variables that have heterogeneous effects across different environments. The proposed Focused Adversarial Invariant Regularization (FAIR) framework utilizes an innovative minimax optimization approach that drives regression models toward prediction-invariant solutions through adversarial testing. Leveraging the representation power of neural networks, FAIR neural networks (FAIR-NN) are introduced for causality pursuit. It is shown that FAIR-NN can find the invariant variables and quasi-causal variables under a minimal identification condition and that the resulting procedure is adaptive to low-dimensional composition structures in a non-asymptotic analysis. Under a structural causal model, variables identified by FAIR-NN represent pragmatic causality and provably align with exact causal mechanisms under conditions of sufficient heterogeneity. Computationally, FAIR-NN employs a novel Gumbel approximation with decreased temperature and a stochastic gradient descent ascent algorithm. The procedures are demonstrated using simulated and real-data examples.", "date": "2024-05-07", "categories": []} | |
| {"id": "2505.17708", "title": "The Third Pillar of Causal Analysis? A Measurement Perspective on Causal Representations", "abstract": "Causal reasoning and discovery, two fundamental tasks of causal analysis, often face challenges in applications due to the complexity, noisiness, and high-dimensionality of real-world data. Despite recent progress in identifying latent causal structures using causal representation learning (CRL), what makes learned representations useful for causal downstream tasks and how to evaluate them are still not well understood. In this paper, we reinterpret CRL using a measurement model framework, where the learned representations are viewed as proxy measurements of the latent causal variables. Our approach clarifies the conditions under which learned representations support downstream causal reasoning and provides a principled basis for quantitatively assessing the quality of representations using a new Test-based Measurement EXclusivity (T-MEX) score. We validate T-MEX across diverse causal inference scenarios, including numerical simulations and real-world ecological video analysis, demonstrating that the proposed framework and corresponding score effectively assess the identification of learned representations and their usefulness for causal downstream tasks.", "date": "2025-05-23", "categories": []} | |
| {"id": "2511.04790", "title": "Causal Structure and Representation Learning with Biomedical Applications", "abstract": "Massive data collection holds the promise of a better understanding of complex phenomena and, ultimately, better decisions. Representation learning has become a key driver of deep learning applications, as it allows learning latent spaces that capture important properties of the data without requiring any supervised annotations. Although representation learning has been hugely successful in predictive tasks, it can fail miserably in causal tasks including predicting the effect of a perturbation/intervention. This calls for a marriage between representation learning and causal inference. An exciting opportunity in this regard stems from the growing availability of multi-modal data (observational and perturbational, imaging-based and sequencing-based, at the single-cell level, tissue-level, and organism-level). We outline a statistical and computational framework for causal structure and representation learning motivated by fundamental biomedical questions: how to effectively use observational and perturbational data to perform causal discovery on observed causal variables; how to use multi-modal views of the system to learn causal variables; and how to design optimal perturbations.", "date": "2025-11-06", "categories": []} | |
| {"id": "2511.03831", "title": "Higher-Order Causal Structure Learning with Additive Models", "abstract": "Causal structure learning has long been the central task of inferring causal insights from data. Despite the abundance of real-world processes exhibiting higher-order mechanisms, however, an explicit treatment of interactions in causal discovery has received little attention. In this work, we focus on extending the causal additive model (CAM) to additive models with higher-order interactions. This second level of modularity we introduce to the structure learning problem is most easily represented by a directed acyclic hypergraph which extends the DAG. We introduce the necessary definitions and theoretical tools to handle the novel structure we introduce and then provide identifiability results for the hyper DAG, extending the typical Markov equivalence classes. We next provide insights into why learning the more complex hypergraph structure may actually lead to better empirical results. In particular, more restrictive assumptions like CAM correspond to easier-to-learn hyper DAGs and better finite sample complexity. We finally develop an extension of the greedy CAM algorithm which can handle the more complex hyper DAG search space and demonstrate its empirical usefulness in synthetic experiments.", "date": "2025-11-05", "categories": []} | |
| {"id": "2510.16026", "title": "A tutorial on discovering and quantifying the effect of latent causal sources of multimodal EHR data", "abstract": "We provide an accessible description of a peer-reviewed generalizable causal machine learning pipeline to (i) discover latent causal sources of large-scale electronic health records observations, and (ii) quantify the source causal effects on clinical outcomes. We illustrate how imperfect multimodal clinical data can be processed, decomposed into probabilistic independent latent sources, and used to train taskspecific causal models from which individual causal effects can be estimated. We summarize the findings of the two real-world applications of the approach to date as a demonstration of its versatility and utility for medical discovery at scale.", "date": "2025-10-15", "categories": []} | |
| {"id": "2405.20482", "title": "Sparsity regularization via tree-structured environments for disentangled representations", "abstract": "Many causal systems such as biological processes in cells can only be observed indirectly via measurements, such as gene expression. Causal representation learning -- the task of correctly mapping low-level observations to latent causal variables -- could advance scientific understanding by enabling inference of latent variables such as pathway activation. In this paper, we develop methods for inferring latent variables from multiple related datasets (environments) and tasks. As a running example, we consider the task of predicting a phenotype from gene expression, where we often collect data from multiple cell types or organisms that are related in known ways. The key insight is that the mapping from latent variables driven by gene expression to the phenotype of interest changes sparsely across closely related environments. To model sparse changes, we introduce Tree-Based Regularization (TBR), an objective that minimizes both prediction error and regularizes closely related environments to learn similar predictors. We prove that under assumptions about the degree of sparse changes, TBR identifies the true latent variables up to some simple transformations. We evaluate the theory empirically with both simulations and ground-truth gene expression data. We find that TBR recovers the latent causal variables better than related methods across these settings, even under settings that violate some assumptions of the theory.", "date": "2024-05-30", "categories": []} | |
| {"id": "2503.04852", "title": "CAUSAL3D: A Comprehensive Benchmark for Causal Learning from Visual Data", "abstract": "True intelligence hinges on the ability to uncover and leverage hidden causal relations. Despite significant progress in AI and computer vision (CV), there remains a lack of benchmarks for assessing models' abilities to infer latent causality from complex visual data. In this paper, we introduce \\textsc{\\textbf{Causal3D}}, a novel and comprehensive benchmark that integrates structured data (tables) with corresponding visual representations (images) to evaluate causal reasoning. Designed within a systematic framework, Causal3D comprises 19 3D-scene datasets capturing diverse causal relations, views, and backgrounds, enabling evaluations across scenes of varying complexity. We assess multiple state-of-the-art methods, including classical causal discovery, causal representation learning, and large/vision-language models (LLMs/VLMs). Our experiments show that as causal structures grow more complex without prior knowledge, performance declines significantly, highlighting the challenges even advanced methods face in complex causal scenarios. Causal3D serves as a vital resource for advancing causal reasoning in CV and fostering trustworthy AI in critical domains.", "date": "2025-03-06", "categories": []} | |
| {"id": "2510.22050", "title": "Towards Error-Centric Intelligence II: Energy-Structured Causal Models", "abstract": "Contemporary machine learning optimizes for predictive accuracy, yet systems that achieve state of the art performance remain causally opaque: their internal representations provide no principled handle for intervention. We can retrain such models, but we cannot surgically edit specific mechanisms while holding others fixed, because learned latent variables lack causal semantics. We argue for a conceptual reorientation: intelligence is the ability to build and refine explanations, falsifiable claims about manipulable structure that specify what changes and what remains invariant under intervention. Explanations subsume prediction but demand more: causal commitments that can be independently tested and corrected at the level of mechanisms. We introduce computational explanations, mappings from observations to intervention ready causal accounts. We instantiate these explanations with Energy Structured Causal Models (ESCMs), in which mechanisms are expressed as constraints (energy functions or vector fields) rather than explicit input output maps, and interventions act by local surgery on those constraints. This shift makes internal structure manipulable at the level where explanations live: which relations must hold, which can change, and what follows when they do. We provide concrete instantiations of the structural-causal principles LAP and ICM in the ESCM context, and also argue that empirical risk minimization systematically produces fractured, entangled representations, a failure we analyze as gauge ambiguity in encoder energy pairs. Finally, we show that under mild conditions, ESCMs recover standard SCM semantics. Building on Part I's principles (LAP, ICM, CAP) and its definition of intelligence as explanation-building under criticism, this paper offers a formal language for causal reasoning in systems that aspire to understand, not merely to predict.", "date": "2025-10-24", "categories": []} | |
| {"id": "2506.09891", "title": "Causal Climate Emulation with Bayesian Filtering", "abstract": "Traditional models of climate change use complex systems of coupled equations to simulate physical processes across the Earth system. These simulations are highly computationally expensive, limiting our predictions of climate change and analyses of its causes and effects. Machine learning has the potential to quickly emulate data from climate models, but current approaches are not able to incorporate physically-based causal relationships. Here, we develop an interpretable climate model emulator based on causal representation learning. We derive a novel approach including a Bayesian filter for stable long-term autoregressive emulation. We demonstrate that our emulator learns accurate climate dynamics, and we show the importance of each one of its components on a realistic synthetic dataset and data from two widely deployed climate models.", "date": "2025-06-11", "categories": []} | |
| {"id": "2510.20884", "title": "ROPES: Robotic Pose Estimation via Score-Based Causal Representation Learning", "abstract": "Causal representation learning (CRL) has emerged as a powerful unsupervised framework that (i) disentangles the latent generative factors underlying high-dimensional data, and (ii) learns the cause-and-effect interactions among the disentangled variables. Despite extensive recent advances in identifiability and some practical progress, a substantial gap remains between theory and real-world practice. This paper takes a step toward closing that gap by bringing CRL to robotics, a domain that has motivated CRL. Specifically, this paper addresses the well-defined robot pose estimation -- the recovery of position and orientation from raw images -- by introducing Robotic Pose Estimation via Score-Based CRL (ROPES). Being an unsupervised framework, ROPES embodies the essence of interventional CRL by identifying those generative factors that are actuated: images are generated by intrinsic and extrinsic latent factors (e.g., joint angles, arm/limb geometry, lighting, background, and camera configuration) and the objective is to disentangle and recover the controllable latent variables, i.e., those that can be directly manipulated (intervened upon) through actuation. Interventional CRL theory shows that variables that undergo variations via interventions can be identified. In robotics, such interventions arise naturally by commanding actuators of various joints and recording images under varied controls. Empirical evaluations in semi-synthetic manipulator experiments demonstrate that ROPES successfully disentangles latent generative factors with high fidelity with respect to the ground truth. Crucially, this is achieved by leveraging only distributional changes, without using any labeled data. The paper also includes a comparison with a baseline based on a recently proposed semi-supervised framework. This paper concludes by positioning robot pose estimation as a near-practical testbed for CRL.", "date": "2025-10-23", "categories": []} | |
| {"id": "2510.18052", "title": "Measure-Theoretic Anti-Causal Representation Learning", "abstract": "Causal representation learning in the anti-causal setting (labels cause features rather than the reverse) presents unique challenges requiring specialized approaches. We propose Anti-Causal Invariant Abstractions (ACIA), a novel measure-theoretic framework for anti-causal representation learning. ACIA employs a two-level design, low-level representations capture how labels generate observations, while high-level representations learn stable causal patterns across environment-specific variations. ACIA addresses key limitations of existing approaches by accommodating prefect and imperfect interventions through interventional kernels, eliminating dependency on explicit causal structures, handling high-dimensional data effectively, and providing theoretical guarantees for out-of-distribution generalization. Experiments on synthetic and real-world medical datasets demonstrate that ACIA consistently outperforms state-of-the-art methods in both accuracy and invariance metrics. Furthermore, our theoretical results establish tight bounds on performance gaps between training and unseen environments, confirming the efficacy of our approach for robust anti-causal learning.", "date": "2025-10-16", "categories": []} | |
| {"id": "2510.18310", "title": "Towards Identifiability of Hierarchical Temporal Causal Representation Learning", "abstract": "Modeling hierarchical latent dynamics behind time series data is critical for capturing temporal dependencies across multiple levels of abstraction in real-world tasks. However, existing temporal causal representation learning methods fail to capture such dynamics, as they fail to recover the joint distribution of hierarchical latent variables from \\textit{single-timestep observed variables}. Interestingly, we find that the joint distribution of hierarchical latent variables can be uniquely determined using three conditionally independent observations. Building on this insight, we propose a Causally Hierarchical Latent Dynamic (CHiLD) identification framework. Our approach first employs temporal contextual observed variables to identify the joint distribution of multi-layer latent variables. Sequentially, we exploit the natural sparsity of the hierarchical structure among latent variables to identify latent variables within each layer. Guided by the theoretical results, we develop a time series generative model grounded in variational inference. This model incorporates a contextual encoder to reconstruct multi-layer latent variables and normalize flow-based hierarchical prior networks to impose the independent noise condition of hierarchical latent dynamics. Empirical evaluations on both synthetic and real-world datasets validate our theoretical claims and demonstrate the effectiveness of CHiLD in modeling hierarchical latent dynamics.", "date": "2025-10-21", "categories": []} | |
| {"id": "2510.16382", "title": "Humanoid-inspired Causal Representation Learning for Domain Generalization", "abstract": "This paper proposes the Humanoid-inspired Structural Causal Model (HSCM), a novel causal framework inspired by human intelligence, designed to overcome the limitations of conventional domain generalization models. Unlike approaches that rely on statistics to capture data-label dependencies and learn distortion-invariant representations, HSCM replicates the hierarchical processing and multi-level learning of human vision systems, focusing on modeling fine-grained causal mechanisms. By disentangling and reweighting key image attributes such as color, texture, and shape, HSCM enhances generalization across diverse domains, ensuring robust performance and interpretability. Leveraging the flexibility and adaptability of human intelligence, our approach enables more effective transfer and learning in dynamic, complex environments. Through both theoretical and empirical evaluations, we demonstrate that HSCM outperforms existing domain generalization models, providing a more principled method for capturing causal relationships and improving model robustness. The code is available at https://github.com/lambett/HSCM.", "date": "2025-10-18", "categories": []} | |
| {"id": "2510.14049", "title": "CausalVerse: Benchmarking Causal Representation Learning with Configurable High-Fidelity Simulations", "abstract": "Causal Representation Learning (CRL) aims to uncover the data-generating process and identify the underlying causal variables and relations, whose evaluation remains inherently challenging due to the requirement of known ground-truth causal variables and causal structure. Existing evaluations often rely on either simplistic synthetic datasets or downstream performance on real-world tasks, generally suffering a dilemma between realism and evaluative precision. In this paper, we introduce a new benchmark for CRL using high-fidelity simulated visual data that retains both realistic visual complexity and, more importantly, access to ground-truth causal generating processes. The dataset comprises around 200 thousand images and 3 million video frames across 24 sub-scenes in four domains: static image generation, dynamic physical simulations, robotic manipulations, and traffic situation analysis. These scenarios range from static to dynamic settings, simple to complex structures, and single to multi-agent interactions, offering a comprehensive testbed that hopefully bridges the gap between rigorous evaluation and real-world applicability. In addition, we provide flexible access to the underlying causal structures, allowing users to modify or configure them to align with the required assumptions in CRL, such as available domain labels, temporal dependencies, or intervention histories. Leveraging this benchmark, we evaluated representative CRL methods across diverse paradigms and offered empirical insights to assist practitioners and newcomers in choosing or extending appropriate CRL frameworks to properly address specific types of real problems that can benefit from the CRL perspective. Welcome to visit our: Project page:https://causal-verse.github.io/, Dataset:https://huggingface.co/CausalVerse.", "date": "2025-10-15", "categories": []} | |
| {"id": "2509.19058", "title": "Towards Causal Representation Learning with Observable Sources as Auxiliaries", "abstract": "Causal representation learning seeks to recover latent factors that generate observational data through a mixing function. Needing assumptions on latent structures or relationships to achieve identifiability in general, prior works often build upon conditional independence given known auxiliary variables. However, prior frameworks limit the scope of auxiliary variables to be external to the mixing function. Yet, in some cases, system-driving latent factors can be easily observed or extracted from data, possibly facilitating identification. In this paper, we introduce a framework of observable sources being auxiliaries, serving as effective conditioning variables. Our main results show that one can identify entire latent variables up to subspace-wise transformations and permutations using volume-preserving encoders. Moreover, when multiple known auxiliary variables are available, we offer a variable-selection scheme to choose those that maximize recoverability of the latent factors given knowledge of the latent causal graph. Finally, we demonstrate the effectiveness of our framework through experiments on synthetic graph and image data, thereby extending the boundaries of current approaches.", "date": "2025-09-23", "categories": []} | |
| {"id": "2509.17228", "title": "Causal Representation Learning from Multimodal Clinical Records under Non-Random Modality Missingness", "abstract": "Clinical notes contain rich patient information, such as diagnoses or medications, making them valuable for patient representation learning. Recent advances in large language models have further improved the ability to extract meaningful representations from clinical texts. However, clinical notes are often missing. For example, in our analysis of the MIMIC-IV dataset, 24.5% of patients have no available discharge summaries. In such cases, representations can be learned from other modalities such as structured data, chest X-rays, or radiology reports. Yet the availability of these modalities is influenced by clinical decision-making and varies across patients, resulting in modality missing-not-at-random (MMNAR) patterns. We propose a causal representation learning framework that leverages observed data and informative missingness in multimodal clinical records. It consists of: (1) an MMNAR-aware modality fusion component that integrates structured data, imaging, and text while conditioning on missingness patterns to capture patient health and clinician-driven assignment; (2) a modality reconstruction component with contrastive learning to ensure semantic sufficiency in representation learning; and (3) a multitask outcome prediction model with a rectifier that corrects for residual bias from specific modality observation patterns. Comprehensive evaluations across MIMIC-IV and eICU show consistent gains over the strongest baselines, achieving up to 13.8% AUC improvement for hospital readmission and 13.1% for ICU admission.", "date": "2025-09-21", "categories": []} | |
| {"id": "2509.03780", "title": "Natural Latents: Latent Variables Stable Across Ontologies", "abstract": "Suppose two Bayesian agents each learn a generative model of the same environment. We will assume the two have converged on the predictive distribution, i.e. distribution over some observables in the environment, but may have different generative models containing different latent variables. Under what conditions can one agent guarantee that their latents are a function of the other agents latents?\n We give simple conditions under which such translation is guaranteed to be possible: the natural latent conditions. We also show that, absent further constraints, these are the most general conditions under which translatability is guaranteed. Crucially for practical application, our theorems are robust to approximation error in the natural latent conditions.", "date": "2025-09-04", "categories": []} | |
| {"id": "2410.00258", "title": "Possible Principles for Aligned Structure Learning Agents", "abstract": "This paper offers a roadmap for the development of scalable aligned artificial intelligence (AI) from first principle descriptions of natural intelligence. In brief, a possible path toward scalable aligned AI rests upon enabling artificial agents to learn a good model of the world that includes a good model of our preferences. For this, the main objective is creating agents that learn to represent the world and other agents' world models; a problem that falls under structure learning (a.k.a. causal representation learning or model discovery). We expose the structure learning and alignment problems with this goal in mind, as well as principles to guide us forward, synthesizing various ideas across mathematics, statistics, and cognitive science. 1) We discuss the essential role of core knowledge, information geometry and model reduction in structure learning, and suggest core structural modules to learn a wide range of naturalistic worlds. 2) We outline a way toward aligned agents through structure learning and theory of mind. As an illustrative example, we mathematically sketch Asimov's Laws of Robotics, which prescribe agents to act cautiously to minimize the ill-being of other agents. We supplement this example by proposing refined approaches to alignment. These observations may guide the development of artificial intelligence in helping to scale existing -- or design new -- aligned structure learning systems.", "date": "2024-09-30", "categories": []} | |
| {"id": "2508.06742", "title": "Learning Causal Structure Distributions for Robust Planning", "abstract": "Structural causal models describe how the components of a robotic system interact. They provide both structural and functional information about the relationships that are present in the system. The structural information outlines the variables among which there is interaction. The functional information describes how such interactions work, via equations or learned models. In this paper we find that learning the functional relationships while accounting for the uncertainty about the structural information leads to more robust dynamics models which improves downstream planning, while using significantly lower computational resources. This in contrast with common model-learning methods that ignore the causal structure and fail to leverage the sparsity of interactions in robotic systems. We achieve this by estimating a causal structure distribution that is used to sample causal graphs that inform the latent-space representations in an encoder-multidecoder probabilistic model. We show that our model can be used to learn the dynamics of a robot, which together with a sampling-based planner can be used to perform new tasks in novel environments, provided an objective function for the new requirement is available. We validate our method using manipulators and mobile robots in both simulation and the real-world. Additionally, we validate the learned dynamics' adaptability and increased robustness to corrupted inputs and changes in the environment, which is highly desirable in challenging real-world robotics scenarios. Video: https://youtu.be/X6k5t7OOnNc.", "date": "2025-08-08", "categories": []} | |
| {"id": "2402.00849", "title": "Score-based Causal Representation Learning: Linear and General Transformations", "abstract": "This paper addresses intervention-based causal representation learning (CRL) under a general nonparametric latent causal model and an unknown transformation that maps the latent variables to the observed variables. Linear and general transformations are investigated. The paper addresses both the identifiability and achievability aspects. Identifiability refers to determining algorithm-agnostic conditions that ensure the recovery of the true latent causal variables and the underlying latent causal graph. Achievability refers to the algorithmic aspects and addresses designing algorithms that achieve identifiability guarantees. By drawing novel connections between score functions (i.e., the gradients of the logarithm of density functions) and CRL, this paper designs a score-based class of algorithms that ensures both identifiability and achievability. First, the paper focuses on linear transformations and shows that one stochastic hard intervention per node suffices to guarantee identifiability. It also provides partial identifiability guarantees for soft interventions, including identifiability up to mixing with parents for general causal models and perfect recovery of the latent graph for sufficiently nonlinear causal models. Secondly, it focuses on general transformations and demonstrates that two stochastic hard interventions per node are sufficient for identifiability. This is achieved by defining a differentiable loss function whose global optima ensure identifiability for general CRL. Notably, one does not need to know which pair of interventional environments has the same node intervened. Finally, the theoretical results are empirically validated via experiments on structured synthetic data and image data.", "date": "2024-02-01", "categories": []} | |
| {"id": "2507.14126", "title": "Toward Temporal Causal Representation Learning with Tensor Decomposition", "abstract": "Temporal causal representation learning is a powerful tool for uncovering complex patterns in observational studies, which are often represented as low-dimensional time series. However, in many real-world applications, data are high-dimensional with varying input lengths and naturally take the form of irregular tensors. To analyze such data, irregular tensor decomposition is critical for extracting meaningful clusters that capture essential information. In this paper, we focus on modeling causal representation learning based on the transformed information. First, we present a novel causal formulation for a set of latent clusters. We then propose CaRTeD, a joint learning framework that integrates temporal causal representation learning with irregular tensor decomposition. Notably, our framework provides a blueprint for downstream tasks using the learned tensor factors, such as modeling latent structures and extracting causal information, and offers a more flexible regularization design to enhance tensor decomposition. Theoretically, we show that our algorithm converges to a stationary point. More importantly, our results fill the gap in theoretical guarantees for the convergence of state-of-the-art irregular tensor decomposition. Experimental results on synthetic and real-world electronic health record (EHR) datasets (MIMIC-III), with extensive benchmarks from both phenotyping and network recovery perspectives, demonstrate that our proposed method outperforms state-of-the-art techniques and enhances the explainability of causal representations.", "date": "2025-07-18", "categories": []} | |
| {"id": "2507.05333", "title": "Causal Foundation Models: Disentangling Physics from Instrument Properties", "abstract": "Foundation models for structured time series data must contend with a fundamental challenge: observations often conflate the true underlying physical phenomena with systematic distortions introduced by measurement instruments. This entanglement limits model generalization, especially in heterogeneous or multi-instrument settings. We present a causally-motivated foundation model that explicitly disentangles physical and instrumental factors using a dual-encoder architecture trained with structured contrastive learning. Leveraging naturally occurring observational triplets (i.e., where the same target is measured under varying conditions, and distinct targets are measured under shared conditions) our model learns separate latent representations for the underlying physical signal and instrument effects. Evaluated on simulated astronomical time series designed to resemble the complexity of variable stars observed by missions like NASA's Transiting Exoplanet Survey Satellite (TESS), our method significantly outperforms traditional single-latent space foundation models on downstream prediction tasks, particularly in low-data regimes. These results demonstrate that our model supports key capabilities of foundation models, including few-shot generalization and efficient adaptation, and highlight the importance of encoding causal structure into representation learning for structured data.", "date": "2025-07-07", "categories": []} | |
| {"id": "2412.07446", "title": "A Causal World Model Underlying Next Token Prediction: Exploring GPT in a Controlled Environment", "abstract": "Are generative pre-trained transformer (GPT) models, trained only to predict the next token, implicitly learning a world model from which sequences are generated one token at a time? We address this question by deriving a causal interpretation of the attention mechanism in GPT and presenting a causal world model that arises from this interpretation. Furthermore, we propose that GPT models, at inference time, can be utilized for zero-shot causal structure learning for input sequences, and introduce a corresponding confidence score. Empirical tests were conducted in controlled environments using the setups of the Othello and Chess strategy games. A GPT, pre-trained on real-world games played with the intention of winning, was tested on out-of-distribution synthetic data consisting of sequences of random legal moves. We find that the GPT model is likely to generate legal next moves for out-of-distribution sequences for which a causal structure is encoded in the attention mechanism with high confidence. In cases where it generates illegal moves, it also fails to capture a causal structure.", "date": "2024-12-10", "categories": []} | |
| {"id": "2507.00440", "title": "A Recipe for Causal Graph Regression: Confounding Effects Revisited", "abstract": "Through recognizing causal subgraphs, causal graph learning (CGL) has risen to be a promising approach for improving the generalizability of graph neural networks under out-of-distribution (OOD) scenarios. However, the empirical successes of CGL techniques are mostly exemplified in classification settings, while regression tasks, a more challenging setting in graph learning, are overlooked. We thus devote this work to tackling causal graph regression (CGR); to this end we reshape the processing of confounding effects in existing CGL studies, which mainly deal with classification. Specifically, we reflect on the predictive power of confounders in graph-level regression, and generalize classification-specific causal intervention techniques to regression through a lens of contrastive learning. Extensive experiments on graph OOD benchmarks validate the efficacy of our proposals for CGR. The model implementation and the code are provided on https://github.com/causal-graph/CGR.", "date": "2025-07-01", "categories": []} | |
| {"id": "2506.17718", "title": "Learning Time-Aware Causal Representation for Model Generalization in Evolving Domains", "abstract": "Endowing deep models with the ability to generalize in dynamic scenarios is of vital significance for real-world deployment, given the continuous and complex changes in data distribution. Recently, evolving domain generalization (EDG) has emerged to address distribution shifts over time, aiming to capture evolving patterns for improved model generalization. However, existing EDG methods may suffer from spurious correlations by modeling only the dependence between data and targets across domains, creating a shortcut between task-irrelevant factors and the target, which hinders generalization. To this end, we design a time-aware structural causal model (SCM) that incorporates dynamic causal factors and the causal mechanism drifts, and propose \\textbf{S}tatic-D\\textbf{YN}amic \\textbf{C}ausal Representation Learning (\\textbf{SYNC}), an approach that effectively learns time-aware causal representations. Specifically, it integrates specially designed information-theoretic objectives into a sequential VAE framework which captures evolving patterns, and produces the desired representations by preserving intra-class compactness of causal factors both across and within domains. Moreover, we theoretically show that our method can yield the optimal causal predictor for each time domain. Results on both synthetic and real-world datasets exhibit that SYNC can achieve superior temporal generalization performance.", "date": "2025-06-21", "categories": []} | |
| {"id": "2411.05331", "title": "Discovering Latent Causal Graphs from Spatiotemporal Data", "abstract": "Many important phenomena in scientific fields like climate, neuroscience, and epidemiology are naturally represented as spatiotemporal gridded data with complex interactions. Inferring causal relationships from these data is a challenging problem compounded by the high dimensionality of such data and the correlations between spatially proximate points. We present SPACY (SPAtiotemporal Causal discoverY), a novel framework based on variational inference, designed to model latent time series and their causal relationships from spatiotemporal data. SPACY alleviates the high-dimensional challenge by discovering causal structures in the latent space. To aggregate spatially proximate, correlated grid points, we use spatial factors, parametrized by spatial kernel functions, to map observational time series to latent representations. Theoretically, we generalize the problem to a continuous spatial domain and establish identifiability when the observations arise from a nonlinear, invertible function of the product of latent series and spatial factors. Using this approach, we avoid assumptions that are often unverifiable, including those about instantaneous effects or sufficient variability. Empirically, SPACY outperforms state-of-the-art baselines on synthetic data, even in challenging settings where existing methods struggle, while remaining scalable for large grids. SPACY also identifies key known phenomena from real-world climate data. An implementation of SPACY is available at https://github.com/Rose-STL-Lab/SPACY/", "date": "2024-11-08", "categories": []} | |
| {"id": "2506.12439", "title": "Interpretable Causal Representation Learning for Biological Data in the Pathway Space", "abstract": "Predicting the impact of genomic and drug perturbations in cellular function is crucial for understanding gene functions and drug effects, ultimately leading to improved therapies. To this end, Causal Representation Learning (CRL) constitutes one of the most promising approaches, as it aims to identify the latent factors that causally govern biological systems, thus facilitating the prediction of the effect of unseen perturbations. Yet, current CRL methods fail in reconciling their principled latent representations with known biological processes, leading to models that are not interpretable. To address this major issue, we present SENA-discrepancy-VAE, a model based on the recently proposed CRL method discrepancy-VAE, that produces representations where each latent factor can be interpreted as the (linear) combination of the activity of a (learned) set of biological processes. To this extent, we present an encoder, SENA-{\\delta}, that efficiently compute and map biological processes' activity levels to the latent causal factors. We show that SENA-discrepancy-VAE achieves predictive performances on unseen combinations of interventions that are comparable with its original, non-interpretable counterpart, while inferring causal latent factors that are biologically meaningful.", "date": "2025-06-14", "categories": []} | |
| {"id": "2506.10378", "title": "Discovering Hierarchical Latent Capabilities of Language Models via Causal Representation Learning", "abstract": "Faithful evaluation of language model capabilities is crucial for deriving actionable insights that can inform model development. However, rigorous causal evaluations in this domain face significant methodological challenges, including complex confounding effects and prohibitive computational costs associated with extensive retraining. To tackle these challenges, we propose a causal representation learning framework wherein observed benchmark performance is modeled as a linear transformation of a few latent capability factors. Crucially, these latent factors are identified as causally interrelated after appropriately controlling for the base model as a common confounder. Applying this approach to a comprehensive dataset encompassing over 1500 models evaluated across six benchmarks from the Open LLM Leaderboard, we identify a concise three-node linear causal structure that reliably explains the observed performance variations. Further interpretation of this causal structure provides substantial scientific insights beyond simple numerical rankings: specifically, we reveal a clear causal direction starting from general problem-solving capabilities, advancing through instruction-following proficiency, and culminating in mathematical reasoning ability. Our results underscore the essential role of carefully controlling base model variations during evaluation, a step critical to accurately uncovering the underlying causal relationships among latent model capabilities.", "date": "2025-06-12", "categories": []} | |
| {"id": "2410.00535", "title": "The Causal Information Bottleneck and Optimal Causal Variable Abstractions", "abstract": "To effectively study complex causal systems, it is often useful to construct abstractions of parts of the system by discarding irrelevant details while preserving key features. The Information Bottleneck (IB) method is a widely used approach to construct variable abstractions by compressing random variables while retaining predictive power over a target variable. Traditional methods like IB are purely statistical and ignore underlying causal structures, making them ill-suited for causal tasks. We propose the Causal Information Bottleneck (CIB), a causal extension of the IB, which compresses a set of chosen variables while maintaining causal control over a target variable. This method produces abstractions of (sets of) variables which are causally interpretable, give us insight about the interactions between the abstracted variables and the target variable, and can be used when reasoning about interventions. We present experimental results demonstrating that the learned abstractions accurately capture causal relations as intended.", "date": "2024-10-01", "categories": []} | |
| {"id": "2502.00407", "title": "Causal Abstraction Learning based on the Semantic Embedding Principle", "abstract": "Structural causal models (SCMs) allow us to investigate complex systems at multiple levels of resolution. The causal abstraction (CA) framework formalizes the mapping between high- and low-level SCMs. We address CA learning in a challenging and realistic setting, where SCMs are inaccessible, interventional data is unavailable, and sample data is misaligned. A key principle of our framework is semantic embedding, formalized as the high-level distribution lying on a subspace of the low-level one. This principle naturally links linear CA to the geometry of the Stiefel manifold. We present a category-theoretic approach to SCMs that enables the learning of a CA by finding a morphism between the low- and high-level probability measures, adhering to the semantic embedding principle. Consequently, we formulate a general CA learning problem. As an application, we solve the latter problem for linear CA; considering Gaussian measures and the Kullback-Leibler divergence as an objective. Given the nonconvexity of the learning task, we develop three algorithms building upon existing paradigms for Riemannian optimization. We demonstrate that the proposed methods succeed on both synthetic and real-world brain data with different degrees of prior information about the structure of CA.", "date": "2025-02-01", "categories": []} | |
| {"id": "2407.14058", "title": "Towards the Causal Complete Cause of Multi-Modal Representation Learning", "abstract": "Multi-Modal Learning (MML) aims to learn effective representations across modalities for accurate predictions. Existing methods typically focus on modality consistency and specificity to learn effective representations. However, from a causal perspective, they may lead to representations that contain insufficient and unnecessary information. To address this, we propose that effective MML representations should be causally sufficient and necessary. Considering practical issues like spurious correlations and modality conflicts, we relax the exogeneity and monotonicity assumptions prevalent in prior works and explore the concepts specific to MML, i.e., Causal Complete Cause $C^3$. We begin by defining $C^3$, which quantifies the probability of representations being causally sufficient and necessary. We then discuss the identifiability of $C^3$ and introduce an instrumental variable to support identifying $C^3$ with non-exogeneity and non-monotonicity. Building on this, we conduct the $C^3$ measurement, i.e., \\(C^3\\) risk. We propose a twin network to estimate it through (i) the real-world branch: utilizing the instrumental variable for sufficiency, and (ii) the hypothetical-world branch: applying gradient-based counterfactual modeling for necessity. Theoretical analyses confirm its reliability. Based on these results, we propose $C^3$ Regularization, a plug-and-play method that enforces the causal completeness of the learned representations by minimizing $C^3$ risk. Extensive experiments demonstrate its effectiveness.", "date": "2024-07-19", "categories": []} | |
| {"id": "2505.12868", "title": "Causality-Inspired Robustness for Nonlinear Models via Representation Learning", "abstract": "Distributional robustness is a central goal of prediction algorithms due to the prevalent distribution shifts in real-world data. The prediction model aims to minimize the worst-case risk among a class of distributions, a.k.a., an uncertainty set. Causality provides a modeling framework with a rigorous robustness guarantee in the above sense, where the uncertainty set is data-driven rather than pre-specified as in traditional distributional robustness optimization. However, current causality-inspired robustness methods possess finite-radius robustness guarantees only in the linear settings, where the causal relationships among the covariates and the response are linear. In this work, we propose a nonlinear method under a causal framework by incorporating recent developments in identifiable representation learning and establish a distributional robustness guarantee. To our best knowledge, this is the first causality-inspired robustness method with such a finite-radius robustness guarantee in nonlinear settings. Empirical validation of the theoretical findings is conducted on both synthetic data and real-world single-cell data, also illustrating that finite-radius robustness is crucial.", "date": "2025-05-19", "categories": []} | |
| {"id": "2505.05192", "title": "Long-Term Individual Causal Effect Estimation via Identifiable Latent Representation Learning", "abstract": "Estimating long-term causal effects by combining long-term observational and short-term experimental data is a crucial but challenging problem in many real-world scenarios. In existing methods, several ideal assumptions, e.g. latent unconfoundedness assumption or additive equi-confounding bias assumption, are proposed to address the latent confounder problem raised by the observational data. However, in real-world applications, these assumptions are typically violated which limits their practical effectiveness. In this paper, we tackle the problem of estimating the long-term individual causal effects without the aforementioned assumptions. Specifically, we propose to utilize the natural heterogeneity of data, such as data from multiple sources, to identify latent confounders, thereby significantly avoiding reliance on idealized assumptions. Practically, we devise a latent representation learning-based estimator of long-term causal effects. Theoretically, we establish the identifiability of latent confounders, with which we further achieve long-term effect identification. Extensive experimental studies, conducted on multiple synthetic and semi-synthetic datasets, demonstrate the effectiveness of our proposed method.", "date": "2025-05-08", "categories": []} | |
| {"id": "2311.03382", "title": "Causal Structure Representation Learning of Confounders in Latent Space\n for Recommendation", "abstract": " Inferring user preferences from the historical feedback of users is a\nvaluable problem in recommender systems. Conventional approaches often rely on\nthe assumption that user preferences in the feedback data are equivalent to the\nreal user preferences without additional noise, which simplifies the problem\nmodeling. However, there are various confounders during user-item interactions,\nsuch as weather and even the recommendation system itself. Therefore,\nneglecting the influence of confounders will result in inaccurate user\npreferences and suboptimal performance of the model. Furthermore, the\nunobservability of confounders poses a challenge in further addressing the\nproblem. To address these issues, we refine the problem and propose a more\nrational solution. Specifically, we consider the influence of confounders,\ndisentangle them from user preferences in the latent space, and employ causal\ngraphs to model their interdependencies without specific labels. By cleverly\ncombining local and global causal graphs, we capture the user-specificity of\nconfounders on user preferences. We theoretically demonstrate the\nidentifiability of the obtained causal graph. Finally, we propose our model\nbased on Variational Autoencoders, named Causal Structure representation\nlearning of Confounders in latent space (CSC). We conducted extensive\nexperiments on one synthetic dataset and five real-world datasets,\ndemonstrating the superiority of our model. Furthermore, we demonstrate that\nthe learned causal representations of confounders are controllable, potentially\noffering users fine-grained control over the objectives of their recommendation\nlists with the learned causal graphs.\n", "date": "2023-11-02", "categories": []} | |
| {"id": "2504.20579", "title": "Representation Learning Preserving Ignorability and Covariate Matching\n for Treatment Effects", "abstract": " Estimating treatment effects from observational data is challenging due to\ntwo main reasons: (a) hidden confounding, and (b) covariate mismatch (control\nand treatment groups not having identical distributions). Long lines of works\nexist that address only either of these issues. To address the former,\nconventional techniques that require detailed knowledge in the form of causal\ngraphs have been proposed. For the latter, covariate matching and importance\nweighting methods have been used. Recently, there has been progress in\ncombining testable independencies with partial side information for tackling\nhidden confounding. A common framework to address both hidden confounding and\nselection bias is missing. We propose neural architectures that aim to learn a\nrepresentation of pre-treatment covariates that is a valid adjustment and also\nsatisfies covariate matching constraints. We combine two different neural\narchitectures: one based on gradient matching across domains created by\nsubsampling a suitable anchor variable that assumes causal side information,\nfollowed by the other, a covariate matching transformation. We prove that\napproximately invariant representations yield approximate valid adjustment sets\nwhich would enable an interval around the true causal effect. In contrast to\nusual sensitivity analysis, where an unknown nuisance parameter is varied, we\nhave a testable approximation yielding a bound on the effect estimate. We also\noutperform various baselines with respect to ATE and PEHE errors on causal\nbenchmarks that include IHDP, Jobs, Cattaneo, and an image-based Crowd\nManagement dataset.\n", "date": "2025-04-29", "categories": []} | |
| {"id": "2502.20099", "title": "Sanity Checking Causal Representation Learning on a Simple Real-World\n System", "abstract": " We evaluate methods for causal representation learning (CRL) on a simple,\nreal-world system where these methods are expected to work. The system consists\nof a controlled optical experiment specifically built for this purpose, which\nsatisfies the core assumptions of CRL and where the underlying causal factors\n(the inputs to the experiment) are known, providing a ground truth. We select\nmethods representative of different approaches to CRL and find that they all\nfail to recover the underlying causal factors. To understand the failure modes\nof the evaluated algorithms, we perform an ablation on the data by substituting\nthe real data-generating process with a simpler synthetic equivalent. The\nresults reveal a reproducibility problem, as most methods already fail on this\nsynthetic ablation despite its simple data-generating process. Additionally, we\nobserve that common assumptions on the mixing function are crucial for the\nperformance of some of the methods but do not hold in the real data. Our\nefforts highlight the contrast between the theoretical promise of the state of\nthe art and the challenges in its application. We hope the benchmark serves as\na simple, real-world sanity check to further develop and validate methodology,\nbridging the gap towards CRL methods that work in practice. We make all code\nand datasets publicly available at github.com/simonbing/CRLSanityCheck\n", "date": "2025-02-27", "categories": []} | |
| {"id": "2402.04869", "title": "Learning by Doing: An Online Causal Reinforcement Learning Framework\n with Causal-Aware Policy", "abstract": " As a key component to intuitive cognition and reasoning solutions in human\nintelligence, causal knowledge provides great potential for reinforcement\nlearning (RL) agents' interpretability towards decision-making by helping\nreduce the searching space. However, there is still a considerable gap in\ndiscovering and incorporating causality into RL, which hinders the rapid\ndevelopment of causal RL. In this paper, we consider explicitly modeling the\ngeneration process of states with the causal graphical model, based on which we\naugment the policy. We formulate the causal structure updating into the RL\ninteraction process with active intervention learning of the environment. To\noptimize the derived objective, we propose a framework with theoretical\nperformance guarantees that alternates between two steps: using interventions\nfor causal structure learning during exploration and using the learned causal\nstructure for policy guidance during exploitation. Due to the lack of public\nbenchmarks that allow direct intervention in the state space, we design the\nroot cause localization task in our simulated fault alarm environment and then\nempirically show the effectiveness and robustness of the proposed method\nagainst state-of-the-art baselines. Theoretical analysis shows that our\nperformance improvement attributes to the virtuous cycle of causal-guided\npolicy learning and causal structure learning, which aligns with our\nexperimental results. Codes are available at\nhttps://github.com/DMIRLAB-Group/FaultAlarm_RL.\n", "date": "2024-02-07", "categories": []} | |
| {"id": "2503.16865", "title": "Nonparametric Factor Analysis and Beyond", "abstract": " Nearly all identifiability results in unsupervised representation learning\ninspired by, e.g., independent component analysis, factor analysis, and causal\nrepresentation learning, rely on assumptions of additive independent noise or\nnoiseless regimes. In contrast, we study the more general case where noise can\ntake arbitrary forms, depend on latent variables, and be non-invertibly\nentangled within a nonlinear function. We propose a general framework for\nidentifying latent variables in the nonparametric noisy settings. We first show\nthat, under suitable conditions, the generative model is identifiable up to\ncertain submanifold indeterminacies even in the presence of non-negligible\nnoise. Furthermore, under the structural or distributional variability\nconditions, we prove that latent variables of the general nonlinear models are\nidentifiable up to trivial indeterminacies. Based on the proposed theoretical\nframework, we have also developed corresponding estimation methods and\nvalidated them in various synthetic and real-world settings. Interestingly, our\nestimate of the true GDP growth from alternative measurements suggests more\ninsightful information on the economies than official reports. We expect our\nframework to provide new insight into how both researchers and practitioners\ndeal with latent variables in real-world scenarios.\n", "date": "2025-03-21", "categories": []} | |
| {"id": "2411.06518", "title": "Causal Representation Learning from Multimodal Biomedical Observations", "abstract": " Prevalent in biomedical applications (e.g., human phenotype research),\nmultimodal datasets can provide valuable insights into the underlying\nphysiological mechanisms. However, current machine learning (ML) models\ndesigned to analyze these datasets often lack interpretability and\nidentifiability guarantees, which are essential for biomedical research. Recent\nadvances in causal representation learning have shown promise in identifying\ninterpretable latent causal variables with formal theoretical guarantees.\nUnfortunately, most current work on multimodal distributions either relies on\nrestrictive parametric assumptions or yields only coarse identification\nresults, limiting their applicability to biomedical research that favors a\ndetailed understanding of the mechanisms.\n In this work, we aim to develop flexible identification conditions for\nmultimodal data and principled methods to facilitate the understanding of\nbiomedical datasets. Theoretically, we consider a nonparametric latent\ndistribution (c.f., parametric assumptions in previous work) that allows for\ncausal relationships across potentially different modalities. We establish\nidentifiability guarantees for each latent component, extending the subspace\nidentification results from previous work. Our key theoretical contribution is\nthe structural sparsity of causal connections between modalities, which, as we\nwill discuss, is natural for a large collection of biomedical systems.\nEmpirically, we present a practical framework to instantiate our theoretical\ninsights. We demonstrate the effectiveness of our approach through extensive\nexperiments on both numerical and synthetic datasets. Results on a real-world\nhuman phenotype dataset are consistent with established biomedical research,\nvalidating our theoretical and methodological framework.\n", "date": "2024-11-10", "categories": []} | |
| {"id": "2503.12784", "title": "Causal Feature Learning in the Social Sciences", "abstract": " Variable selection poses a significant challenge in causal modeling,\nparticularly within the social sciences, where constructs often rely on\ninter-related factors such as age, socioeconomic status, gender, and race.\nIndeed, it has been argued that such attributes must be modeled as macro-level\nabstractions of lower-level manipulable features, in order to preserve the\nmodularity assumption essential to causal inference. This paper accordingly\nextends the theoretical framework of Causal Feature Learning (CFL).\nEmpirically, we apply the CFL algorithm to diverse social science datasets,\nevaluating how CFL-derived macrostates compare with traditional microstates in\ndownstream modeling tasks.\n", "date": "2025-03-17", "categories": []} | |
| {"id": "2503.10834", "title": "On the Identifiability of Causal Abstractions", "abstract": " Causal representation learning (CRL) enhances machine learning models'\nrobustness and generalizability by learning structural causal models associated\nwith data-generating processes. We focus on a family of CRL methods that uses\ncontrastive data pairs in the observable space, generated before and after a\nrandom, unknown intervention, to identify the latent causal model. (Brehmer et\nal., 2022) showed that this is indeed possible, given that all latent variables\ncan be intervened on individually. However, this is a highly restrictive\nassumption in many systems. In this work, we instead assume interventions on\narbitrary subsets of latent variables, which is more realistic. We introduce a\ntheoretical framework that calculates the degree to which we can identify a\ncausal model, given a set of possible interventions, up to an abstraction that\ndescribes the system at a higher level of granularity.\n", "date": "2025-03-13", "categories": []} | |
| {"id": "2503.04358", "title": "Learning Causal Response Representations through Direct Effect Analysis", "abstract": " We propose a novel approach for learning causal response representations. Our\nmethod aims to extract directions in which a multidimensional outcome is most\ndirectly caused by a treatment variable. By bridging conditional independence\ntesting with causal representation learning, we formulate an optimisation\nproblem that maximises the evidence against conditional independence between\nthe treatment and outcome, given a conditioning set. This formulation employs\nflexible regression models tailored to specific applications, creating a\nversatile framework. The problem is addressed through a generalised eigenvalue\ndecomposition. We show that, under mild assumptions, the distribution of the\nlargest eigenvalue can be bounded by a known $F$-distribution, enabling\ntestable conditional independence. We also provide theoretical guarantees for\nthe optimality of the learned representation in terms of signal-to-noise ratio\nand Fisher information maximisation. Finally, we demonstrate the empirical\neffectiveness of our approach in simulation and real-world experiments. Our\nresults underscore the utility of this framework in uncovering direct causal\neffects within complex, multivariate settings.\n", "date": "2025-03-06", "categories": []} | |
| {"id": "2409.02772", "title": "Unifying Causal Representation Learning with the Invariance Principle", "abstract": " Causal representation learning (CRL) aims at recovering latent causal\nvariables from high-dimensional observations to solve causal downstream tasks,\nsuch as predicting the effect of new interventions or more robust\nclassification. A plethora of methods have been developed, each tackling\ncarefully crafted problem settings that lead to different types of\nidentifiability. These different settings are widely assumed to be important\nbecause they are often linked to different rungs of Pearl's causal hierarchy,\neven though this correspondence is not always exact. This work shows that\ninstead of strictly conforming to this hierarchical mapping, many causal\nrepresentation learning approaches methodologically align their representations\nwith inherent data symmetries. Identification of causal variables is guided by\ninvariance principles that are not necessarily causal. This result allows us to\nunify many existing approaches in a single method that can mix and match\ndifferent assumptions, including non-causal ones, based on the invariance\nrelevant to the problem at hand. It also significantly benefits applicability,\nwhich we demonstrate by improving treatment effect estimation on real-world\nhigh-dimensional ecological data. Overall, this paper clarifies the role of\ncausal assumptions in the discovery of causal variables and shifts the focus to\npreserving data symmetries.\n", "date": "2024-09-04", "categories": []} | |
| {"id": "2503.03634", "title": "Feature Matching Intervention: Leveraging Observational Data for Causal\n Representation Learning", "abstract": " A major challenge in causal discovery from observational data is the absence\nof perfect interventions, making it difficult to distinguish causal features\nfrom spurious ones. We propose an innovative approach, Feature Matching\nIntervention (FMI), which uses a matching procedure to mimic perfect\ninterventions. We define causal latent graphs, extending structural causal\nmodels to latent feature space, providing a framework that connects FMI with\ncausal graph learning. Our feature matching procedure emulates perfect\ninterventions within these causal latent graphs. Theoretical results\ndemonstrate that FMI exhibits strong out-of-distribution (OOD)\ngeneralizability. Experiments further highlight FMI's superior performance in\neffectively identifying causal features solely from observational data.\n", "date": "2025-03-05", "categories": []} | |
| {"id": "2406.14302", "title": "Identifiable Exchangeable Mechanisms for Causal Structure and\n Representation Learning", "abstract": " Identifying latent representations or causal structures is important for good\ngeneralization and downstream task performance. However, both fields have been\ndeveloped rather independently. We observe that several methods in both\nrepresentation and causal structure learning rely on the same data-generating\nprocess (DGP), namely, exchangeable but not i.i.d. (independent and identically\ndistributed) data. We provide a unified framework, termed Identifiable\nExchangeable Mechanisms (IEM), for representation and structure learning under\nthe lens of exchangeability. IEM provides new insights that let us relax the\nnecessary conditions for causal structure identification in exchangeable\nnon--i.i.d. data. We also demonstrate the existence of a duality condition in\nidentifiable representation learning, leading to new identifiability results.\nWe hope this work will pave the way for further research in causal\nrepresentation learning.\n", "date": "2024-06-20", "categories": []} | |
| {"id": "2407.16975", "title": "On the Parameter Identifiability of Partially Observed Linear Causal\n Models", "abstract": " Linear causal models are important tools for modeling causal dependencies and\nyet in practice, only a subset of the variables can be observed. In this paper,\nwe examine the parameter identifiability of these models by investigating\nwhether the edge coefficients can be recovered given the causal structure and\npartially observed data. Our setting is more general than that of prior\nresearch - we allow all variables, including both observed and latent ones, to\nbe flexibly related, and we consider the coefficients of all edges, whereas\nmost existing works focus only on the edges between observed variables.\nTheoretically, we identify three types of indeterminacy for the parameters in\npartially observed linear causal models. We then provide graphical conditions\nthat are sufficient for all parameters to be identifiable and show that some of\nthem are provably necessary. Methodologically, we propose a novel\nlikelihood-based parameter estimation method that addresses the variance\nindeterminacy of latent variables in a specific way and can asymptotically\nrecover the underlying parameters up to trivial indeterminacy. Empirical\nstudies on both synthetic and real-world datasets validate our identifiability\ntheory and the effectiveness of the proposed method in the finite-sample\nregime. Code: https://github.com/dongxinshuai/scm-identify.\n", "date": "2024-07-24", "categories": []} | |
| {"id": "2501.11622", "title": "Causal Learning for Heterogeneous Subgroups Based on Nonlinear Causal\n Kernel Clustering", "abstract": " Due to the challenge posed by multi-source and heterogeneous data collected\nfrom diverse environments, causal relationships among features can exhibit\nvariations influenced by different time spans, regions, or strategies. This\ndiversity makes a single causal model inadequate for accurately representing\ncomplex causal relationships in all observational data, a crucial consideration\nin causal learning. To address this challenge, the nonlinear Causal Kernel\nClustering method is introduced for heterogeneous subgroup causal learning,\nhighlighting variations in causal relationships across diverse subgroups. The\nmain component for clustering heterogeneous subgroups lies in the construction\nof the $u$-centered sample mapping function with the property of unbiased\nestimation, which assesses the differences in potential nonlinear causal\nrelationships in various samples and supported by causal identifiability\ntheory. Experimental results indicate that the method performs well in\nidentifying heterogeneous subgroups and enhancing causal learning, leading to a\nreduction in prediction error.\n", "date": "2025-01-20", "categories": []} | |
| {"id": "2502.05037", "title": "Leveraging a Simulator for Learning Causal Representations from\n Post-Treatment Covariates for CATE", "abstract": " Treatment effect estimation involves assessing the impact of different\ntreatments on individual outcomes. Current methods estimate Conditional Average\nTreatment Effect (CATE) using observational datasets where covariates are\ncollected before treatment assignment and outcomes are observed afterward,\nunder assumptions like positivity and unconfoundedness. In this paper, we\naddress a scenario where both covariates and outcomes are gathered after\ntreatment. We show that post-treatment covariates render CATE unidentifiable,\nand recovering CATE requires learning treatment-independent causal\nrepresentations. Prior work shows that such representations can be learned\nthrough contrastive learning if counterfactual supervision is available in\nobservational data. However, since counterfactuals are rare, other works have\nexplored using simulators that offer synthetic counterfactual supervision. Our\ngoal in this paper is to systematically analyze the role of simulators in\nestimating CATE. We analyze the CATE error of several baselines and highlight\ntheir limitations. We then establish a generalization bound that characterizes\nthe CATE error from jointly training on real and simulated distributions, as a\nfunction of the real-simulator mismatch. Finally, we introduce SimPONet, a\nnovel method whose loss function is inspired from our generalization bound. We\nfurther show how SimPONet adjusts the simulator's influence on the learning\nobjective based on the simulator's relevance to the CATE task. We experiment\nwith various DGPs, by systematically varying the real-simulator distribution\ngap to evaluate SimPONet's efficacy against state-of-the-art CATE baselines.\n", "date": "2025-02-07", "categories": []} | |
| {"id": "2405.13888", "title": "Marrying Causal Representation Learning with Dynamical Systems for\n Science", "abstract": " Causal representation learning promises to extend causal models to hidden\ncausal variables from raw entangled measurements. However, most progress has\nfocused on proving identifiability results in different settings, and we are\nnot aware of any successful real-world application. At the same time, the field\nof dynamical systems benefited from deep learning and scaled to countless\napplications but does not allow parameter identification. In this paper, we\ndraw a clear connection between the two and their key assumptions, allowing us\nto apply identifiable methods developed in causal representation learning to\ndynamical systems. At the same time, we can leverage scalable differentiable\nsolvers developed for differential equations to build models that are both\nidentifiable and practical. Overall, we learn explicitly controllable models\nthat isolate the trajectory-specific parameters for further downstream tasks\nsuch as out-of-distribution classification or treatment effect estimation. We\nexperiment with a wind simulator with partially known factors of variation. We\nalso apply the resulting model to real-world climate data and successfully\nanswer downstream causal questions in line with existing literature on climate\nchange.\n", "date": "2024-05-22", "categories": []} | |
| {"id": "2501.15705", "title": "Disentanglement Analysis in Deep Latent Variable Models Matching\n Aggregate Posterior Distributions", "abstract": " Deep latent variable models (DLVMs) are designed to learn meaningful\nrepresentations in an unsupervised manner, such that the hidden explanatory\nfactors are interpretable by independent latent variables (aka\ndisentanglement). The variational autoencoder (VAE) is a popular DLVM widely\nstudied in disentanglement analysis due to the modeling of the posterior\ndistribution using a factorized Gaussian distribution that encourages the\nalignment of the latent factors with the latent axes. Several metrics have been\nproposed recently, assuming that the latent variables explaining the variation\nin data are aligned with the latent axes (cardinal directions). However, there\nare other DLVMs, such as the AAE and WAE-MMD (matching the aggregate posterior\nto the prior), where the latent variables might not be aligned with the latent\naxes. In this work, we propose a statistical method to evaluate disentanglement\nfor any DLVMs in general. The proposed technique discovers the latent vectors\nrepresenting the generative factors of a dataset that can be different from the\ncardinal latent axes. We empirically demonstrate the advantage of the method on\ntwo datasets.\n", "date": "2025-01-26", "categories": []} | |
| {"id": "2501.09163", "title": "Towards Understanding Extrapolation: a Causal Lens", "abstract": " Canonical work handling distribution shifts typically necessitates an entire\ntarget distribution that lands inside the training distribution. However,\npractical scenarios often involve only a handful of target samples, potentially\nlying outside the training support, which requires the capability of\nextrapolation. In this work, we aim to provide a theoretical understanding of\nwhen extrapolation is possible and offer principled methods to achieve it\nwithout requiring an on-support target distribution. To this end, we formulate\nthe extrapolation problem with a latent-variable model that embodies the\nminimal change principle in causal mechanisms. Under this formulation, we cast\nthe extrapolation problem into a latent-variable identification problem. We\nprovide realistic conditions on shift properties and the estimation objectives\nthat lead to identification even when only one off-support target sample is\navailable, tackling the most challenging scenarios. Our theory reveals the\nintricate interplay between the underlying manifold's smoothness and the shift\nproperties. We showcase how our theoretical results inform the design of\npractical adaptation algorithms. Through experiments on both synthetic and\nreal-world data, we validate our theoretical findings and their practical\nimplications.\n", "date": "2025-01-15", "categories": []} | |
| {"id": "2406.00519", "title": "Learning Discrete Concepts in Latent Hierarchical Models", "abstract": " Learning concepts from natural high-dimensional data (e.g., images) holds\npotential in building human-aligned and interpretable machine learning models.\nDespite its encouraging prospect, formalization and theoretical insights into\nthis crucial task are still lacking. In this work, we formalize concepts as\ndiscrete latent causal variables that are related via a hierarchical causal\nmodel that encodes different abstraction levels of concepts embedded in\nhigh-dimensional data (e.g., a dog breed and its eye shapes in natural images).\nWe formulate conditions to facilitate the identification of the proposed causal\nmodel, which reveals when learning such concepts from unsupervised data is\npossible. Our conditions permit complex causal hierarchical structures beyond\nlatent trees and multi-level directed acyclic graphs in prior work and can\nhandle high-dimensional, continuous observed variables, which is well-suited\nfor unstructured data modalities such as images. We substantiate our\ntheoretical claims with synthetic data experiments. Further, we discuss our\ntheory's implications for understanding the underlying mechanisms of latent\ndiffusion models and provide corresponding empirical evidence for our\ntheoretical insights.\n", "date": "2024-06-01", "categories": []} | |
| {"id": "2310.18634", "title": "SSL Framework for Causal Inconsistency between Structures and\n Representations", "abstract": " The cross-pollination between causal discovery and deep learning has led to\nincreasingly extensive interactions. It results in a large number of deep\nlearning data types (such as images, text, etc.) extending into the field of\ncausal discovery, and a multitude of deep learning tasks have begun to utilize\ncausal discovery to explore the internal causal structure and causal\nrepresentation of data. In this paper, we first identified that a complex data\ntype, ``Indefinite Data\", has conflicts between causal relationships expressed\nby the causal structure and causal representation generated by deep learning\nmodels, a phenomenon referred to as causal inconsistency. We thoroughly\nanalyzed related work to explain why only Indefinite Data exhibits causal\ninconsistency while other data types do not. Furthermore, to alleviate causal\ninconsistency, we proposed a self-supervised learning (SSL) framework based on\nintervention, hoping to provide more causal information from different\nintervention views to promote consistency between structure and representation.\nExtensive experiments have shown that the SSL framework enhances causal\nconsistency and can further improve causal structure and representation\nlearning performance. Additionally, we extended the SSL framework to three\ndifferent downstream tasks and LLM instructions. The quantitative results of\nthese applications all reflect the performance improvement brought about by\ncausal consistency.\n", "date": "2023-10-28", "categories": []} | |
| {"id": "2304.09010", "title": "Causal Flow-based Variational Auto-Encoder for Disentangled Causal\n Representation Learning", "abstract": " Disentangled representation learning aims to learn low-dimensional\nrepresentations where each dimension corresponds to an underlying generative\nfactor. While the Variational Auto-Encoder (VAE) is widely used for this\npurpose, most existing methods assume independence among factors, a\nsimplification that does not hold in many real-world scenarios where factors\nare often interdependent and exhibit causal relationships. To overcome this\nlimitation, we propose the Disentangled Causal Variational Auto-Encoder\n(DCVAE), a novel supervised VAE framework that integrates causal flows into the\nrepresentation learning process, enabling the learning of more meaningful and\ninterpretable disentangled representations. We evaluate DCVAE on both synthetic\nand real-world datasets, demonstrating its superior ability in causal\ndisentanglement and intervention experiments. Furthermore, DCVAE outperforms\nstate-of-the-art methods in various downstream tasks, highlighting its\npotential for learning true causal structures among factors.\n", "date": "2023-04-18", "categories": []} | |
| {"id": "2410.23620", "title": "Identifiability Guarantees for Causal Disentanglement from Purely\n Observational Data", "abstract": " Causal disentanglement aims to learn about latent causal factors behind data,\nholding the promise to augment existing representation learning methods in\nterms of interpretability and extrapolation. Recent advances establish\nidentifiability results assuming that interventions on (single) latent factors\nare available; however, it remains debatable whether such assumptions are\nreasonable due to the inherent nature of intervening on latent variables.\nAccordingly, we reconsider the fundamentals and ask what can be learned using\njust observational data.\n We provide a precise characterization of latent factors that can be\nidentified in nonlinear causal models with additive Gaussian noise and linear\nmixing, without any interventions or graphical restrictions. In particular, we\nshow that the causal variables can be identified up to a layer-wise\ntransformation and that further disentanglement is not possible. We transform\nthese theoretical results into a practical algorithm consisting of solving a\nquadratic program over the score estimation of the observed data. We provide\nsimulation results to support our theoretical guarantees and demonstrate that\nour algorithm can derive meaningful causal representations from purely\nobservational data.\n", "date": "2024-10-31", "categories": []} | |
| {"id": "2412.02579", "title": "Factored space models: Towards causality between levels of abstraction", "abstract": " Causality plays an important role in understanding intelligent behavior, and\nthere is a wealth of literature on mathematical models for causality, most of\nwhich is focused on causal graphs. Causal graphs are a powerful tool for a wide\nrange of applications, in particular when the relevant variables are known and\nat the same level of abstraction. However, the given variables can also be\nunstructured data, like pixels of an image. Meanwhile, the causal variables,\nsuch as the positions of objects in the image, can be arbitrary deterministic\nfunctions of the given variables. Moreover, the causal variables may form a\nhierarchy of abstractions, in which the macro-level variables are deterministic\nfunctions of the micro-level variables. Causal graphs are limited when it comes\nto modeling this kind of situation. In the presence of deterministic\nrelationships there is generally no causal graph that satisfies both the Markov\ncondition and the faithfulness condition. We introduce factored space models as\nan alternative to causal graphs which naturally represent both probabilistic\nand deterministic relationships at all levels of abstraction. Moreover, we\nintroduce structural independence and establish that it is equivalent to\nstatistical independence in every distribution that factorizes over the\nfactored space. This theorem generalizes the classical soundness and\ncompleteness theorem for d-separation.\n", "date": "2024-12-03", "categories": []} | |
| {"id": "2412.09843", "title": "Learning Structural Causal Models from Ordering: Identifiable Flow\n Models", "abstract": " In this study, we address causal inference when only observational data and a\nvalid causal ordering from the causal graph are available. We introduce a set\nof flow models that can recover component-wise, invertible transformation of\nexogenous variables. Our flow-based methods offer flexible model design while\nmaintaining causal consistency regardless of the number of discretization\nsteps. We propose design improvements that enable simultaneous learning of all\ncausal mechanisms and reduce abduction and prediction complexity to linear O(n)\nrelative to the number of layers, independent of the number of causal\nvariables. Empirically, we demonstrate that our method outperforms previous\nstate-of-the-art approaches and delivers consistent performance across a wide\nrange of structural causal models in answering observational, interventional,\nand counterfactual questions. Additionally, our method achieves a significant\nreduction in computational time compared to existing diffusion-based\ntechniques, making it practical for large structural causal models.\n", "date": "2024-12-13", "categories": []} | |
| {"id": "2402.09236", "title": "Learning Interpretable Concepts: Unifying Causal Representation Learning\n and Foundation Models", "abstract": " To build intelligent machine learning systems, there are two broad\napproaches. One approach is to build inherently interpretable models, as\nendeavored by the growing field of causal representation learning. The other\napproach is to build highly-performant foundation models and then invest\nefforts into understanding how they work. In this work, we relate these two\napproaches and study how to learn human-interpretable concepts from data.\nWeaving together ideas from both fields, we formally define a notion of\nconcepts and show that they can be provably recovered from diverse data.\nExperiments on synthetic data and large language models show the utility of our\nunified approach.\n", "date": "2024-02-14", "categories": []} | |
| {"id": "2406.05937", "title": "Linear Causal Representation Learning from Unknown Multi-node\n Interventions", "abstract": " Despite the multifaceted recent advances in interventional causal\nrepresentation learning (CRL), they primarily focus on the stylized assumption\nof single-node interventions. This assumption is not valid in a wide range of\napplications, and generally, the subset of nodes intervened in an\ninterventional environment is fully unknown. This paper focuses on\ninterventional CRL under unknown multi-node (UMN) interventional environments\nand establishes the first identifiability results for general latent causal\nmodels (parametric or nonparametric) under stochastic interventions (soft or\nhard) and linear transformation from the latent to observed space.\nSpecifically, it is established that given sufficiently diverse interventional\nenvironments, (i) identifiability up to ancestors is possible using only soft\ninterventions, and (ii) perfect identifiability is possible using hard\ninterventions. Remarkably, these guarantees match the best-known results for\nmore restrictive single-node interventions. Furthermore, CRL algorithms are\nalso provided that achieve the identifiability guarantees. A central step in\ndesigning these algorithms is establishing the relationships between UMN\ninterventional CRL and score functions associated with the statistical models\nof different interventional environments. Establishing these relationships also\nserves as constructive proof of the identifiability guarantees.\n", "date": "2024-06-09", "categories": []} | |
| {"id": "2410.17882", "title": "Identifiable Representation and Model Learning for Latent Dynamic\n Systems", "abstract": " Learning identifiable representations and models from low-level observations\nis helpful for an intelligent spacecraft to complete downstream tasks reliably.\nFor temporal observations, to ensure that the data generating process is\nprovably inverted, most existing works either assume the noise variables in the\ndynamic mechanisms are (conditionally) independent or require that the\ninterventions can directly affect each latent variable. However, in practice,\nthe relationship between the exogenous inputs/interventions and the latent\nvariables may follow some complex deterministic mechanisms. In this work, we\nstudy the problem of identifiable representation and model learning for latent\ndynamic systems. The key idea is to use an inductive bias inspired by\ncontrollable canonical forms, which are sparse and input-dependent by\ndefinition. We prove that, for linear and affine nonlinear latent dynamic\nsystems with sparse input matrices, it is possible to identify the latent\nvariables up to scaling and determine the dynamic models up to some simple\ntransformations. The results have the potential to provide some theoretical\nguarantees for developing more trustworthy decision-making and control methods\nfor intelligent spacecrafts.\n", "date": "2024-10-23", "categories": []} | |
| {"id": "2410.22472", "title": "Learning Identifiable Factorized Causal Representations of Cellular\n Responses", "abstract": " The study of cells and their responses to genetic or chemical perturbations\npromises to accelerate the discovery of therapeutic targets. However, designing\nadequate and insightful models for such data is difficult because the response\nof a cell to perturbations essentially depends on its biological context (e.g.,\ngenetic background or cell type). For example, while discovering therapeutic\ntargets, one may want to enrich for drugs that specifically target a certain\ncell type. This challenge emphasizes the need for methods that explicitly take\ninto account potential interactions between drugs and contexts. Towards this\ngoal, we propose a novel Factorized Causal Representation (FCR) learning method\nthat reveals causal structure in single-cell perturbation data from several\ncell lines. Based on the framework of identifiable deep generative models, FCR\nlearns multiple cellular representations that are disentangled, comprised of\ncovariate-specific ($\\mathbf{z}_x$), treatment-specific ($\\mathbf{z}_{t}$), and\ninteraction-specific ($\\mathbf{z}_{tx}$) blocks. Based on recent advances in\nnon-linear ICA theory, we prove the component-wise identifiability of\n$\\mathbf{z}_{tx}$ and block-wise identifiability of $\\mathbf{z}_t$ and\n$\\mathbf{z}_x$. Then, we present our implementation of FCR, and empirically\ndemonstrate that it outperforms state-of-the-art baselines in various tasks\nacross four single-cell datasets.\n", "date": "2024-10-29", "categories": []} | |
| {"id": "2310.15580", "title": "Identifiable Latent Polynomial Causal Models Through the Lens of Change", "abstract": " Causal representation learning aims to unveil latent high-level causal\nrepresentations from observed low-level data. One of its primary tasks is to\nprovide reliable assurance of identifying these latent causal models, known as\nidentifiability. A recent breakthrough explores identifiability by leveraging\nthe change of causal influences among latent causal variables across multiple\nenvironments \\citep{liu2022identifying}. However, this progress rests on the\nassumption that the causal relationships among latent causal variables adhere\nstrictly to linear Gaussian models. In this paper, we extend the scope of\nlatent causal models to involve nonlinear causal relationships, represented by\npolynomial models, and general noise distributions conforming to the\nexponential family. Additionally, we investigate the necessity of imposing\nchanges on all causal parameters and present partial identifiability results\nwhen part of them remains unchanged. Further, we propose a novel empirical\nestimation method, grounded in our theoretical finding, that enables learning\nconsistent latent causal representations. Our experimental results, obtained\nfrom both synthetic and real-world data, validate our theoretical contributions\nconcerning identifiability and consistency.\n", "date": "2023-10-24", "categories": []} | |
| {"id": "2411.19245", "title": "Contrastive representations of high-dimensional, structured treatments", "abstract": " Estimating causal effects is vital for decision making. In standard causal\neffect estimation, treatments are usually binary- or continuous-valued.\nHowever, in many important real-world settings, treatments can be structured,\nhigh-dimensional objects, such as text, video, or audio. This provides a\nchallenge to traditional causal effect estimation. While leveraging the shared\nstructure across different treatments can help generalize to unseen treatments\nat test time, we show in this paper that using such structure blindly can lead\nto biased causal effect estimation. We address this challenge by devising a\nnovel contrastive approach to learn a representation of the high-dimensional\ntreatments, and prove that it identifies underlying causal factors and discards\nnon-causally relevant factors. We prove that this treatment representation\nleads to unbiased estimates of the causal effect, and empirically validate and\nbenchmark our results on synthetic and real-world datasets.\n", "date": "2024-11-28", "categories": []} | |
| {"id": "2410.24059", "title": "Identifying General Mechanism Shifts in Linear Causal Representations", "abstract": " We consider the linear causal representation learning setting where we\nobserve a linear mixing of $d$ unknown latent factors, which follow a linear\nstructural causal model. Recent work has shown that it is possible to recover\nthe latent factors as well as the underlying structural causal model over them,\nup to permutation and scaling, provided that we have at least $d$ environments,\neach of which corresponds to perfect interventions on a single latent node\n(factor). After this powerful result, a key open problem faced by the community\nhas been to relax these conditions: allow for coarser than perfect single-node\ninterventions, and allow for fewer than $d$ of them, since the number of latent\nfactors $d$ could be very large. In this work, we consider precisely such a\nsetting, where we allow a smaller than $d$ number of environments, and also\nallow for very coarse interventions that can very coarsely \\textit{change the\nentire causal graph over the latent factors}. On the flip side, we relax what\nwe wish to extract to simply the \\textit{list of nodes that have shifted\nbetween one or more environments}. We provide a surprising identifiability\nresult that it is indeed possible, under some very mild standard assumptions,\nto identify the set of shifted nodes. Our identifiability proof moreover is a\nconstructive one: we explicitly provide necessary and sufficient conditions for\na node to be a shifted node, and show that we can check these conditions given\nobserved data. Our algorithm lends itself very naturally to the sample setting\nwhere instead of just interventional distributions, we are provided datasets of\nsamples from each of these distributions. We corroborate our results on both\nsynthetic experiments as well as an interesting psychometric dataset. The code\ncan be found at https://github.com/TianyuCodings/iLCS.\n", "date": "2024-10-31", "categories": []} | |
| {"id": "2410.20405", "title": "Causal Modeling in Multi-Context Systems: Distinguishing Multiple\n Context-Specific Causal Graphs which Account for Observational Support", "abstract": " Causal structure learning with data from multiple contexts carries both\nopportunities and challenges. Opportunities arise from considering shared and\ncontext-specific causal graphs enabling to generalize and transfer causal\nknowledge across contexts. However, a challenge that is currently understudied\nin the literature is the impact of differing observational support between\ncontexts on the identifiability of causal graphs. Here we study in detail\nrecently introduced [6] causal graph objects that capture both causal\nmechanisms and data support, allowing for the analysis of a larger class of\ncontext-specific changes, characterizing distribution shifts more precisely. We\nthereby extend results on the identifiability of context-specific causal\nstructures and propose a framework to model context-specific independence (CSI)\nwithin structural causal models (SCMs) in a refined way that allows to explore\nscenarios where these graph objects differ. We demonstrate how this framework\ncan help explaining phenomena like anomalies or extreme events, where causal\nmechanisms change or appear to change under different conditions. Our results\ncontribute to the theoretical foundations for understanding causal relations in\nmulti-context systems, with implications for generalization, transfer learning,\nand anomaly detection. Future work may extend this approach to more complex\ndata types, such as time-series.\n", "date": "2024-10-27", "categories": []} | |
| {"id": "2410.15491", "title": "Structural Causality-based Generalizable Concept Discovery Models", "abstract": " The rising need for explainable deep neural network architectures has\nutilized semantic concepts as explainable units. Several approaches utilizing\ndisentangled representation learning estimate the generative factors and\nutilize them as concepts for explaining DNNs. However, even though the\ngenerative factors for a dataset remain fixed, concepts are not fixed entities\nand vary based on downstream tasks. In this paper, we propose a disentanglement\nmechanism utilizing a variational autoencoder (VAE) for learning mutually\nindependent generative factors for a given dataset and subsequently learning\ntask-specific concepts using a structural causal model (SCM). Our method\nassumes generative factors and concepts to form a bipartite graph, with\ndirected causal edges from generative factors to concepts. Experiments are\nconducted on datasets with known generative factors: D-sprites and Shapes3D. On\nspecific downstream tasks, our proposed method successfully learns\ntask-specific concepts which are explained well by the causal edges from the\ngenerative factors. Lastly, separate from current causal concept discovery\nmethods, our methodology is generalizable to an arbitrary number of concepts\nand flexible to any downstream tasks.\n", "date": "2024-10-20", "categories": []} | |
| {"id": "2410.07013", "title": "Causal Representation Learning in Temporal Data via Single-Parent\n Decoding", "abstract": " Scientific research often seeks to understand the causal structure underlying\nhigh-level variables in a system. For example, climate scientists study how\nphenomena, such as El Ni\\~no, affect other climate processes at remote\nlocations across the globe. However, scientists typically collect low-level\nmeasurements, such as geographically distributed temperature readings. From\nthese, one needs to learn both a mapping to causally-relevant latent variables,\nsuch as a high-level representation of the El Ni\\~no phenomenon and other\nprocesses, as well as the causal model over them. The challenge is that this\ntask, called causal representation learning, is highly underdetermined from\nobservational data alone, requiring other constraints during learning to\nresolve the indeterminacies. In this work, we consider a temporal model with a\nsparsity assumption, namely single-parent decoding: each observed low-level\nvariable is only affected by a single latent variable. Such an assumption is\nreasonable in many scientific applications that require finding groups of\nlow-level variables, such as extracting regions from geographically gridded\nmeasurement data in climate research or capturing brain regions from neural\nactivity data. We demonstrate the identifiability of the resulting model and\npropose a differentiable method, Causal Discovery with Single-parent Decoding\n(CDSD), that simultaneously learns the underlying latents and a causal graph\nover them. We assess the validity of our theoretical results using simulated\ndata and showcase the practical validity of our method in an application to\nreal-world data from the climate science field.\n", "date": "2024-10-09", "categories": []} | |
| {"id": "2409.08419", "title": "Introducing CausalBench: A Flexible Benchmark Framework for Causal\n Analysis and Machine Learning", "abstract": " While witnessing the exceptional success of machine learning (ML)\ntechnologies in many applications, users are starting to notice a critical\nshortcoming of ML: correlation is a poor substitute for causation. The\nconventional way to discover causal relationships is to use randomized\ncontrolled experiments (RCT); in many situations, however, these are\nimpractical or sometimes unethical. Causal learning from observational data\noffers a promising alternative. While being relatively recent, causal learning\naims to go far beyond conventional machine learning, yet several major\nchallenges remain. Unfortunately, advances are hampered due to the lack of\nunified benchmark datasets, algorithms, metrics, and evaluation service\ninterfaces for causal learning. In this paper, we introduce {\\em CausalBench},\na transparent, fair, and easy-to-use evaluation platform, aiming to (a) enable\nthe advancement of research in causal learning by facilitating scientific\ncollaboration in novel algorithms, datasets, and metrics and (b) promote\nscientific objectivity, reproducibility, fairness, and awareness of bias in\ncausal learning research. CausalBench provides services for benchmarking data,\nalgorithms, models, and metrics, impacting the needs of a broad of scientific\nand engineering disciplines.\n", "date": "2024-09-12", "categories": []} | |
| {"id": "2409.10271", "title": "Causal Discovery in Recommender Systems: Example and Discussion", "abstract": " Causality is receiving increasing attention by the artificial intelligence\nand machine learning communities. This paper gives an example of modelling a\nrecommender system problem using causal graphs. Specifically, we approached the\ncausal discovery task to learn a causal graph by combining observational data\nfrom an open-source dataset with prior knowledge. The resulting causal graph\nshows that only a few variables effectively influence the analysed feedback\nsignals. This contrasts with the recent trend in the machine learning community\nto include more and more variables in massive models, such as neural networks.\n", "date": "2024-09-16", "categories": []} | |
| {"id": "2409.05804", "title": "Celcomen: spatial causal disentanglement for single-cell and tissue\n perturbation modeling", "abstract": " Celcomen leverages a mathematical causality framework to disentangle intra-\nand inter- cellular gene regulation programs in spatial transcriptomics and\nsingle-cell data through a generative graph neural network. It can learn\ngene-gene interactions, as well as generate post-perturbation counterfactual\nspatial transcriptomics, thereby offering access to experimentally inaccessible\nsamples. We validated its disentanglement, identifiability, and counterfactual\nprediction capabilities through simulations and in clinically relevant human\nglioblastoma, human fetal spleen, and mouse lung cancer samples. Celcomen\nprovides the means to model disease and therapy induced changes allowing for\nnew insights into single-cell spatially resolved tissue responses relevant to\nhuman health.\n", "date": "2024-09-09", "categories": []} | |
| {"id": "2409.03142", "title": "Causal Temporal Representation Learning with Nonstationary Sparse\n Transition", "abstract": " Causal Temporal Representation Learning (Ctrl) methods aim to identify the\ntemporal causal dynamics of complex nonstationary temporal sequences. Despite\nthe success of existing Ctrl methods, they require either directly observing\nthe domain variables or assuming a Markov prior on them. Such requirements\nlimit the application of these methods in real-world scenarios when we do not\nhave such prior knowledge of the domain variables. To address this problem,\nthis work adopts a sparse transition assumption, aligned with intuitive human\nunderstanding, and presents identifiability results from a theoretical\nperspective. In particular, we explore under what conditions on the\nsignificance of the variability of the transitions we can build a model to\nidentify the distribution shifts. Based on the theoretical result, we introduce\na novel framework, Causal Temporal Representation Learning with Nonstationary\nSparse Transition (CtrlNS), designed to leverage the constraints on transition\nsparsity and conditional independence to reliably identify both distribution\nshifts and latent factors. Our experimental evaluations on synthetic and\nreal-world datasets demonstrate significant improvements over existing\nbaselines, highlighting the effectiveness of our approach.\n", "date": "2024-09-05", "categories": []} | |
| {"id": "2408.17363", "title": "Look, Learn and Leverage (L$^3$): Mitigating Visual-Domain Shift and\n Discovering Intrinsic Relations via Symbolic Alignment", "abstract": " Modern deep learning models have demonstrated outstanding performance on\ndiscovering the underlying mechanisms when both visual appearance and intrinsic\nrelations (e.g., causal structure) data are sufficient, such as Disentangled\nRepresentation Learning (DRL), Causal Representation Learning (CRL) and Visual\nQuestion Answering (VQA) methods. However, generalization ability of these\nmodels is challenged when the visual domain shifts and the relations data is\nabsent during finetuning. To address this challenge, we propose a novel\nlearning framework, Look, Learn and Leverage (L$^3$), which decomposes the\nlearning process into three distinct phases and systematically utilize the\nclass-agnostic segmentation masks as the common symbolic space to align visual\ndomains. Thus, a relations discovery model can be trained on the source domain,\nand when the visual domain shifts and the intrinsic relations are absent, the\npretrained relations discovery model can be directly reused and maintain a\nsatisfactory performance. Extensive performance evaluations are conducted on\nthree different tasks: DRL, CRL and VQA, and show outstanding results on all\nthree tasks, which reveals the advantages of L$^3$.\n", "date": "2024-08-30", "categories": []} | |
| {"id": "2306.01213", "title": "Learning Causally Disentangled Representations via the Principle of\n Independent Causal Mechanisms", "abstract": " Learning disentangled causal representations is a challenging problem that\nhas gained significant attention recently due to its implications for\nextracting meaningful information for downstream tasks. In this work, we define\na new notion of causal disentanglement from the perspective of independent\ncausal mechanisms. We propose ICM-VAE, a framework for learning causally\ndisentangled representations supervised by causally related observed labels. We\nmodel causal mechanisms using nonlinear learnable flow-based diffeomorphic\nfunctions to map noise variables to latent causal variables. Further, to\npromote the disentanglement of causal factors, we propose a causal\ndisentanglement prior learned from auxiliary labels and the latent causal\nstructure. We theoretically show the identifiability of causal factors and\nmechanisms up to permutation and elementwise reparameterization. We empirically\ndemonstrate that our framework induces highly disentangled causal factors,\nimproves interventional robustness, and is compatible with counterfactual\ngeneration.\n", "date": "2023-06-02", "categories": []} | |
| {"id": "2404.17735", "title": "Causal Diffusion Autoencoders: Toward Counterfactual Generation via\n Diffusion Probabilistic Models", "abstract": " Diffusion probabilistic models (DPMs) have become the state-of-the-art in\nhigh-quality image generation. However, DPMs have an arbitrary noisy latent\nspace with no interpretable or controllable semantics. Although there has been\nsignificant research effort to improve image sample quality, there is little\nwork on representation-controlled generation using diffusion models.\nSpecifically, causal modeling and controllable counterfactual generation using\nDPMs is an underexplored area. In this work, we propose CausalDiffAE, a\ndiffusion-based causal representation learning framework to enable\ncounterfactual generation according to a specified causal model. Our key idea\nis to use an encoder to extract high-level semantically meaningful causal\nvariables from high-dimensional data and model stochastic variation using\nreverse diffusion. We propose a causal encoding mechanism that maps\nhigh-dimensional data to causally related latent factors and parameterize the\ncausal mechanisms among latent factors using neural networks. To enforce the\ndisentanglement of causal variables, we formulate a variational objective and\nleverage auxiliary label information in a prior to regularize the latent space.\nWe propose a DDIM-based counterfactual generation procedure subject to\ndo-interventions. Finally, to address the limited label supervision scenario,\nwe also study the application of CausalDiffAE when a part of the training data\nis unlabeled, which also enables granular control over the strength of\ninterventions in generating counterfactuals during inference. We empirically\nshow that CausalDiffAE learns a disentangled latent space and is capable of\ngenerating high-quality counterfactual images.\n", "date": "2024-04-27", "categories": []} | |
| {"id": "2306.14861", "title": "Leveraging Task Structures for Improved Identifiability in Neural\n Network Representations", "abstract": " This work extends the theory of identifiability in supervised learning by\nconsidering the consequences of having access to a distribution of tasks. In\nsuch cases, we show that linear identifiability is achievable in the general\nmulti-task regression setting. Furthermore, we show that the existence of a\ntask distribution which defines a conditional prior over latent factors reduces\nthe equivalence class for identifiability to permutations and scaling of the\ntrue latent factors, a stronger and more useful result than linear\nidentifiability. Crucially, when we further assume a causal structure over\nthese tasks, our approach enables simple maximum marginal likelihood\noptimization, and suggests potential downstream applications to causal\nrepresentation learning. Empirically, we find that this straightforward\noptimization procedure enables our model to outperform more general\nunsupervised models in recovering canonical representations for both synthetic\ndata and real-world molecular data.\n", "date": "2023-06-26", "categories": []} | |
| {"id": "2402.11124", "title": "Implicit Causal Representation Learning via Switchable Mechanisms", "abstract": " Learning causal representations from observational and interventional data in\nthe absence of known ground-truth graph structures necessitates implicit latent\ncausal representation learning. Implicit learning of causal mechanisms\ntypically involves two categories of interventional data: hard and soft\ninterventions. In real-world scenarios, soft interventions are often more\nrealistic than hard interventions, as the latter require fully controlled\nenvironments. Unlike hard interventions, which directly force changes in a\ncausal variable, soft interventions exert influence indirectly by affecting the\ncausal mechanism. However, the subtlety of soft interventions impose several\nchallenges for learning causal models. One challenge is that soft\nintervention's effects are ambiguous, since parental relations remain intact.\nIn this paper, we tackle the challenges of learning causal models using soft\ninterventions while retaining implicit modelling. We propose ICLR-SM, which\nmodels the effects of soft interventions by employing a causal mechanism switch\nvariable designed to toggle between different causal mechanisms. In our\nexperiments, we consistently observe improved learning of identifiable, causal\nrepresentations, compared to baseline approaches.\n", "date": "2024-02-16", "categories": []} | |
| {"id": "2408.07219", "title": "Causal Effect Estimation using identifiable Variational AutoEncoder with\n Latent Confounders and Post-Treatment Variables", "abstract": " Estimating causal effects from observational data is challenging, especially\nin the presence of latent confounders. Much work has been done on addressing\nthis challenge, but most of the existing research ignores the bias introduced\nby the post-treatment variables. In this paper, we propose a novel method of\njoint Variational AutoEncoder (VAE) and identifiable Variational AutoEncoder\n(iVAE) for learning the representations of latent confounders and latent\npost-treatment variables from their proxy variables, termed CPTiVAE, to achieve\nunbiased causal effect estimation from observational data. We further prove the\nidentifiability in terms of the representation of latent post-treatment\nvariables. Extensive experiments on synthetic and semi-synthetic datasets\ndemonstrate that the CPTiVAE outperforms the state-of-the-art methods in the\npresence of latent confounders and post-treatment variables. We further apply\nCPTiVAE to a real-world dataset to show its potential application.\n", "date": "2024-08-13", "categories": []} | |
| {"id": "2402.05052", "title": "Causal Representation Learning from Multiple Distributions: A General\n Setting", "abstract": " In many problems, the measured variables (e.g., image pixels) are just\nmathematical functions of the latent causal variables (e.g., the underlying\nconcepts or objects). For the purpose of making predictions in changing\nenvironments or making proper changes to the system, it is helpful to recover\nthe latent causal variables $Z_i$ and their causal relations represented by\ngraph $\\mathcal{G}_Z$. This problem has recently been known as causal\nrepresentation learning. This paper is concerned with a general, completely\nnonparametric setting of causal representation learning from multiple\ndistributions (arising from heterogeneous data or nonstationary time series),\nwithout assuming hard interventions behind distribution changes. We aim to\ndevelop general solutions in this fundamental case; as a by product, this helps\nsee the unique benefit offered by other assumptions such as parametric causal\nmodels or hard interventions. We show that under the sparsity constraint on the\nrecovered graph over the latent variables and suitable sufficient change\nconditions on the causal influences, interestingly, one can recover the\nmoralized graph of the underlying directed acyclic graph, and the recovered\nlatent variables and their relations are related to the underlying causal model\nin a specific, nontrivial way. In some cases, most latent variables can even be\nrecovered up to component-wise transformations. Experimental results verify our\ntheoretical claims.\n", "date": "2024-02-07", "categories": []} | |
| {"id": "2408.05788", "title": "Continual Learning of Nonlinear Independent Representations", "abstract": " Identifying the causal relations between interested variables plays a pivotal\nrole in representation learning as it provides deep insights into the dataset.\nIdentifiability, as the central theme of this approach, normally hinges on\nleveraging data from multiple distributions (intervention, distribution shift,\ntime series, etc.). Despite the exciting development in this field, a practical\nbut often overlooked problem is: what if those distribution shifts happen\nsequentially? In contrast, any intelligence possesses the capacity to abstract\nand refine learned knowledge sequentially -- lifelong learning. In this paper,\nwith a particular focus on the nonlinear independent component analysis (ICA)\nframework, we move one step forward toward the question of enabling models to\nlearn meaningful (identifiable) representations in a sequential manner, termed\ncontinual causal representation learning. We theoretically demonstrate that\nmodel identifiability progresses from a subspace level to a component-wise\nlevel as the number of distributions increases. Empirically, we show that our\nmethod achieves performance comparable to nonlinear ICA methods trained jointly\non multiple offline distributions and, surprisingly, the incoming new\ndistribution does not necessarily benefit the identification of all latent\nvariables.\n", "date": "2024-08-11", "categories": []} | |
| {"id": "2310.18615", "title": "Temporally Disentangled Representation Learning under Unknown\n Nonstationarity", "abstract": " In unsupervised causal representation learning for sequential data with\ntime-delayed latent causal influences, strong identifiability results for the\ndisentanglement of causally-related latent variables have been established in\nstationary settings by leveraging temporal structure. However, in nonstationary\nsetting, existing work only partially addressed the problem by either utilizing\nobserved auxiliary variables (e.g., class labels and/or domain indexes) as side\ninformation or assuming simplified latent causal dynamics. Both constrain the\nmethod to a limited range of scenarios. In this study, we further explored the\nMarkov Assumption under time-delayed causally related process in nonstationary\nsetting and showed that under mild conditions, the independent latent\ncomponents can be recovered from their nonlinear mixture up to a permutation\nand a component-wise transformation, without the observation of auxiliary\nvariables. We then introduce NCTRL, a principled estimation framework, to\nreconstruct time-delayed latent causal variables and identify their relations\nfrom measured sequential data only. Empirical evaluations demonstrated the\nreliable identification of time-delayed latent causal influences, with our\nmethodology substantially outperforming existing baselines that fail to exploit\nthe nonstationarity adequately and then, consequently, cannot distinguish\ndistribution shifts.\n", "date": "2023-10-28", "categories": []} | |
| {"id": "2211.03374", "title": "Deep Causal Learning: Representation, Discovery and Inference", "abstract": " Causal learning has garnered significant attention in recent years because it\nreveals the essential relationships that underpin phenomena and delineates the\nmechanisms by which the world evolves. Nevertheless, traditional causal\nlearning methods face numerous challenges and limitations, including\nhigh-dimensional, unstructured variables, combinatorial optimization problems,\nunobserved confounders, selection biases, and estimation inaccuracies. Deep\ncausal learning, which leverages deep neural networks, offers innovative\ninsights and solutions for addressing these challenges. Although numerous deep\nlearning-based methods for causal discovery and inference have been proposed,\nthere remains a dearth of reviews examining the underlying mechanisms by which\ndeep learning can enhance causal learning. In this article, we comprehensively\nreview how deep learning can contribute to causal learning by tackling\ntraditional challenges across three key dimensions: representation, discovery,\nand inference. We emphasize that deep causal learning is pivotal for advancing\nthe theoretical frontiers and broadening the practical applications of causal\nscience. We conclude by summarizing open issues and outlining potential\ndirections for future research.\n", "date": "2022-11-07", "categories": []} | |
| {"id": "2407.13091", "title": "On Causally Disentangled State Representation Learning for Reinforcement\n Learning based Recommender Systems", "abstract": " In Reinforcement Learning-based Recommender Systems (RLRS), the complexity\nand dynamism of user interactions often result in high-dimensional and noisy\nstate spaces, making it challenging to discern which aspects of the state are\ntruly influential in driving the decision-making process. This issue is\nexacerbated by the evolving nature of user preferences and behaviors, requiring\nthe recommender system to adaptively focus on the most relevant information for\ndecision-making while preserving generaliability. To tackle this problem, we\nintroduce an innovative causal approach for decomposing the state and\nextracting \\textbf{C}ausal-\\textbf{I}n\\textbf{D}ispensable \\textbf{S}tate\nRepresentations (CIDS) in RLRS. Our method concentrates on identifying the\n\\textbf{D}irectly \\textbf{A}ction-\\textbf{I}nfluenced \\textbf{S}tate Variables\n(DAIS) and \\textbf{A}ction-\\textbf{I}nfluence \\textbf{A}ncestors (AIA), which\nare essential for making effective recommendations. By leveraging conditional\nmutual information, we develop a framework that not only discerns the causal\nrelationships within the generative process but also isolates critical state\nvariables from the typically dense and high-dimensional state representations.\nWe provide theoretical evidence for the identifiability of these variables.\nThen, by making use of the identified causal relationship, we construct\ncausal-indispensable state representations, enabling the training of policies\nover a more advantageous subset of the agent's state space. We demonstrate the\nefficacy of our approach through extensive experiments, showcasing our method\noutperforms state-of-the-art methods.\n", "date": "2024-07-18", "categories": []} | |
| {"id": "2307.12625", "title": "De-confounding Representation Learning for Counterfactual Inference on\n Continuous Treatment via Generative Adversarial Network", "abstract": " Counterfactual inference for continuous rather than binary treatment\nvariables is more common in real-world causal inference tasks. While there are\nalready some sample reweighting methods based on Marginal Structural Model for\neliminating the confounding bias, they generally focus on removing the\ntreatment's linear dependence on confounders and rely on the accuracy of the\nassumed parametric models, which are usually unverifiable. In this paper, we\npropose a de-confounding representation learning (DRL) framework for\ncounterfactual outcome estimation of continuous treatment by generating the\nrepresentations of covariates disentangled with the treatment variables. The\nDRL is a non-parametric model that eliminates both linear and nonlinear\ndependence between treatment and covariates. Specifically, we train the\ncorrelations between the de-confounded representations and the treatment\nvariables against the correlations between the covariate representations and\nthe treatment variables to eliminate confounding bias. Further, a\ncounterfactual inference network is embedded into the framework to make the\nlearned representations serve both de-confounding and trusted inference.\nExtensive experiments on synthetic datasets show that the DRL model performs\nsuperiorly in learning de-confounding representations and outperforms\nstate-of-the-art counterfactual inference models for continuous treatment\nvariables. In addition, we apply the DRL model to a real-world medical dataset\nMIMIC and demonstrate a detailed causal relationship between red cell width\ndistribution and mortality.\n", "date": "2023-07-24", "categories": []} | |
| {"id": "2407.06015", "title": "Simulation-based Benchmarking for Causal Structure Learning in Gene\n Perturbation Experiments", "abstract": " Causal structure learning (CSL) refers to the task of learning causal\nrelationships from data. Advances in CSL now allow learning of causal graphs in\ndiverse application domains, which has the potential to facilitate data-driven\ncausal decision-making. Real-world CSL performance depends on a number of\n$\\textit{context-specific}$ factors, including context-specific data\ndistributions and non-linear dependencies, that are important in practical\nuse-cases. However, our understanding of how to assess and select CSL methods\nin specific contexts remains limited. To address this gap, we present\n$\\textit{CausalRegNet}$, a multiplicative effect structural causal model that\nallows for generating observational and interventional data incorporating\ncontext-specific properties, with a focus on the setting of gene perturbation\nexperiments. Using real-world gene perturbation data, we show that CausalRegNet\ngenerates accurate distributions and scales far better than current simulation\nframeworks. We illustrate the use of CausalRegNet in assessing CSL methods in\nthe context of interventional experiments in biology.\n", "date": "2024-07-08", "categories": []} | |
| {"id": "2407.04605", "title": "Linear causal disentanglement via higher-order cumulants", "abstract": " Linear causal disentanglement is a recent method in causal representation\nlearning to describe a collection of observed variables via latent variables\nwith causal dependencies between them. It can be viewed as a generalization of\nboth independent component analysis and linear structural equation models. We\nstudy the identifiability of linear causal disentanglement, assuming access to\ndata under multiple contexts, each given by an intervention on a latent\nvariable. We show that one perfect intervention on each latent variable is\nsufficient and in the worst case necessary to recover parameters under perfect\ninterventions, generalizing previous work to allow more latent than observed\nvariables. We give a constructive proof that computes parameters via a coupled\ntensor decomposition. For soft interventions, we find the equivalence class of\nlatent graphs and parameters that are consistent with observed data, via the\nstudy of a system of polynomial equations. Our results hold assuming the\nexistence of non-zero higher-order cumulants, which implies non-Gaussianity of\nvariables.\n", "date": "2024-07-05", "categories": []} | |
| {"id": "2407.02825", "title": "Representation learning with CGAN for casual inference", "abstract": " Conditional Generative Adversarial Nets (CGAN) is often used to improve\nconditional image generation performance. However, there is little research on\nRepresentation learning with CGAN for causal inference. This paper proposes a\nnew method for finding representation learning functions by adopting the\nadversarial idea. We apply the pattern of CGAN and theoretically emonstrate the\nfeasibility of finding a suitable representation function in the context of two\ndistributions being balanced. The theoretical result shows that when two\ndistributions are balanced, the ideal representation function can be found and\nthus can be used to further research.\n", "date": "2024-07-03", "categories": []} | |
| {"id": "2407.00744", "title": "Disentangled Representations for Causal Cognition", "abstract": " Complex adaptive agents consistently achieve their goals by solving problems\nthat seem to require an understanding of causal information, information\npertaining to the causal relationships that exist among elements of combined\nagent-environment systems. Causal cognition studies and describes the main\ncharacteristics of causal learning and reasoning in human and non-human\nanimals, offering a conceptual framework to discuss cognitive performances\nbased on the level of apparent causal understanding of a task. Despite the use\nof formal intervention-based models of causality, including causal Bayesian\nnetworks, psychological and behavioural research on causal cognition does not\nyet offer a computational account that operationalises how agents acquire a\ncausal understanding of the world. Machine and reinforcement learning research\non causality, especially involving disentanglement as a candidate process to\nbuild causal representations, represent on the one hand a concrete attempt at\ndesigning causal artificial agents that can shed light on the inner workings of\nnatural causal cognition. In this work, we connect these two areas of research\nto build a unifying framework for causal cognition that will offer a\ncomputational perspective on studies of animal cognition, and provide insights\nin the development of new algorithms for causal reinforcement learning in AI.\n", "date": "2024-06-30", "categories": []} | |
| {"id": "2211.11695", "title": "Disentangled Representation Learning", "abstract": " Disentangled Representation Learning (DRL) aims to learn a model capable of\nidentifying and disentangling the underlying factors hidden in the observable\ndata in representation form. The process of separating underlying factors of\nvariation into variables with semantic meaning benefits in learning explainable\nrepresentations of data, which imitates the meaningful understanding process of\nhumans when observing an object or relation. As a general learning strategy,\nDRL has demonstrated its power in improving the model explainability,\ncontrolability, robustness, as well as generalization capacity in a wide range\nof scenarios such as computer vision, natural language processing, and data\nmining. In this article, we comprehensively investigate DRL from various\naspects including motivations, definitions, methodologies, evaluations,\napplications, and model designs. We first present two well-recognized\ndefinitions, i.e., Intuitive Definition and Group Theory Definition for\ndisentangled representation learning. We further categorize the methodologies\nfor DRL into four groups from the following perspectives, the model type,\nrepresentation structure, supervision signal, and independence assumption. We\nalso analyze principles to design different DRL models that may benefit\ndifferent tasks in practical applications. Finally, we point out challenges in\nDRL as well as potential research directions deserving future investigations.\nWe believe this work may provide insights for promoting the DRL research in the\ncommunity.\n", "date": "2022-11-21", "categories": []} | |
| {"id": "2406.13371", "title": "Identifiable Causal Representation Learning: Unsupervised, Multi-View,\n and Multi-Environment", "abstract": " Causal models provide rich descriptions of complex systems as sets of\nmechanisms by which each variable is influenced by its direct causes. They\nsupport reasoning about manipulating parts of the system and thus hold promise\nfor addressing some of the open challenges of artificial intelligence (AI),\nsuch as planning, transferring knowledge in changing environments, or\nrobustness to distribution shifts. However, a key obstacle to more widespread\nuse of causal models in AI is the requirement that the relevant variables be\nspecified a priori, which is typically not the case for the high-dimensional,\nunstructured data processed by modern AI systems. At the same time, machine\nlearning (ML) has proven quite successful at automatically extracting useful\nand compact representations of such complex data. Causal representation\nlearning (CRL) aims to combine the core strengths of ML and causality by\nlearning representations in the form of latent variables endowed with causal\nmodel semantics.\n In this thesis, we study and present new results for different CRL settings.\nA central theme is the question of identifiability: Given infinite data, when\nare representations satisfying the same learning objective guaranteed to be\nequivalent? This is an important prerequisite for CRL, as it formally\ncharacterises if and when a learning task is, at least in principle, feasible.\nSince learning causal models, even without a representation learning component,\nis notoriously difficult, we require additional assumptions on the model class\nor rich data beyond the classical i.i.d. setting. By partially characterising\nidentifiability for different settings, this thesis investigates what is\npossible for CRL without direct supervision, and thus contributes to its\ntheoretical foundations. Ideally, the developed insights can help inform data\ncollection practices or inspire the design of new practical estimation methods.\n", "date": "2024-06-19", "categories": []} | |
| {"id": "2403.08335", "title": "A Sparsity Principle for Partially Observable Causal Representation\n Learning", "abstract": " Causal representation learning aims at identifying high-level causal\nvariables from perceptual data. Most methods assume that all latent causal\nvariables are captured in the high-dimensional observations. We instead\nconsider a partially observed setting, in which each measurement only provides\ninformation about a subset of the underlying causal state. Prior work has\nstudied this setting with multiple domains or views, each depending on a fixed\nsubset of latents. Here, we focus on learning from unpaired observations from a\ndataset with an instance-dependent partial observability pattern. Our main\ncontribution is to establish two identifiability results for this setting: one\nfor linear mixing functions without parametric assumptions on the underlying\ncausal model, and one for piecewise linear mixing functions with Gaussian\nlatent causal variables. Based on these insights, we propose two methods for\nestimating the underlying causal variables by enforcing sparsity in the\ninferred representation. Experiments on different simulated datasets and\nestablished benchmarks highlight the effectiveness of our approach in\nrecovering the ground-truth latents.\n", "date": "2024-03-13", "categories": []} | |
| {"id": "2406.05855", "title": "Self-Distilled Disentangled Learning for Counterfactual Prediction", "abstract": " The advancements in disentangled representation learning significantly\nenhance the accuracy of counterfactual predictions by granting precise control\nover instrumental variables, confounders, and adjustable variables. An\nappealing method for achieving the independent separation of these factors is\nmutual information minimization, a task that presents challenges in numerous\nmachine learning scenarios, especially within high-dimensional spaces. To\ncircumvent this challenge, we propose the Self-Distilled Disentanglement\nframework, referred to as $SD^2$. Grounded in information theory, it ensures\ntheoretically sound independent disentangled representations without intricate\nmutual information estimator designs for high-dimensional representations. Our\ncomprehensive experiments, conducted on both synthetic and real-world datasets,\nconfirms the effectiveness of our approach in facilitating counterfactual\ninference in the presence of both observed and unobserved confounders.\n", "date": "2024-06-09", "categories": []} | |
| {"id": "2310.15709", "title": "Causal Representation Learning Made Identifiable by Grouping of\n Observational Variables", "abstract": " A topic of great current interest is Causal Representation Learning (CRL),\nwhose goal is to learn a causal model for hidden features in a data-driven\nmanner. Unfortunately, CRL is severely ill-posed since it is a combination of\nthe two notoriously ill-posed problems of representation learning and causal\ndiscovery. Yet, finding practical identifiability conditions that guarantee a\nunique solution is crucial for its practical applicability. Most approaches so\nfar have been based on assumptions on the latent causal mechanisms, such as\ntemporal causality, or existence of supervision or interventions; these can be\ntoo restrictive in actual applications. Here, we show identifiability based on\nnovel, weak constraints, which requires no temporal structure, intervention,\nnor weak supervision. The approach is based on assuming the observational\nmixing exhibits a suitable grouping of the observational variables. We also\npropose a novel self-supervised estimation framework consistent with the model,\nprove its statistical consistency, and experimentally show its superior CRL\nperformances compared to the state-of-the-art baselines. We further demonstrate\nits robustness against latent confounders and causal cycles.\n", "date": "2023-10-24", "categories": []} | |
| {"id": "2311.18639", "title": "Targeted Reduction of Causal Models", "abstract": " Why does a phenomenon occur? Addressing this question is central to most\nscientific inquiries and often relies on simulations of scientific models. As\nmodels become more intricate, deciphering the causes behind phenomena in\nhigh-dimensional spaces of interconnected variables becomes increasingly\nchallenging. Causal Representation Learning (CRL) offers a promising avenue to\nuncover interpretable causal patterns within these simulations through an\ninterventional lens. However, developing general CRL frameworks suitable for\npractical applications remains an open challenge. We introduce Targeted Causal\nReduction (TCR), a method for condensing complex intervenable models into a\nconcise set of causal factors that explain a specific target phenomenon. We\npropose an information theoretic objective to learn TCR from interventional\ndata of simulations, establish identifiability for continuous variables under\nshift interventions and present a practical algorithm for learning TCRs. Its\nability to generate interpretable high-level explanations from complex models\nis demonstrated on toy and mechanical systems, illustrating its potential to\nassist scientists in the study of complex phenomena in a broad range of\ndisciplines.\n", "date": "2023-11-30", "categories": []} | |
| {"id": "2401.14535", "title": "CaRiNG: Learning Temporal Causal Representation under Non-Invertible\n Generation Process", "abstract": " Identifying the underlying time-delayed latent causal processes in sequential\ndata is vital for grasping temporal dynamics and making downstream reasoning.\nWhile some recent methods can robustly identify these latent causal variables,\nthey rely on strict assumptions about the invertible generation process from\nlatent variables to observed data. However, these assumptions are often hard to\nsatisfy in real-world applications containing information loss. For instance,\nthe visual perception process translates a 3D space into 2D images, or the\nphenomenon of persistence of vision incorporates historical data into current\nperceptions. To address this challenge, we establish an identifiability theory\nthat allows for the recovery of independent latent components even when they\ncome from a nonlinear and non-invertible mix. Using this theory as a\nfoundation, we propose a principled approach, CaRiNG, to learn the CAusal\nRepresentatIon of Non-invertible Generative temporal data with identifiability\nguarantees. Specifically, we utilize temporal context to recover lost latent\ninformation and apply the conditions in our theory to guide the training\nprocess. Through experiments conducted on synthetic datasets, we validate that\nour CaRiNG method reliably identifies the causal process, even when the\ngeneration process is non-invertible. Moreover, we demonstrate that our\napproach considerably improves temporal understanding and reasoning in\npractical applications.\n", "date": "2024-01-25", "categories": []} | |
| {"id": "2310.11011", "title": "From Identifiable Causal Representations to Controllable Counterfactual\n Generation: A Survey on Causal Generative Modeling", "abstract": " Deep generative models have shown tremendous capability in data density\nestimation and data generation from finite samples. While these models have\nshown impressive performance by learning correlations among features in the\ndata, some fundamental shortcomings are their lack of explainability, tendency\nto induce spurious correlations, and poor out-of-distribution extrapolation. To\nremedy such challenges, recent work has proposed a shift toward causal\ngenerative models. Causal models offer several beneficial properties to deep\ngenerative models, such as distribution shift robustness, fairness, and\ninterpretability. Structural causal models (SCMs) describe data-generating\nprocesses and model complex causal relationships and mechanisms among variables\nin a system. Thus, SCMs can naturally be combined with deep generative models.\nWe provide a technical survey on causal generative modeling categorized into\ncausal representation learning and controllable counterfactual generation\nmethods. We focus on fundamental theory, methodology, drawbacks, datasets, and\nmetrics. Then, we cover applications of causal generative models in fairness,\nprivacy, out-of-distribution generalization, precision medicine, and biological\nsciences. Lastly, we discuss open problems and fruitful research directions for\nfuture work in the field.\n", "date": "2023-10-17", "categories": []} | |
| {"id": "2306.11281", "title": "Towards Characterizing Domain Counterfactuals For Invertible Latent\n Causal Models", "abstract": " Answering counterfactual queries has important applications such as\nexplainability, robustness, and fairness but is challenging when the causal\nvariables are unobserved and the observations are non-linear mixtures of these\nlatent variables, such as pixels in images. One approach is to recover the\nlatent Structural Causal Model (SCM), which may be infeasible in practice due\nto requiring strong assumptions, e.g., linearity of the causal mechanisms or\nperfect atomic interventions. Meanwhile, more practical ML-based approaches\nusing naive domain translation models to generate counterfactual samples lack\ntheoretical grounding and may construct invalid counterfactuals. In this work,\nwe strive to strike a balance between practicality and theoretical guarantees\nby analyzing a specific type of causal query called domain counterfactuals,\nwhich hypothesizes what a sample would have looked like if it had been\ngenerated in a different domain (or environment). We show that recovering the\nlatent SCM is unnecessary for estimating domain counterfactuals, thereby\nsidestepping some of the theoretic challenges. By assuming invertibility and\nsparsity of intervention, we prove domain counterfactual estimation error can\nbe bounded by a data fit term and intervention sparsity term. Building upon our\ntheoretical results, we develop a theoretically grounded practical algorithm\nthat simplifies the modeling process to generative model estimation under\nautoregressive and shared parameter constraints that enforce intervention\nsparsity. Finally, we show an improvement in counterfactual estimation over\nbaseline methods through extensive simulated and image-based experiments.\n", "date": "2023-06-20", "categories": []} | |
| {"id": "2311.02695", "title": "Identifying Linearly-Mixed Causal Representations from Multi-Node\n Interventions", "abstract": " The task of inferring high-level causal variables from low-level\nobservations, commonly referred to as causal representation learning, is\nfundamentally underconstrained. As such, recent works to address this problem\nfocus on various assumptions that lead to identifiability of the underlying\nlatent causal variables. A large corpus of these preceding approaches consider\nmulti-environment data collected under different interventions on the causal\nmodel. What is common to virtually all of these works is the restrictive\nassumption that in each environment, only a single variable is intervened on.\nIn this work, we relax this assumption and provide the first identifiability\nresult for causal representation learning that allows for multiple variables to\nbe targeted by an intervention within one environment. Our approach hinges on a\ngeneral assumption on the coverage and diversity of interventions across\nenvironments, which also includes the shared assumption of single-node\ninterventions of previous works. The main idea behind our approach is to\nexploit the trace that interventions leave on the variance of the ground truth\ncausal variables and regularizing for a specific notion of sparsity with\nrespect to this trace. In addition to and inspired by our theoretical\ncontributions, we present a practical algorithm to learn causal representations\nfrom multi-node interventional data and provide empirical evidence that\nvalidates our identifiability results.\n", "date": "2023-11-05", "categories": []} | |
| {"id": "2403.09830", "title": "Towards the Reusability and Compositionality of Causal Representations", "abstract": " Causal Representation Learning (CRL) aims at identifying high-level causal\nfactors and their relationships from high-dimensional observations, e.g.,\nimages. While most CRL works focus on learning causal representations in a\nsingle environment, in this work we instead propose a first step towards\nlearning causal representations from temporal sequences of images that can be\nadapted in a new environment, or composed across multiple related environments.\nIn particular, we introduce DECAF, a framework that detects which causal\nfactors can be reused and which need to be adapted from previously learned\ncausal representations. Our approach is based on the availability of\nintervention targets, that indicate which variables are perturbed at each time\nstep. Experiments on three benchmark datasets show that integrating our\nframework with four state-of-the-art CRL approaches leads to accurate\nrepresentations in a new environment with only a few samples.\n", "date": "2024-03-14", "categories": []} | |
| {"id": "2311.04056", "title": "Multi-View Causal Representation Learning with Partial Observability", "abstract": " We present a unified framework for studying the identifiability of\nrepresentations learned from simultaneously observed views, such as different\ndata modalities. We allow a partially observed setting in which each view\nconstitutes a nonlinear mixture of a subset of underlying latent variables,\nwhich can be causally related. We prove that the information shared across all\nsubsets of any number of views can be learned up to a smooth bijection using\ncontrastive learning and a single encoder per view. We also provide graphical\ncriteria indicating which latent variables can be identified through a simple\nset of rules, which we refer to as identifiability algebra. Our general\nframework and theoretical results unify and extend several previous works on\nmulti-view nonlinear ICA, disentanglement, and causal representation learning.\nWe experimentally validate our claims on numerical, image, and multi-modal data\nsets. Further, we demonstrate that the performance of prior methods is\nrecovered in different special cases of our setup. Overall, we find that access\nto multiple partial views enables us to identify a more fine-grained\nrepresentation, under the generally milder assumption of partial observability.\n", "date": "2023-11-07", "categories": []} | |
| {"id": "2310.04295", "title": "Identifying Representations for Intervention Extrapolation", "abstract": " The premise of identifiable and causal representation learning is to improve\nthe current representation learning paradigm in terms of generalizability or\nrobustness. Despite recent progress in questions of identifiability, more\ntheoretical results demonstrating concrete advantages of these methods for\ndownstream tasks are needed. In this paper, we consider the task of\nintervention extrapolation: predicting how interventions affect an outcome,\neven when those interventions are not observed at training time, and show that\nidentifiable representations can provide an effective solution to this task\neven if the interventions affect the outcome non-linearly. Our setup includes\nan outcome Y, observed features X, which are generated as a non-linear\ntransformation of latent features Z, and exogenous action variables A, which\ninfluence Z. The objective of intervention extrapolation is to predict how\ninterventions on A that lie outside the training support of A affect Y. Here,\nextrapolation becomes possible if the effect of A on Z is linear and the\nresidual when regressing Z on A has full support. As Z is latent, we combine\nthe task of intervention extrapolation with identifiable representation\nlearning, which we call Rep4Ex: we aim to map the observed features X into a\nsubspace that allows for non-linear extrapolation in A. We show that the hidden\nrepresentation is identifiable up to an affine transformation in Z-space, which\nis sufficient for intervention extrapolation. The identifiability is\ncharacterized by a novel constraint describing the linearity assumption of A on\nZ. Based on this insight, we propose a method that enforces the linear\ninvariance constraint and can be combined with any type of autoencoder. We\nvalidate our theoretical findings through synthetic experiments and show that\nour approach succeeds in predicting the effects of unseen interventions.\n", "date": "2023-10-06", "categories": []} | |
| {"id": "2307.16387", "title": "Relation-First Modeling Paradigm for Causal Representation Learning\n toward the Development of AGI", "abstract": " The traditional i.i.d.-based learning paradigm faces inherent challenges in\naddressing causal relationships, which has become increasingly evident with the\nrise of applications in causal representation learning.\n Our understanding of causality naturally requires a perspective as the\ncreator rather than observer, as the ``what...if'' questions only hold within\nthe possible world we conceive. The traditional perspective limits capturing\ndynamic causal outcomes and leads to compensatory efforts such as the reliance\non hidden confounders. This paper lays the groundwork for the new perspective,\nwhich enables the \\emph{relation-first} modeling paradigm for causality. Also,\nit introduces the Relation-Indexed Representation Learning (RIRL) as a\npractical implementation, supported by experiments that validate its efficacy.\n", "date": "2023-07-31", "categories": []} | |
| {"id": "2209.11924", "title": "Interventional Causal Representation Learning", "abstract": " Causal representation learning seeks to extract high-level latent factors\nfrom low-level sensory data. Most existing methods rely on observational data\nand structural assumptions (e.g., conditional independence) to identify the\nlatent factors. However, interventional data is prevalent across applications.\nCan interventional data facilitate causal representation learning? We explore\nthis question in this paper. The key observation is that interventional data\noften carries geometric signatures of the latent factors' support (i.e. what\nvalues each latent can possibly take). For example, when the latent factors are\ncausally connected, interventions can break the dependency between the\nintervened latents' support and their ancestors'. Leveraging this fact, we\nprove that the latent causal factors can be identified up to permutation and\nscaling given data from perfect $do$ interventions. Moreover, we can achieve\nblock affine identification, namely the estimated latent factors are only\nentangled with a few other latents if we have access to data from imperfect\ninterventions. These results highlight the unique power of interventional data\nin causal representation learning; they can enable provable identification of\nlatent factors without any assumptions about their distributions or dependency\nstructure.\n", "date": "2022-09-24", "categories": []} | |
| {"id": "2402.15309", "title": "Counterfactual Generation with Identifiability Guarantees", "abstract": " Counterfactual generation lies at the core of various machine learning tasks,\nincluding image translation and controllable text generation. This generation\nprocess usually requires the identification of the disentangled latent\nrepresentations, such as content and style, that underlie the observed data.\nHowever, it becomes more challenging when faced with a scarcity of paired data\nand labeling information. Existing disentangled methods crucially rely on\noversimplified assumptions, such as assuming independent content and style\nvariables, to identify the latent variables, even though such assumptions may\nnot hold for complex data distributions. For instance, food reviews tend to\ninvolve words like tasty, whereas movie reviews commonly contain words such as\nthrilling for the same positive sentiment. This problem is exacerbated when\ndata are sampled from multiple domains since the dependence between content and\nstyle may vary significantly over domains. In this work, we tackle the\ndomain-varying dependence between the content and the style variables inherent\nin the counterfactual generation task. We provide identification guarantees for\nsuch latent-variable models by leveraging the relative sparsity of the\ninfluences from different latent variables. Our theoretical insights enable the\ndevelopment of a doMain AdapTive counTerfactual gEneration model, called\n(MATTE). Our theoretically grounded framework achieves state-of-the-art\nperformance in unsupervised style transfer tasks, where neither paired data nor\nstyle labels are utilized, across four large-scale datasets. Code is available\nat https://github.com/hanqi-qi/Matte.git\n", "date": "2024-02-23", "categories": []} | |
| {"id": "2303.02186", "title": "Causal Deep Learning", "abstract": " Causality has the potential to truly transform the way we solve a large\nnumber of real-world problems. Yet, so far, its potential largely remains to be\nunlocked as causality often requires crucial assumptions which cannot be tested\nin practice. To address this challenge, we propose a new way of thinking about\ncausality -- we call this causal deep learning. Our causal deep learning\nframework spans three dimensions: (1) a structural dimension, which\nincorporates partial yet testable causal knowledge rather than assuming either\ncomplete or no causal knowledge among the variables of interest; (2) a\nparametric dimension, which encompasses parametric forms that capture the type\nof relationships among the variables of interest; and (3) a temporal dimension,\nwhich captures exposure times or how the variables of interest interact\n(possibly causally) over time. Causal deep learning enables us to make progress\non a variety of real-world problems by leveraging partial causal knowledge\n(including independencies among variables) and quantitatively characterising\ncausal relationships among variables of interest (possibly over time). Our\nframework clearly identifies which assumptions are testable and which ones are\nnot, such that the resulting solutions can be judiciously adopted in practice.\nUsing our formulation we can combine or chain together causal representations\nto solve specific problems without losing track of which assumptions are\nrequired to build these solutions, pushing real-world impact in healthcare,\neconomics and business, environmental sciences and education, through causal\ndeep learning.\n", "date": "2023-03-03", "categories": []} | |
| {"id": "2310.15450", "title": "General Identifiability and Achievability for Causal Representation\n Learning", "abstract": " This paper focuses on causal representation learning (CRL) under a general\nnonparametric latent causal model and a general transformation model that maps\nthe latent data to the observational data. It establishes identifiability and\nachievability results using two hard uncoupled interventions per node in the\nlatent causal graph. Notably, one does not know which pair of intervention\nenvironments have the same node intervened (hence, uncoupled). For\nidentifiability, the paper establishes that perfect recovery of the latent\ncausal model and variables is guaranteed under uncoupled interventions. For\nachievability, an algorithm is designed that uses observational and\ninterventional data and recovers the latent causal model and variables with\nprovable guarantees. This algorithm leverages score variations across different\nenvironments to estimate the inverse of the transformer and, subsequently, the\nlatent variables. The analysis, additionally, recovers the identifiability\nresult for two hard coupled interventions, that is when metadata about the pair\nof environments that have the same node intervened is known. This paper also\nshows that when observational data is available, additional faithfulness\nassumptions that are adopted by the existing literature are unnecessary.\n", "date": "2023-10-24", "categories": []} | |
| {"id": "2303.10931", "title": "Approaching an unknown communication system by latent space exploration\n and causal inference", "abstract": " This paper proposes a methodology for discovering meaningful properties in\ndata by exploring the latent space of unsupervised deep generative models. We\ncombine manipulation of individual latent variables to extreme values with\nmethods inspired by causal inference into an approach we call causal\ndisentanglement with extreme values (CDEV) and show that this method yields\ninsights for model interpretability. With this, we can test for what properties\nof unknown data the model encodes as meaningful, using it to glean insight into\nthe communication system of sperm whales (Physeter macrocephalus), one of the\nmost intriguing and understudied animal communication systems. The network\narchitecture used has been shown to learn meaningful representations of speech;\nhere, it is used as a learning mechanism to decipher the properties of another\nvocal communication system in which case we have no ground truth. The proposed\nmethodology suggests that sperm whales encode information using the number of\nclicks in a sequence, the regularity of their timing, and audio properties such\nas the spectral mean and the acoustic regularity of the sequences. Some of\nthese findings are consistent with existing hypotheses, while others are\nproposed for the first time. We also argue that our models uncover rules that\ngovern the structure of units in the communication system and apply them while\ngenerating innovative data not shown during training. This paper suggests that\nan interpretation of the outputs of deep neural networks with causal inference\nmethodology can be a viable strategy for approaching data about which little is\nknown and presents another case of how deep learning can limit the hypothesis\nspace. Finally, the proposed approach can be extended to other architectures\nand datasets.\n", "date": "2023-03-20", "categories": []} | |
| {"id": "2311.12267", "title": "Learning Causal Representations from General Environments:\n Identifiability and Intrinsic Ambiguity", "abstract": " We study causal representation learning, the task of recovering high-level\nlatent variables and their causal relationships in the form of a causal graph\nfrom low-level observed data (such as text and images), assuming access to\nobservations generated from multiple environments. Prior results on the\nidentifiability of causal representations typically assume access to\nsingle-node interventions which is rather unrealistic in practice, since the\nlatent variables are unknown in the first place. In this work, we provide the\nfirst identifiability results based on data that stem from general\nenvironments. We show that for linear causal models, while the causal graph can\nbe fully recovered, the latent variables are only identified up to the\nsurrounded-node ambiguity (SNA) \\citep{varici2023score}. We provide a\ncounterpart of our guarantee, showing that SNA is basically unavoidable in our\nsetting. We also propose an algorithm, \\texttt{LiNGCReL} which provably\nrecovers the ground-truth model up to SNA, and we demonstrate its effectiveness\nvia numerical experiments. Finally, we consider general non-parametric causal\nmodels and show that the same identification barrier holds when assuming access\nto groups of soft single-node interventions.\n", "date": "2023-11-21", "categories": []} | |
| {"id": "2305.17225", "title": "Causal Component Analysis", "abstract": " Independent Component Analysis (ICA) aims to recover independent latent\nvariables from observed mixtures thereof. Causal Representation Learning (CRL)\naims instead to infer causally related (thus often statistically dependent)\nlatent variables, together with the unknown graph encoding their causal\nrelationships. We introduce an intermediate problem termed Causal Component\nAnalysis (CauCA). CauCA can be viewed as a generalization of ICA, modelling the\ncausal dependence among the latent components, and as a special case of CRL. In\ncontrast to CRL, it presupposes knowledge of the causal graph, focusing solely\non learning the unmixing function and the causal mechanisms. Any impossibility\nresults regarding the recovery of the ground truth in CauCA also apply for CRL,\nwhile possibility results may serve as a stepping stone for extensions to CRL.\nWe characterize CauCA identifiability from multiple datasets generated through\ndifferent types of interventions on the latent causal variables. As a\ncorollary, this interventional perspective also leads to new identifiability\nresults for nonlinear ICA -- a special case of CauCA with an empty graph --\nrequiring strictly fewer datasets than previous results. We introduce a\nlikelihood-based approach using normalizing flows to estimate both the unmixing\nfunction and the causal mechanisms, and demonstrate its effectiveness through\nextensive synthetic experiments in the CauCA and ICA setting.\n", "date": "2023-05-26", "categories": []} | |
| {"id": "2401.08221", "title": "Towards Causal Relationship in Indefinite Data: Baseline Model and New\n Datasets", "abstract": " Integrating deep learning and causal discovery has encouraged us to spot that\nlearning causal structures and representations in dialogue and video is full of\nchallenges. We defined These data forms as \"Indefinite Data\", characterized by\nmulti-structure data and multi-value representations. Unlike existing adaptable\ndata forms, Indefinite Data still faces gaps in datasets and methods. To\naddress the dataset gap, we release two high-quality datasets - Causalogue and\nCausaction, containing text dialogue samples and video action samples with\ncausal annotations respectively. Moreover, the method gap arises from the\ncoexistence of multi-structure data and multi-value representations, breaking\nthe assumptions of all current methods and rendering them infeasible on\nIndefinite Data. To this end, we propose a probabilistic framework as a\nbaseline, incorporating three designed highlights for this gap: 1) establishing\nCausation Condition of representations using the independence of noise terms\nunder non-fixed causal structures, 2) treating causal strength as a latent\nvariable and measuring the reconstruction loss in the correlation space, and 3)\nestimating the effects of latent confounders. These highpoints make the\nprobabilistic model capable of overcoming challenges brought by the coexistence\nof multi-structure data and multi-value representations and pave the way for\nthe extension of latent confounders. Comprehensive experiments have evaluated\nbaseline results of causal structures, causal representations, and confounding\ndisentanglement.\n", "date": "2024-01-16", "categories": []} | |
| {"id": "2401.04890", "title": "Nonparametric Partial Disentanglement via Mechanism Sparsity: Sparse\n Actions, Interventions and Sparse Temporal Dependencies", "abstract": " This work introduces a novel principle for disentanglement we call mechanism\nsparsity regularization, which applies when the latent factors of interest\ndepend sparsely on observed auxiliary variables and/or past latent factors. We\npropose a representation learning method that induces disentanglement by\nsimultaneously learning the latent factors and the sparse causal graphical\nmodel that explains them. We develop a nonparametric identifiability theory\nthat formalizes this principle and shows that the latent factors can be\nrecovered by regularizing the learned causal graph to be sparse. More\nprecisely, we show identifiablity up to a novel equivalence relation we call\n\"consistency\", which allows some latent factors to remain entangled (hence the\nterm partial disentanglement). To describe the structure of this entanglement,\nwe introduce the notions of entanglement graphs and graph preserving functions.\nWe further provide a graphical criterion which guarantees complete\ndisentanglement, that is identifiability up to permutations and element-wise\ntransformations. We demonstrate the scope of the mechanism sparsity principle\nas well as the assumptions it relies on with several worked out examples. For\ninstance, the framework shows how one can leverage multi-node interventions\nwith unknown targets on the latent factors to disentangle them. We further draw\nconnections between our nonparametric results and the now popular exponential\nfamily assumption. Lastly, we propose an estimation procedure based on\nvariational autoencoders and a sparsity constraint and demonstrate it on\nvarious synthetic datasets. This work is meant to be a significantly extended\nversion of Lachapelle et al. (2022).\n", "date": "2024-01-10", "categories": []} | |
| {"id": "2212.09897", "title": "Inducing Character-level Structure in Subword-based Language Models with\n Type-level Interchange Intervention Training", "abstract": " Language tasks involving character-level manipulations (e.g., spelling\ncorrections, arithmetic operations, word games) are challenging for models\noperating on subword units. To address this, we develop a causal intervention\nframework to learn robust and interpretable character representations inside\nsubword-based language models. Our method treats each character as a typed\nvariable in a causal model and learns such causal structures by adapting the\ninterchange intervention training method of Geiger et al. (2021). We\nadditionally introduce a suite of character-level tasks that systematically\nvary in their dependence on meaning and sequence-level context. While\ncharacter-level models still perform best on purely form-based tasks like\nstring reversal, our method outperforms character-level models on more complex\ntasks that blend form, meaning, and context, such as spelling correction in\ncontext and word search games. Compared with standard subword-based models, our\napproach also significantly improves robustness on unseen token sequences and\nleads to human-interpretable internal representations of characters.\n", "date": "2022-12-19", "categories": []} | |
| {"id": "2306.02235", "title": "Learning Linear Causal Representations from Interventions under General\n Nonlinear Mixing", "abstract": " We study the problem of learning causal representations from unknown, latent\ninterventions in a general setting, where the latent distribution is Gaussian\nbut the mixing function is completely general. We prove strong identifiability\nresults given unknown single-node interventions, i.e., without having access to\nthe intervention targets. This generalizes prior works which have focused on\nweaker classes, such as linear maps or paired counterfactual data. This is also\nthe first instance of causal identifiability from non-paired interventions for\ndeep neural network embeddings. Our proof relies on carefully uncovering the\nhigh-dimensional geometric structure present in the data distribution after a\nnon-linear density transformation, which we capture by analyzing quadratic\nforms of precision matrices of the latent distributions. Finally, we propose a\ncontrastive algorithm to identify the latent variables in practice and evaluate\nits performance on various tasks.\n", "date": "2023-06-04", "categories": []} | |
| {"id": "2312.08107", "title": "Causal Optimal Transport of Abstractions", "abstract": " Causal abstraction (CA) theory establishes formal criteria for relating\nmultiple structural causal models (SCMs) at different levels of granularity by\ndefining maps between them. These maps have significant relevance for\nreal-world challenges such as synthesizing causal evidence from multiple\nexperimental environments, learning causally consistent representations at\ndifferent resolutions, and linking interventions across multiple SCMs. In this\nwork, we propose COTA, the first method to learn abstraction maps from\nobservational and interventional data without assuming complete knowledge of\nthe underlying SCMs. In particular, we introduce a multi-marginal Optimal\nTransport (OT) formulation that enforces do-calculus causal constraints,\ntogether with a cost function that relies on interventional information. We\nextensively evaluate COTA on synthetic and real world problems, and showcase\nits advantages over non-causal, independent and aggregated COTA formulations.\nFinally, we demonstrate the efficiency of our method as a data augmentation\ntool by comparing it against the state-of-the-art CA learning framework, which\nassumes fully specified SCMs, on a real-world downstream task.\n", "date": "2023-12-13", "categories": []} | |
| {"id": "2310.02854", "title": "Multi-Domain Causal Representation Learning via Weak Distributional\n Invariances", "abstract": " Causal representation learning has emerged as the center of action in causal\nmachine learning research. In particular, multi-domain datasets present a\nnatural opportunity for showcasing the advantages of causal representation\nlearning over standard unsupervised representation learning. While recent works\nhave taken crucial steps towards learning causal representations, they often\nlack applicability to multi-domain datasets due to over-simplifying assumptions\nabout the data; e.g. each domain comes from a different single-node perfect\nintervention. In this work, we relax these assumptions and capitalize on the\nfollowing observation: there often exists a subset of latents whose certain\ndistributional properties (e.g., support, variance) remain stable across\ndomains; this property holds when, for example, each domain comes from a\nmulti-node imperfect intervention. Leveraging this observation, we show that\nautoencoders that incorporate such invariances can provably identify the stable\nset of latents from the rest across different settings.\n", "date": "2023-10-04", "categories": []} | |
| {"id": "2306.05415", "title": "Causal normalizing flows: from theory to practice", "abstract": " In this work, we deepen on the use of normalizing flows for causal reasoning.\nSpecifically, we first leverage recent results on non-linear ICA to show that\ncausal models are identifiable from observational data given a causal ordering,\nand thus can be recovered using autoregressive normalizing flows (NFs). Second,\nwe analyze different design and learning choices for causal normalizing flows\nto capture the underlying causal data-generating process. Third, we describe\nhow to implement the do-operator in causal NFs, and thus, how to answer\ninterventional and counterfactual questions. Finally, in our experiments, we\nvalidate our design and training choices through a comprehensive ablation\nstudy; compare causal NFs to other approaches for approximating causal models;\nand empirically demonstrate that causal NFs can be used to address real-world\nproblems, where the presence of mixed discrete-continuous data and partial\nknowledge on the causal graph is the norm. The code for this work can be found\nat https://github.com/psanch21/causal-flows.\n", "date": "2023-06-08", "categories": []} | |
| {"id": "2312.02858", "title": "Towards Causal Representations of Climate Model Data", "abstract": " Climate models, such as Earth system models (ESMs), are crucial for\nsimulating future climate change based on projected Shared Socioeconomic\nPathways (SSP) greenhouse gas emissions scenarios. While ESMs are sophisticated\nand invaluable, machine learning-based emulators trained on existing simulation\ndata can project additional climate scenarios much faster and are\ncomputationally efficient. However, they often lack generalizability and\ninterpretability. This work delves into the potential of causal representation\nlearning, specifically the \\emph{Causal Discovery with Single-parent Decoding}\n(CDSD) method, which could render climate model emulation efficient\n\\textit{and} interpretable. We evaluate CDSD on multiple climate datasets,\nfocusing on emissions, temperature, and precipitation. Our findings shed light\non the challenges, limitations, and promise of using CDSD as a stepping stone\ntowards more interpretable and robust climate model emulation.\n", "date": "2023-12-05", "categories": []} | |
| {"id": "2312.03580", "title": "Invariance & Causal Representation Learning: Prospects and Limitations", "abstract": " In causal models, a given mechanism is assumed to be invariant to changes of\nother mechanisms. While this principle has been utilized for inference in\nsettings where the causal variables are observed, theoretical insights when the\nvariables of interest are latent are largely missing. We assay the connection\nbetween invariance and causal representation learning by establishing\nimpossibility results which show that invariance alone is insufficient to\nidentify latent causal variables. Together with practical considerations, we\nuse these theoretical findings to highlight the need for additional constraints\nin order to identify representations by exploiting invariance.\n", "date": "2023-12-06", "categories": []} | |
| {"id": "2311.06567", "title": "SCADI: Self-supervised Causal Disentanglement in Latent Variable Models", "abstract": " Causal disentanglement has great potential for capturing complex situations.\nHowever, there is a lack of practical and efficient approaches. It is already\nknown that most unsupervised disentangling methods are unable to produce\nidentifiable results without additional information, often leading to randomly\ndisentangled output. Therefore, most existing models for disentangling are\nweakly supervised, providing information about intrinsic factors, which incurs\nexcessive costs. Therefore, we propose a novel model, SCADI(SElf-supervised\nCAusal DIsentanglement), that enables the model to discover semantic factors\nand learn their causal relationships without any supervision. This model\ncombines a masked structural causal model (SCM) with a pseudo-label generator\nfor causal disentanglement, aiming to provide a new direction for\nself-supervised causal disentanglement models.\n", "date": "2023-11-11", "categories": []} | |
| {"id": "2307.06250", "title": "Identifiability Guarantees for Causal Disentanglement from Soft\n Interventions", "abstract": " Causal disentanglement aims to uncover a representation of data using latent\nvariables that are interrelated through a causal model. Such a representation\nis identifiable if the latent model that explains the data is unique. In this\npaper, we focus on the scenario where unpaired observational and interventional\ndata are available, with each intervention changing the mechanism of a latent\nvariable. When the causal variables are fully observed, statistically\nconsistent algorithms have been developed to identify the causal model under\nfaithfulness assumptions. We here show that identifiability can still be\nachieved with unobserved causal variables, given a generalized notion of\nfaithfulness. Our results guarantee that we can recover the latent causal model\nup to an equivalence class and predict the effect of unseen combinations of\ninterventions, in the limit of infinite data. We implement our causal\ndisentanglement framework by developing an autoencoding variational Bayes\nalgorithm and apply it to the problem of predicting combinatorial perturbation\neffects in genomics.\n", "date": "2023-07-12", "categories": []} | |
| {"id": "2311.05421", "title": "Diffusion Based Causal Representation Learning", "abstract": " Causal reasoning can be considered a cornerstone of intelligent systems.\nHaving access to an underlying causal graph comes with the promise of\ncause-effect estimation and the identification of efficient and safe\ninterventions. However, learning causal representations remains a major\nchallenge, due to the complexity of many real-world systems. Previous works on\ncausal representation learning have mostly focused on Variational Auto-Encoders\n(VAE). These methods only provide representations from a point estimate, and\nthey are unsuitable to handle high dimensions. To overcome these problems, we\nproposed a new Diffusion-based Causal Representation Learning (DCRL) algorithm.\nThis algorithm uses diffusion-based representations for causal discovery. DCRL\noffers access to infinite dimensional latent codes, which encode different\nlevels of information in the latent code. In a first proof of principle, we\ninvestigate the use of DCRL for causal representation learning. We further\ndemonstrate experimentally that this approach performs comparably well in\nidentifying the causal structure and causal variables.\n", "date": "2023-11-09", "categories": []} | |
| {"id": "2310.18471", "title": "Causal disentanglement of multimodal data", "abstract": " Causal representation learning algorithms discover lower-dimensional\nrepresentations of data that admit a decipherable interpretation of cause and\neffect; as achieving such interpretable representations is challenging, many\ncausal learning algorithms utilize elements indicating prior information, such\nas (linear) structural causal models, interventional data, or weak supervision.\nUnfortunately, in exploratory causal representation learning, such elements and\nprior information may not be available or warranted. Alternatively, scientific\ndatasets often have multiple modalities or physics-based constraints, and the\nuse of such scientific, multimodal data has been shown to improve\ndisentanglement in fully unsupervised settings. Consequently, we introduce a\ncausal representation learning algorithm (causalPIMA) that can use multimodal\ndata and known physics to discover important features with causal\nrelationships. Our innovative algorithm utilizes a new differentiable\nparametrization to learn a directed acyclic graph (DAG) together with a latent\nspace of a variational autoencoder in an end-to-end differentiable framework\nvia a single, tractable evidence lower bound loss function. We place a Gaussian\nmixture prior on the latent space and identify each of the mixtures with an\noutcome of the DAG nodes; this novel identification enables feature discovery\nwith causal relationships. Tested against a synthetic and a scientific dataset,\nour results demonstrate the capability of learning an interpretable causal\nstructure while simultaneously discovering key features in a fully unsupervised\nsetting.\n", "date": "2023-10-27", "categories": []} | |
| {"id": "2306.02899", "title": "Learning nonparametric latent causal graphs with unknown interventions", "abstract": " We establish conditions under which latent causal graphs are\nnonparametrically identifiable and can be reconstructed from unknown\ninterventions in the latent space. Our primary focus is the identification of\nthe latent structure in measurement models without parametric assumptions such\nas linearity or Gaussianity. Moreover, we do not assume the number of hidden\nvariables is known, and we show that at most one unknown intervention per\nhidden variable is needed. This extends a recent line of work on learning\ncausal representations from observations and interventions. The proofs are\nconstructive and introduce two new graphical concepts -- imaginary subsets and\nisolated edges -- that may be useful in their own right. As a matter of\nindependent interest, the proofs also involve a novel characterization of the\nlimits of edge orientations within the equivalence class of DAGs induced by\nunknown interventions. These are the first results to characterize the\nconditions under which causal representations are identifiable without making\nany parametric assumptions in a general setting with unknown interventions and\nwithout faithfulness.\n", "date": "2023-06-05", "categories": []} | |
| {"id": "2311.00923", "title": "A Review and Roadmap of Deep Causal Model from Different Causal\n Structures and Representations", "abstract": " The fusion of causal models with deep learning introducing increasingly\nintricate data sets, such as the causal associations within images or between\ntextual components, has surfaced as a focal research area. Nonetheless, the\nbroadening of original causal concepts and theories to such complex,\nnon-statistical data has been met with serious challenges. In response, our\nstudy proposes redefinitions of causal data into three distinct categories from\nthe standpoint of causal structure and representation: definite data,\nsemi-definite data, and indefinite data. Definite data chiefly pertains to\nstatistical data used in conventional causal scenarios, while semi-definite\ndata refers to a spectrum of data formats germane to deep learning, including\ntime-series, images, text, and others. Indefinite data is an emergent research\nsphere inferred from the progression of data forms by us. To comprehensively\npresent these three data paradigms, we elaborate on their formal definitions,\ndifferences manifested in datasets, resolution pathways, and development of\nresearch. We summarize key tasks and achievements pertaining to definite and\nsemi-definite data from myriad research undertakings, present a roadmap for\nindefinite data, beginning with its current research conundrums. Lastly, we\nclassify and scrutinize the key datasets presently utilized within these three\nparadigms.\n", "date": "2023-11-02", "categories": []} | |
| {"id": "2306.07916", "title": "Identification of Nonlinear Latent Hierarchical Models", "abstract": " Identifying latent variables and causal structures from observational data is\nessential to many real-world applications involving biological data, medical\ndata, and unstructured data such as images and languages. However, this task\ncan be highly challenging, especially when observed variables are generated by\ncausally related latent variables and the relationships are nonlinear. In this\nwork, we investigate the identification problem for nonlinear latent\nhierarchical causal models in which observed variables are generated by a set\nof causally related latent variables, and some latent variables may not have\nobserved children.\n We show that the identifiability of causal structures and latent variables\n(up to invertible transformations) can be achieved under mild assumptions: on\ncausal structures, we allow for multiple paths between any pair of variables in\nthe graph, which relaxes latent tree assumptions in prior work; on structural\nfunctions, we permit general nonlinearity and multi-dimensional continuous\nvariables, alleviating existing work's parametric assumptions. Specifically, we\nfirst develop an identification criterion in the form of novel identifiability\nguarantees for an elementary latent variable model. Leveraging this criterion,\nwe show that both causal structures and latent variables of the hierarchical\nmodel can be identified asymptotically by explicitly constructing an estimation\nprocedure. To the best of our knowledge, our work is the first to establish\nidentifiability guarantees for both causal structures and latent variables in\nnonlinear latent hierarchical models.\n", "date": "2023-06-13", "categories": []} | |
| {"id": "2208.10027", "title": "Learning Invariant Representations under General Interventions on the\n Response", "abstract": " It has become increasingly common nowadays to collect observations of feature\nand response pairs from different environments. As a consequence, one has to\napply learned predictors to data with a different distribution due to\ndistribution shifts. One principled approach is to adopt the structural causal\nmodels to describe training and test models, following the invariance principle\nwhich says that the conditional distribution of the response given its\npredictors remains the same across environments. However, this principle might\nbe violated in practical settings when the response is intervened. A natural\nquestion is whether it is still possible to identify other forms of invariance\nto facilitate prediction in unseen environments. To shed light on this\nchallenging scenario, we focus on linear structural causal models (SCMs) and\nintroduce invariant matching property (IMP), an explicit relation to capture\ninterventions through an additional feature, leading to an alternative form of\ninvariance that enables a unified treatment of general interventions on the\nresponse as well as the predictors. We analyze the asymptotic generalization\nerrors of our method under both the discrete and continuous environment\nsettings, where the continuous case is handled by relating it to the\nsemiparametric varying coefficient models. We present algorithms that show\ncompetitive performance compared to existing methods over various experimental\nsettings including a COVID dataset.\n", "date": "2022-08-22", "categories": []} | |
| {"id": "2306.00542", "title": "Nonparametric Identifiability of Causal Representations from Unknown\n Interventions", "abstract": " We study causal representation learning, the task of inferring latent causal\nvariables and their causal relations from high-dimensional mixtures of the\nvariables. Prior work relies on weak supervision, in the form of counterfactual\npre- and post-intervention views or temporal structure; places restrictive\nassumptions, such as linearity, on the mixing function or latent causal model;\nor requires partial knowledge of the generative process, such as the causal\ngraph or intervention targets. We instead consider the general setting in which\nboth the causal model and the mixing function are nonparametric. The learning\nsignal takes the form of multiple datasets, or environments, arising from\nunknown interventions in the underlying causal model. Our goal is to identify\nboth the ground truth latents and their causal graph up to a set of ambiguities\nwhich we show to be irresolvable from interventional data. We study the\nfundamental setting of two causal variables and prove that the observational\ndistribution and one perfect intervention per node suffice for identifiability,\nsubject to a genericity condition. This condition rules out spurious solutions\nthat involve fine-tuning of the intervened and observational distributions,\nmirroring similar conditions for nonlinear cause-effect inference. For an\narbitrary number of variables, we show that at least one pair of distinct\nperfect interventional domains per node guarantees identifiability. Further, we\ndemonstrate that the strengths of causal influences among the latent variables\nare preserved by all equivalent solutions, rendering the inferred\nrepresentation appropriate for drawing causal conclusions from new data. Our\nstudy provides the first identifiability results for the general nonparametric\nsetting with unknown interventions, and elucidates what is possible and\nimpossible for causal representation learning without more direct supervision.\n", "date": "2023-06-01", "categories": []} | |
| {"id": "2310.19054", "title": "Object-centric architectures enable efficient causal representation\n learning", "abstract": " Causal representation learning has showed a variety of settings in which we\ncan disentangle latent variables with identifiability guarantees (up to some\nreasonable equivalence class). Common to all of these approaches is the\nassumption that (1) the latent variables are represented as $d$-dimensional\nvectors, and (2) that the observations are the output of some injective\ngenerative function of these latent variables. While these assumptions appear\nbenign, we show that when the observations are of multiple objects, the\ngenerative function is no longer injective and disentanglement fails in\npractice. We can address this failure by combining recent developments in\nobject-centric learning and causal representation learning. By modifying the\nSlot Attention architecture arXiv:2006.15055, we develop an object-centric\narchitecture that leverages weak supervision from sparse perturbations to\ndisentangle each object's properties. This approach is more data-efficient in\nthe sense that it requires significantly fewer perturbations than a comparable\napproach that encodes to a Euclidean space and we show that this approach\nsuccessfully disentangles the properties of a set of objects in a series of\nsimple image-based disentanglement experiments.\n", "date": "2023-10-29", "categories": []} | |
| {"id": "2302.00993", "title": "Unpaired Multi-Domain Causal Representation Learning", "abstract": " The goal of causal representation learning is to find a representation of\ndata that consists of causally related latent variables. We consider a setup\nwhere one has access to data from multiple domains that potentially share a\ncausal representation. Crucially, observations in different domains are assumed\nto be unpaired, that is, we only observe the marginal distribution in each\ndomain but not their joint distribution. In this paper, we give sufficient\nconditions for identifiability of the joint distribution and the shared causal\ngraph in a linear setup. Identifiability holds if we can uniquely recover the\njoint distribution and the shared causal representation from the marginal\ndistributions in each domain. We transform our identifiability results into a\npractical method to recover the shared latent causal graph.\n", "date": "2023-02-02", "categories": []} | |
| {"id": "2310.17325", "title": "C-Disentanglement: Discovering Causally-Independent Generative Factors\n under an Inductive Bias of Confounder", "abstract": " Representation learning assumes that real-world data is generated by a few\nsemantically meaningful generative factors (i.e., sources of variation) and\naims to discover them in the latent space. These factors are expected to be\ncausally disentangled, meaning that distinct factors are encoded into separate\nlatent variables, and changes in one factor will not affect the values of the\nothers. Compared to statistical independence, causal disentanglement allows\nmore controllable data generation, improved robustness, and better\ngeneralization. However, most existing work assumes unconfoundedness in the\ndiscovery process, that there are no common causes to the generative factors\nand thus obtain only statistical independence. In this paper, we recognize the\nimportance of modeling confounders in discovering causal generative factors.\nUnfortunately, such factors are not identifiable without proper inductive bias.\nWe fill the gap by introducing a framework entitled Confounded-Disentanglement\n(C-Disentanglement), the first framework that explicitly introduces the\ninductive bias of confounder via labels from domain expertise. In addition, we\naccordingly propose an approach to sufficiently identify the causally\ndisentangled factors under any inductive bias of the confounder. We conduct\nextensive experiments on both synthetic and real-world datasets. Our method\ndemonstrates competitive results compared to various SOTA baselines in\nobtaining causally disentangled features and downstream tasks under domain\nshifts.\n", "date": "2023-10-26", "categories": []} | |
| {"id": "2310.13892", "title": "Specify Robust Causal Representation from Mixed Observations", "abstract": " Learning representations purely from observations concerns the problem of\nlearning a low-dimensional, compact representation which is beneficial to\nprediction models. Under the hypothesis that the intrinsic latent factors\nfollow some casual generative models, we argue that by learning a causal\nrepresentation, which is the minimal sufficient causes of the whole system, we\ncan improve the robustness and generalization performance of machine learning\nmodels. In this paper, we develop a learning method to learn such\nrepresentation from observational data by regularizing the learning procedure\nwith mutual information measures, according to the hypothetical factored causal\ngraph. We theoretically and empirically show that the models trained with the\nlearned causal representations are more robust under adversarial attacks and\ndistribution shifts compared with baselines. The supplementary materials are\navailable at https://github.com/ymy $4323460 / \\mathrm{CaRI} /$.\n", "date": "2023-10-21", "categories": []} | |
| {"id": "2302.11737", "title": "Causally Disentangled Generative Variational AutoEncoder", "abstract": " We present a new supervised learning technique for the Variational\nAutoEncoder (VAE) that allows it to learn a causally disentangled\nrepresentation and generate causally disentangled outcomes simultaneously. We\ncall this approach Causally Disentangled Generation (CDG). CDG is a generative\nmodel that accurately decodes an output based on a causally disentangled\nrepresentation. Our research demonstrates that adding supervised regularization\nto the encoder alone is insufficient for achieving a generative model with CDG,\neven for a simple task. Therefore, we explore the necessary and sufficient\nconditions for achieving CDG within a specific model. Additionally, we\nintroduce a universal metric for evaluating the causal disentanglement of a\ngenerative model. Empirical results from both image and tabular datasets\nsupport our findings.\n", "date": "2023-02-23", "categories": []} | |
| {"id": "2211.08573", "title": "Realization of Causal Representation Learning to Adjust Confounding Bias\n in Latent Space", "abstract": " Causal DAGs(Directed Acyclic Graphs) are usually considered in a 2D plane.\nEdges indicate causal effects' directions and imply their corresponding\ntime-passings. Due to the natural restriction of statistical models, effect\nestimation is usually approximated by averaging the individuals' correlations,\ni.e., observational changes over a specific time. However, in the context of\nMachine Learning on large-scale questions with complex DAGs, such slight biases\ncan snowball to distort global models - More importantly, it has practically\nimpeded the development of AI, for instance, the weak generalizability of\ncausal models. In this paper, we redefine causal DAG as \\emph{do-DAG}, in which\nvariables' values are no longer time-stamp-dependent, and timelines can be seen\nas axes. By geometric explanation of multi-dimensional do-DAG, we identify the\n\\emph{Causal Representation Bias} and its necessary factors, differentiated\nfrom common confounding biases. Accordingly, a DL(Deep Learning)-based\nframework will be proposed as the general solution, along with a realization\nmethod and experiments to verify its feasibility.\n", "date": "2022-11-15", "categories": []} | |
| {"id": "2302.00869", "title": "Disentanglement of Latent Representations via Causal Interventions", "abstract": " The process of generating data such as images is controlled by independent\nand unknown factors of variation. The retrieval of these variables has been\nstudied extensively in the disentanglement, causal representation learning, and\nindependent component analysis fields. Recently, approaches merging these\ndomains together have shown great success. Instead of directly representing the\nfactors of variation, the problem of disentanglement can be seen as finding the\ninterventions on one image that yield a change to a single factor. Following\nthis assumption, we introduce a new method for disentanglement inspired by\ncausal dynamics that combines causality theory with vector-quantized\nvariational autoencoders. Our model considers the quantized vectors as causal\nvariables and links them in a causal graph. It performs causal interventions on\nthe graph and generates atomic transitions affecting a unique factor of\nvariation in the image. We also introduce a new task of action retrieval that\nconsists of finding the action responsible for the transition between two\nimages. We test our method on standard synthetic and real-world disentanglement\ndatasets. We show that it can effectively disentangle the factors of variation\nand perform precise interventions on high-level semantic attributes of an image\nwithout affecting its quality, even with imbalanced data distributions.\n", "date": "2023-02-02", "categories": []} | |
| {"id": "2308.05707", "title": "Shadow Datasets, New challenging datasets for Causal Representation\n Learning", "abstract": " Discovering causal relations among semantic factors is an emergent topic in\nrepresentation learning. Most causal representation learning (CRL) methods are\nfully supervised, which is impractical due to costly labeling. To resolve this\nrestriction, weakly supervised CRL methods were introduced. To evaluate CRL\nperformance, four existing datasets, Pendulum, Flow, CelebA(BEARD) and\nCelebA(SMILE), are utilized. However, existing CRL datasets are limited to\nsimple graphs with few generative factors. Thus we propose two new datasets\nwith a larger number of diverse generative factors and more sophisticated\ncausal graphs. In addition, current real datasets, CelebA(BEARD) and\nCelebA(SMILE), the originally proposed causal graphs are not aligned with the\ndataset distributions. Thus, we propose modifications to them.\n", "date": "2023-08-10", "categories": []} | |
| {"id": "2305.02640", "title": "Towards Causal Representation Learning and Deconfounding from Indefinite\n Data", "abstract": " Owing to the cross-pollination between causal discovery and deep learning,\nnon-statistical data (e.g., images, text, etc.) encounters significant\nconflicts in terms of properties and methods with traditional causal data. To\nunify these data types of varying forms, we redefine causal data from two novel\nperspectives and then propose three data paradigms. Among them, the indefinite\ndata (like dialogues or video sources) induce low sample utilization and\nincapability of the distribution assumption, both leading to the fact that\nlearning causal representation from indefinite data is, as of yet, largely\nunexplored. We design the causal strength variational model to settle down\nthese two problems. Specifically, we leverage the causal strength instead of\nindependent noise as the latent variable to construct evidence lower bound. By\nthis design ethos, The causal strengths of different structures are regarded as\na distribution and can be expressed as a 2D matrix. Moreover, considering the\nlatent confounders, we disentangle the causal graph G into two relation\nsubgraphs O and C. O contains pure relations between observed variables, while\nC represents the relations from latent variables to observed variables. We\nimplement the above designs as a dynamic variational inference model, tailored\nto learn causal representation from indefinite data under latent confounding.\nFinally, we conduct comprehensive experiments on synthetic and real-world data\nto demonstrate the effectiveness of our method.\n", "date": "2023-05-04", "categories": []} | |
| {"id": "2307.05704", "title": "A Causal Ordering Prior for Unsupervised Representation Learning", "abstract": " Unsupervised representation learning with variational inference relies\nheavily on independence assumptions over latent variables. Causal\nrepresentation learning (CRL), however, argues that factors of variation in a\ndataset are, in fact, causally related. Allowing latent variables to be\ncorrelated, as a consequence of causal relationships, is more realistic and\ngeneralisable. So far, provably identifiable methods rely on: auxiliary\ninformation, weak labels, and interventional or even counterfactual data.\nInspired by causal discovery with functional causal models, we propose a fully\nunsupervised representation learning method that considers a data generation\nprocess with a latent additive noise model (ANM). We encourage the latent space\nto follow a causal ordering via loss function based on the Hessian of the\nlatent distribution.\n", "date": "2023-07-11", "categories": []} | |
| {"id": "2306.09643", "title": "BISCUIT: Causal Representation Learning from Binary Interactions", "abstract": " Identifying the causal variables of an environment and how to intervene on\nthem is of core value in applications such as robotics and embodied AI. While\nan agent can commonly interact with the environment and may implicitly perturb\nthe behavior of some of these causal variables, often the targets it affects\nremain unknown. In this paper, we show that causal variables can still be\nidentified for many common setups, e.g., additive Gaussian noise models, if the\nagent's interactions with a causal variable can be described by an unknown\nbinary variable. This happens when each causal variable has two different\nmechanisms, e.g., an observational and an interventional one. Using this\nidentifiability result, we propose BISCUIT, a method for simultaneously\nlearning causal variables and their corresponding binary interaction variables.\nOn three robotic-inspired datasets, BISCUIT accurately identifies causal\nvariables and can even be scaled to complex, realistic environments for\nembodied AI.\n", "date": "2023-06-16", "categories": []} | |
| {"id": "2306.09302", "title": "Knowledge Guided Representation Learning and Causal Structure Learning\n in Soil Science", "abstract": " An improved understanding of soil can enable more sustainable land-use\npractices. Nevertheless, soil is called a complex, living medium due to the\ncomplex interaction of different soil processes that limit our understanding of\nsoil. Process-based models and analyzing observed data provide two avenues for\nimproving our understanding of soil processes. Collecting observed data is\ncost-prohibitive but reflects real-world behavior, while process-based models\ncan be used to generate ample synthetic data which may not be representative of\nreality. We propose a framework, knowledge-guided representation learning, and\ncausal structure learning (KGRCL), to accelerate scientific discoveries in soil\nscience. The framework improves representation learning for simulated soil\nprocesses via conditional distribution matching with observed soil processes.\nSimultaneously, the framework leverages both observed and simulated data to\nlearn a causal structure among the soil processes. The learned causal graph is\nmore representative of ground truth than other graphs generated from other\ncausal discovery methods. Furthermore, the learned causal graph is leveraged in\na supervised learning setup to predict the impact of fertilizer use and\nchanging weather on soil carbon. We present the results in five different\nlocations to show the improvement in the prediction performance in\nout-of-sample and few-shots setting.\n", "date": "2023-06-15", "categories": []} | |
| {"id": "2211.16467", "title": "Linear Causal Disentanglement via Interventions", "abstract": " Causal disentanglement seeks a representation of data involving latent\nvariables that relate to one another via a causal model. A representation is\nidentifiable if both the latent model and the transformation from latent to\nobserved variables are unique. In this paper, we study observed variables that\nare a linear transformation of a linear latent causal model. Data from\ninterventions are necessary for identifiability: if one latent variable is\nmissing an intervention, we show that there exist distinct models that cannot\nbe distinguished. Conversely, we show that a single intervention on each latent\nvariable is sufficient for identifiability. Our proof uses a generalization of\nthe RQ decomposition of a matrix that replaces the usual orthogonal and upper\ntriangular conditions with analogues depending on a partial order on the rows\nof the matrix, with partial order determined by a latent causal model. We\ncorroborate our theoretical results with a method for causal disentanglement\nthat accurately recovers a latent causal model.\n", "date": "2022-11-29", "categories": []} | |
| {"id": "2302.02228", "title": "Counterfactual Identifiability of Bijective Causal Models", "abstract": " We study counterfactual identifiability in causal models with bijective\ngeneration mechanisms (BGM), a class that generalizes several widely-used\ncausal models in the literature. We establish their counterfactual\nidentifiability for three common causal structures with unobserved confounding,\nand propose a practical learning method that casts learning a BGM as structured\ngenerative modeling. Learned BGMs enable efficient counterfactual estimation\nand can be obtained using a variety of deep conditional generative models. We\nevaluate our techniques in a visual task and demonstrate its application in a\nreal-world video streaming simulation task.\n", "date": "2023-02-04", "categories": []} | |
| {"id": "2302.05209", "title": "A Survey on Causal Reinforcement Learning", "abstract": " While Reinforcement Learning (RL) achieves tremendous success in sequential\ndecision-making problems of many domains, it still faces key challenges of data\ninefficiency and the lack of interpretability. Interestingly, many researchers\nhave leveraged insights from the causality literature recently, bringing forth\nflourishing works to unify the merits of causality and address well the\nchallenges from RL. As such, it is of great necessity and significance to\ncollate these Causal Reinforcement Learning (CRL) works, offer a review of CRL\nmethods, and investigate the potential functionality from causality toward RL.\nIn particular, we divide existing CRL approaches into two categories according\nto whether their causality-based information is given in advance or not. We\nfurther analyze each category in terms of the formalization of different\nmodels, ranging from the Markov Decision Process (MDP), Partially Observed\nMarkov Decision Process (POMDP), Multi-Arm Bandits (MAB), and Dynamic Treatment\nRegime (DTR). Moreover, we summarize the evaluation matrices and open sources\nwhile we discuss emerging applications, along with promising prospects for the\nfuture development of CRL.\n", "date": "2023-02-10", "categories": []} | |
| {"id": "2305.19802", "title": "Neuro-Causal Factor Analysis", "abstract": " Factor analysis (FA) is a statistical tool for studying how observed\nvariables with some mutual dependences can be expressed as functions of\nmutually independent unobserved factors, and it is widely applied throughout\nthe psychological, biological, and physical sciences. We revisit this classic\nmethod from the comparatively new perspective given by advancements in causal\ndiscovery and deep learning, introducing a framework for Neuro-Causal Factor\nAnalysis (NCFA). Our approach is fully nonparametric: it identifies factors via\nlatent causal discovery methods and then uses a variational autoencoder (VAE)\nthat is constrained to abide by the Markov factorization of the distribution\nwith respect to the learned graph. We evaluate NCFA on real and synthetic data\nsets, finding that it performs comparably to standard VAEs on data\nreconstruction tasks but with the advantages of sparser architecture, lower\nmodel complexity, and causal interpretability. Unlike traditional FA methods,\nour proposed NCFA method allows learning and reasoning about the latent factors\nunderlying observed data from a justifiably causal perspective, even when the\nrelations between factors and measurements are highly nonlinear.\n", "date": "2023-05-31", "categories": []} | |
| {"id": "2209.01321", "title": "Deep Stable Representation Learning on Electronic Health Records", "abstract": " Deep learning models have achieved promising disease prediction performance\nof the Electronic Health Records (EHR) of patients. However, most models\ndeveloped under the I.I.D. hypothesis fail to consider the agnostic\ndistribution shifts, diminishing the generalization ability of deep learning\nmodels to Out-Of-Distribution (OOD) data. In this setting, spurious statistical\ncorrelations that may change in different environments will be exploited, which\ncan cause sub-optimal performances of deep learning models. The unstable\ncorrelation between procedures and diagnoses existed in the training\ndistribution can cause spurious correlation between historical EHR and future\ndiagnosis. To address this problem, we propose to use a causal representation\nlearning method called Causal Healthcare Embedding (CHE). CHE aims at\neliminating the spurious statistical relationship by removing the dependencies\nbetween diagnoses and procedures. We introduce the Hilbert-Schmidt Independence\nCriterion (HSIC) to measure the degree of independence between the embedded\ndiagnosis and procedure features. Based on causal view analyses, we perform the\nsample weighting technique to get rid of such spurious relationship for the\nstable learning of EHR across different environments. Moreover, our proposed\nCHE method can be used as a flexible plug-and-play module that can enhance\nexisting deep learning models on EHR. Extensive experiments on two public\ndatasets and five state-of-the-art baselines unequivocally show that CHE can\nimprove the prediction accuracy of deep learning models on out-of-distribution\ndata by a large margin. In addition, the interpretability study shows that CHE\ncould successfully leverage causal structures to reflect a more reasonable\ncontribution of historical records for predictions.\n", "date": "2022-09-03", "categories": []} | |
| {"id": "2207.09081", "title": "Generalizing Goal-Conditioned Reinforcement Learning with Variational\n Causal Reasoning", "abstract": " As a pivotal component to attaining generalizable solutions in human\nintelligence, reasoning provides great potential for reinforcement learning\n(RL) agents' generalization towards varied goals by summarizing part-to-whole\narguments and discovering cause-and-effect relations. However, how to discover\nand represent causalities remains a huge gap that hinders the development of\ncausal RL. In this paper, we augment Goal-Conditioned RL (GCRL) with Causal\nGraph (CG), a structure built upon the relation between objects and events. We\nnovelly formulate the GCRL problem into variational likelihood maximization\nwith CG as latent variables. To optimize the derived objective, we propose a\nframework with theoretical performance guarantees that alternates between two\nsteps: using interventional data to estimate the posterior of CG; using CG to\nlearn generalizable models and interpretable policies. Due to the lack of\npublic benchmarks that verify generalization capability under reasoning, we\ndesign nine tasks and then empirically show the effectiveness of the proposed\nmethod against five baselines on these tasks. Further theoretical analysis\nshows that our performance improvement is attributed to the virtuous cycle of\ncausal discovery, transition modeling, and policy training, which aligns with\nthe experimental evidence in extensive ablation studies.\n", "date": "2022-07-19", "categories": []} | |
| {"id": "2301.05893", "title": "Jointly Learning Consistent Causal Abstractions Over Multiple\n Interventional Distributions", "abstract": " An abstraction can be used to relate two structural causal models\nrepresenting the same system at different levels of resolution. Learning\nabstractions which guarantee consistency with respect to interventional\ndistributions would allow one to jointly reason about evidence across multiple\nlevels of granularity while respecting the underlying cause-effect\nrelationships. In this paper, we introduce a first framework for causal\nabstraction learning between SCMs based on the formalization of abstraction\nrecently proposed by Rischel (2020). Based on that, we propose a differentiable\nprogramming solution that jointly solves a number of combinatorial\nsub-problems, and we study its performance and benefits against independent and\nsequential approaches on synthetic settings and on a challenging real-world\nproblem related to electric vehicle battery manufacturing.\n", "date": "2023-01-14", "categories": []} | |
| {"id": "2305.04357", "title": "Quantifying Consistency and Information Loss for Causal Abstraction\n Learning", "abstract": " Structural causal models provide a formalism to express causal relations\nbetween variables of interest. Models and variables can represent a system at\ndifferent levels of abstraction, whereby relations may be coarsened and refined\naccording to the need of a modeller. However, switching between different\nlevels of abstraction requires evaluating a trade-off between the consistency\nand the information loss among different models. In this paper we introduce a\nfamily of interventional measures that an agent may use to evaluate such a\ntrade-off. We consider four measures suited for different tasks, analyze their\nproperties, and propose algorithms to evaluate and learn causal abstractions.\nFinally, we illustrate the flexibility of our setup by empirically showing how\ndifferent measures and algorithmic choices may lead to different abstractions.\n", "date": "2023-05-07", "categories": []} | |
| {"id": "2302.02672", "title": "Identifiability of latent-variable and structural-equation models: from\n linear to nonlinear", "abstract": " An old problem in multivariate statistics is that linear Gaussian models are\noften unidentifiable, i.e. some parameters cannot be uniquely estimated. In\nfactor (component) analysis, an orthogonal rotation of the factors is\nunidentifiable, while in linear regression, the direction of effect cannot be\nidentified. For such linear models, non-Gaussianity of the (latent) variables\nhas been shown to provide identifiability. In the case of factor analysis, this\nleads to independent component analysis, while in the case of the direction of\neffect, non-Gaussian versions of structural equation modelling solve the\nproblem. More recently, we have shown how even general nonparametric nonlinear\nversions of such models can be estimated. Non-Gaussianity is not enough in this\ncase, but assuming we have time series, or that the distributions are suitably\nmodulated by some observed auxiliary variables, the models are identifiable.\nThis paper reviews the identifiability theory for the linear and nonlinear\ncases, considering both factor analytic models and structural equation models.\n", "date": "2023-02-06", "categories": []} | |
| {"id": "2301.08230", "title": "Score-based Causal Representation Learning with Interventions", "abstract": " This paper studies the causal representation learning problem when the latent\ncausal variables are observed indirectly through an unknown linear\ntransformation. The objectives are: (i) recovering the unknown linear\ntransformation (up to scaling) and (ii) determining the directed acyclic graph\n(DAG) underlying the latent variables. Sufficient conditions for DAG recovery\nare established, and it is shown that a large class of non-linear models in the\nlatent space (e.g., causal mechanisms parameterized by two-layer neural\nnetworks) satisfy these conditions. These sufficient conditions ensure that the\neffect of an intervention can be detected correctly from changes in the score.\nCapitalizing on this property, recovering a valid transformation is facilitated\nby the following key property: any valid transformation renders latent\nvariables' score function to necessarily have the minimal variations across\ndifferent interventional environments. This property is leveraged for perfect\nrecovery of the latent DAG structure using only \\emph{soft} interventions. For\nthe special case of stochastic \\emph{hard} interventions, with an additional\nhypothesis testing step, one can also uniquely recover the linear\ntransformation up to scaling and a valid causal ordering.\n", "date": "2023-01-19", "categories": []} | |
| {"id": "2302.08635", "title": "Generative Causal Representation Learning for Out-of-Distribution Motion\n Forecasting", "abstract": " Conventional supervised learning methods typically assume i.i.d samples and\nare found to be sensitive to out-of-distribution (OOD) data. We propose\nGenerative Causal Representation Learning (GCRL) which leverages causality to\nfacilitate knowledge transfer under distribution shifts. While we evaluate the\neffectiveness of our proposed method in human trajectory prediction models,\nGCRL can be applied to other domains as well. First, we propose a novel causal\nmodel that explains the generative factors in motion forecasting datasets using\nfeatures that are common across all environments and with features that are\nspecific to each environment. Selection variables are used to determine which\nparts of the model can be directly transferred to a new environment without\nfine-tuning. Second, we propose an end-to-end variational learning paradigm to\nlearn the causal mechanisms that generate observations from features. GCRL is\nsupported by strong theoretical results that imply identifiability of the\ncausal model under certain assumptions. Experimental results on synthetic and\nreal-world motion forecasting datasets show the robustness and effectiveness of\nour proposed method for knowledge transfer under zero-shot and low-shot\nsettings by substantially outperforming the prior motion forecasting models on\nout-of-distribution prediction. Our code is available at\nhttps://github.com/sshirahmad/GCRL.\n", "date": "2023-02-17", "categories": []} | |
| {"id": "2304.07922", "title": "Causal Disentangled Variational Auto-Encoder for Preference\n Understanding in Recommendation", "abstract": " Recommendation models are typically trained on observational user interaction\ndata, but the interactions between latent factors in users' decision-making\nprocesses lead to complex and entangled data. Disentangling these latent\nfactors to uncover their underlying representation can improve the robustness,\ninterpretability, and controllability of recommendation models. This paper\nintroduces the Causal Disentangled Variational Auto-Encoder (CaD-VAE), a novel\napproach for learning causal disentangled representations from interaction data\nin recommender systems. The CaD-VAE method considers the causal relationships\nbetween semantically related factors in real-world recommendation scenarios,\nrather than enforcing independence as in existing disentanglement methods. The\napproach utilizes structural causal models to generate causal representations\nthat describe the causal relationship between latent factors. The results\ndemonstrate that CaD-VAE outperforms existing methods, offering a promising\nsolution for disentangling complex user behavior data in recommendation\nsystems.\n", "date": "2023-04-17", "categories": []} | |
| {"id": "2301.05169", "title": "Causal Triplet: An Open Challenge for Intervention-centric Causal\n Representation Learning", "abstract": " Recent years have seen a surge of interest in learning high-level causal\nrepresentations from low-level image pairs under interventions. Yet, existing\nefforts are largely limited to simple synthetic settings that are far away from\nreal-world problems. In this paper, we present Causal Triplet, a causal\nrepresentation learning benchmark featuring not only visually more complex\nscenes, but also two crucial desiderata commonly overlooked in previous works:\n(i) an actionable counterfactual setting, where only certain object-level\nvariables allow for counterfactual observations whereas others do not; (ii) an\ninterventional downstream task with an emphasis on out-of-distribution\nrobustness from the independent causal mechanisms principle. Through extensive\nexperiments, we find that models built with the knowledge of disentangled or\nobject-centric representations significantly outperform their distributed\ncounterparts. However, recent causal representation learning methods still\nstruggle to identify such latent structures, indicating substantial challenges\nand opportunities for future work. Our code and datasets will be available at\nhttps://sites.google.com/view/causaltriplet.\n", "date": "2023-01-12", "categories": []} | |
| {"id": "2303.17264", "title": "Multifactor Sequential Disentanglement via Structured Koopman\n Autoencoders", "abstract": " Disentangling complex data to its latent factors of variation is a\nfundamental task in representation learning. Existing work on sequential\ndisentanglement mostly provides two factor representations, i.e., it separates\nthe data to time-varying and time-invariant factors. In contrast, we consider\nmultifactor disentanglement in which multiple (more than two) semantic\ndisentangled components are generated. Key to our approach is a strong\ninductive bias where we assume that the underlying dynamics can be represented\nlinearly in the latent space. Under this assumption, it becomes natural to\nexploit the recently introduced Koopman autoencoder models. However,\ndisentangled representations are not guaranteed in Koopman approaches, and thus\nwe propose a novel spectral loss term which leads to structured Koopman\nmatrices and disentanglement. Overall, we propose a simple and easy to code new\ndeep model that is fully unsupervised and it supports multifactor\ndisentanglement. We showcase new disentangling abilities such as swapping of\nindividual static factors between characters, and an incremental swap of\ndisentangled factors from the source to the target. Moreover, we evaluate our\nmethod extensively on two factor standard benchmark tasks where we\nsignificantly improve over competing unsupervised approaches, and we perform\ncompetitively in comparison to weakly- and self-supervised state-of-the-art\napproaches. The code is available at https://github.com/azencot-group/SKD.\n", "date": "2023-03-30", "categories": []} | |
| {"id": "2206.06169", "title": "Causal Representation Learning for Instantaneous and Temporal Effects in\n Interactive Systems", "abstract": " Causal representation learning is the task of identifying the underlying\ncausal variables and their relations from high-dimensional observations, such\nas images. Recent work has shown that one can reconstruct the causal variables\nfrom temporal sequences of observations under the assumption that there are no\ninstantaneous causal relations between them. In practical applications,\nhowever, our measurement or frame rate might be slower than many of the causal\neffects. This effectively creates \"instantaneous\" effects and invalidates\nprevious identifiability results. To address this issue, we propose iCITRIS, a\ncausal representation learning method that allows for instantaneous effects in\nintervened temporal sequences when intervention targets can be observed, e.g.,\nas actions of an agent. iCITRIS identifies the potentially multidimensional\ncausal variables from temporal observations, while simultaneously using a\ndifferentiable causal discovery method to learn their causal graph. In\nexperiments on three datasets of interactive systems, iCITRIS accurately\nidentifies the causal variables and their causal graph.\n", "date": "2022-06-13", "categories": []} | |
| {"id": "2211.03553", "title": "Learning Causal Representations of Single Cells via Sparse Mechanism\n Shift Modeling", "abstract": " Latent variable models such as the Variational Auto-Encoder (VAE) have become\na go-to tool for analyzing biological data, especially in the field of\nsingle-cell genomics. One remaining challenge is the interpretability of latent\nvariables as biological processes that define a cell's identity. Outside of\nbiological applications, this problem is commonly referred to as learning\ndisentangled representations. Although several disentanglement-promoting\nvariants of the VAE were introduced, and applied to single-cell genomics data,\nthis task has been shown to be infeasible from independent and identically\ndistributed measurements, without additional structure. Instead, recent methods\npropose to leverage non-stationary data, as well as the sparse mechanism shift\nassumption in order to learn disentangled representations with a causal\nsemantic. Here, we extend the application of these methodological advances to\nthe analysis of single-cell genomics data with genetic or chemical\nperturbations. More precisely, we propose a deep generative model of\nsingle-cell gene expression data for which each perturbation is treated as a\nstochastic intervention targeting an unknown, but sparse, subset of latent\nvariables. We benchmark these methods on simulated single-cell data to evaluate\ntheir performance at latent units recovery, causal target identification and\nout-of-domain generalization. Finally, we apply those approaches to two\nreal-world large-scale gene perturbation data sets and find that models that\nexploit the sparse mechanism shift hypothesis surpass contemporary methods on a\ntransfer learning task. We implement our new model and benchmarks using the\nscvi-tools library, and release it as open-source software at\nhttps://github.com/Genentech/sVAE.\n", "date": "2022-11-07", "categories": []} | |
| {"id": "2207.01575", "title": "De-Biasing Generative Models using Counterfactual Methods", "abstract": " Variational autoencoders (VAEs) and other generative methods have garnered\ngrowing interest not just for their generative properties but also for the\nability to dis-entangle a low-dimensional latent variable space. However, few\nexisting generative models take causality into account. We propose a new\ndecoder based framework named the Causal Counterfactual Generative Model\n(CCGM), which includes a partially trainable causal layer in which a part of a\ncausal model can be learned without significantly impacting reconstruction\nfidelity. By learning the causal relationships between image semantic labels or\ntabular variables, we can analyze biases, intervene on the generative model,\nand simulate new scenarios. Furthermore, by modifying the causal structure, we\ncan generate samples outside the domain of the original training data and use\nsuch counterfactual models to de-bias datasets. Thus, datasets with known\nbiases can still be used to train the causal generative model and learn the\ncausal relationships, but we can produce de-biased datasets on the generative\nside. Our proposed method combines a causal latent space VAE model with\nspecific modification to emphasize causal fidelity, enabling finer control over\nthe causal layer and the ability to learn a robust intervention framework. We\nexplore how better disentanglement of causal learning and encoding/decoding\ngenerates higher causal intervention quality. We also compare our model against\nsimilar research to demonstrate the need for explicit generative de-biasing\nbeyond interventions. Our initial experiments show that our model can generate\nimages and tabular data with high fidelity to the causal framework and\naccommodate explicit de-biasing to ignore undesired relationships in the causal\ndata compared to the baseline.\n", "date": "2022-07-04", "categories": []} | |
| {"id": "2208.08584", "title": "Robust Causal Graph Representation Learning against Confounding Effects", "abstract": " The prevailing graph neural network models have achieved significant progress\nin graph representation learning. However, in this paper, we uncover an\never-overlooked phenomenon: the pre-trained graph representation learning model\ntested with full graphs underperforms the model tested with well-pruned graphs.\nThis observation reveals that there exist confounders in graphs, which may\ninterfere with the model learning semantic information, and current graph\nrepresentation learning methods have not eliminated their influence. To tackle\nthis issue, we propose Robust Causal Graph Representation Learning (RCGRL) to\nlearn robust graph representations against confounding effects. RCGRL\nintroduces an active approach to generate instrumental variables under\nunconditional moment restrictions, which empowers the graph representation\nlearning model to eliminate confounders, thereby capturing discriminative\ninformation that is causally related to downstream predictions. We offer\ntheorems and proofs to guarantee the theoretical effectiveness of the proposed\napproach. Empirically, we conduct extensive experiments on a synthetic dataset\nand multiple benchmark datasets. The results demonstrate that compared with\nstate-of-the-art methods, RCGRL achieves better prediction performance and\ngeneralization ability.\n", "date": "2022-08-18", "categories": []} | |
| {"id": "2206.10591", "title": "Can Foundation Models Talk Causality?", "abstract": " Foundation models are subject to an ongoing heated debate, leaving open the\nquestion of progress towards AGI and dividing the community into two camps: the\nones who see the arguably impressive results as evidence to the scaling\nhypothesis, and the others who are worried about the lack of interpretability\nand reasoning capabilities. By investigating to which extent causal\nrepresentations might be captured by these large scale language models, we make\na humble efforts towards resolving the ongoing philosophical conflicts.\n", "date": "2022-06-14", "categories": []} | |
| {"id": "2206.01802", "title": "Do-Operation Guided Causal Representation Learning with Reduced\n Supervision Strength", "abstract": " Causal representation learning has been proposed to encode relationships\nbetween factors presented in the high dimensional data. However, existing\nmethods suffer from merely using a large amount of labeled data and ignore the\nfact that samples generated by the same causal mechanism follow the same causal\nrelationships. In this paper, we seek to explore such information by leveraging\ndo-operation to reduce supervision strength. We propose a framework that\nimplements do-operation by swapping latent cause and effect factors encoded\nfrom a pair of inputs. Moreover, we also identify the inadequacy of existing\ncausal representation metrics empirically and theoretically and introduce new\nmetrics for better evaluation. Experiments conducted on both synthetic and real\ndatasets demonstrate the superiorities of our method compared with\nstate-of-the-art methods.\n", "date": "2022-06-03", "categories": []} | |
| {"id": "2210.16034", "title": "A Survey on Causal Representation Learning and Future Work for Medical\n Image Analysis", "abstract": " Statistical machine learning algorithms have achieved state-of-the-art\nresults on benchmark datasets, outperforming humans in many tasks. However, the\nout-of-distribution data and confounder, which have an unpredictable causal\nrelationship, significantly degrade the performance of the existing models.\nCausal Representation Learning (CRL) has recently been a promising direction to\naddress the causal relationship problem in vision understanding. This survey\npresents recent advances in CRL in vision. Firstly, we introduce the basic\nconcept of causal inference. Secondly, we analyze the CRL theoretical work,\nespecially in invariant risk minimization, and the practical work in feature\nunderstanding and transfer learning. Finally, we propose a future research\ndirection in medical image analysis and CRL general theory.\n", "date": "2022-10-28", "categories": []} | |
| {"id": "2210.13583", "title": "Learning Latent Structural Causal Models", "abstract": " Causal learning has long concerned itself with the accurate recovery of\nunderlying causal mechanisms. Such causal modelling enables better explanations\nof out-of-distribution data. Prior works on causal learning assume that the\nhigh-level causal variables are given. However, in machine learning tasks, one\noften operates on low-level data like image pixels or high-dimensional vectors.\nIn such settings, the entire Structural Causal Model (SCM) -- structure,\nparameters, \\textit{and} high-level causal variables -- is unobserved and needs\nto be learnt from low-level data. We treat this problem as Bayesian inference\nof the latent SCM, given low-level data. For linear Gaussian additive noise\nSCMs, we present a tractable approximate inference method which performs joint\ninference over the causal variables, structure and parameters of the latent SCM\nfrom random, known interventions. Experiments are performed on synthetic\ndatasets and a causally generated image dataset to demonstrate the efficacy of\nour approach. We also perform image generation from unseen interventions,\nthereby verifying out of distribution generalization for the proposed causal\nmodel.\n", "date": "2022-10-24", "categories": []} | |
| {"id": "2210.13647", "title": "Temporally Disentangled Representation Learning", "abstract": " Recently in the field of unsupervised representation learning, strong\nidentifiability results for disentanglement of causally-related latent\nvariables have been established by exploiting certain side information, such as\nclass labels, in addition to independence. However, most existing work is\nconstrained by functional form assumptions such as independent sources or\nfurther with linear transitions, and distribution assumptions such as\nstationary, exponential family distribution. It is unknown whether the\nunderlying latent variables and their causal relations are identifiable if they\nhave arbitrary, nonparametric causal influences in between. In this work, we\nestablish the identifiability theories of nonparametric latent causal processes\nfrom their nonlinear mixtures under fixed temporal causal influences and\nanalyze how distribution changes can further benefit the disentanglement. We\npropose \\textbf{\\texttt{TDRL}}, a principled framework to recover time-delayed\nlatent causal variables and identify their relations from measured sequential\ndata under stationary environments and under different distribution shifts.\nSpecifically, the framework can factorize unknown distribution shifts into\ntransition distribution changes under fixed and time-varying latent causal\nrelations, and under observation changes in observation. Through experiments,\nwe show that time-delayed latent causal influences are reliably identified and\nthat our approach considerably outperforms existing baselines that do not\ncorrectly exploit this modular representation of changes. Our code is available\nat: \\url{https://github.com/weirayao/tdrl}.\n", "date": "2022-10-24", "categories": []} | |
| {"id": "2206.13973", "title": "Towards a Grounded Theory of Causation for Embodied AI", "abstract": " There exist well-developed frameworks for causal modelling, but these require\nrather a lot of human domain expertise to define causal variables and perform\ninterventions. In order to enable autonomous agents to learn abstract causal\nmodels through interactive experience, the existing theoretical foundations\nneed to be extended and clarified. Existing frameworks give no guidance\nregarding variable choice / representation, and more importantly, give no\nindication as to which behaviour policies or physical transformations of state\nspace shall count as interventions. The framework sketched in this paper\ndescribes actions as transformations of state space, for instance induced by an\nagent running a policy. This makes it possible to describe in a uniform way\nboth transformations of the micro-state space and abstract models thereof, and\nsay when the latter is veridical / grounded / natural. We then introduce\n(causal) variables, define a mechanism as an invariant predictor, and say when\nan action can be viewed as a ``surgical intervention'', thus bringing the\nobjective of causal representation \\& intervention skill learning into clearer\nfocus.\n", "date": "2022-06-28", "categories": []} | |
| {"id": "2207.05723", "title": "Latent Variable Models for Bayesian Causal Discovery", "abstract": " Learning predictors that do not rely on spurious correlations involves\nbuilding causal representations. However, learning such a representation is\nvery challenging. We, therefore, formulate the problem of learning a causal\nrepresentation from high dimensional data and study causal recovery with\nsynthetic data. This work introduces a latent variable decoder model, Decoder\nBCD, for Bayesian causal discovery and performs experiments in mildly\nsupervised and unsupervised settings. We present a series of synthetic\nexperiments to characterize important factors for causal discovery and show\nthat using known intervention targets as labels helps in unsupervised Bayesian\ninference over structure and parameters of linear Gaussian additive noise\nlatent structural causal models.\n", "date": "2022-07-12", "categories": []} | |
| {"id": "2207.07732", "title": "Partial Disentanglement via Mechanism Sparsity", "abstract": " Disentanglement via mechanism sparsity was introduced recently as a\nprincipled approach to extract latent factors without supervision when the\ncausal graph relating them in time is sparse, and/or when actions are observed\nand affect them sparsely. However, this theory applies only to ground-truth\ngraphs satisfying a specific criterion. In this work, we introduce a\ngeneralization of this theory which applies to any ground-truth graph and\nspecifies qualitatively how disentangled the learned representation is expected\nto be, via a new equivalence relation over models we call consistency. This\nequivalence captures which factors are expected to remain entangled and which\nare not based on the specific form of the ground-truth graph. We call this\nweaker form of identifiability partial disentanglement. The graphical criterion\nthat allows complete disentanglement, proposed in an earlier work, can be\nderived as a special case of our theory. Finally, we enforce graph sparsity\nwith constrained optimization and illustrate our theory and algorithm in\nsimulations.\n", "date": "2022-07-15", "categories": []} | |
| {"id": "2207.05259", "title": "Language-Based Causal Representation Learning", "abstract": " Consider the finite state graph that results from a simple, discrete,\ndynamical system in which an agent moves in a rectangular grid picking up and\ndropping packages. Can the state variables of the problem, namely, the agent\nlocation and the package locations, be recovered from the structure of the\nstate graph alone without having access to information about the objects, the\nstructure of the states, or any background knowledge? We show that this is\npossible provided that the dynamics is learned over a suitable\ndomain-independent first-order causal language that makes room for objects and\nrelations that are not assumed to be known. The preference for the most compact\nrepresentation in the language that is compatible with the data provides a\nstrong and meaningful learning bias that makes this possible. The language of\nstructured causal models (SCMs) is the standard language for representing\n(static) causal models but in dynamic worlds populated by objects, first-order\ncausal languages such as those used in \"classical AI planning\" are required.\nWhile \"classical AI\" requires handcrafted representations, similar\nrepresentations can be learned from unstructured data over the same languages.\nIndeed, it is the languages and the preference for compact representations in\nthose languages that provide structure to the world, uncovering objects,\nrelations, and causes.\n", "date": "2022-07-12", "categories": []} | |
| {"id": "2207.01603", "title": "Invariant and Transportable Representations for Anti-Causal Domain\n Shifts", "abstract": " Real-world classification problems must contend with domain shift, the\n(potential) mismatch between the domain where a model is deployed and the\ndomain(s) where the training data was gathered. Methods to handle such problems\nmust specify what structure is common between the domains and what varies. A\nnatural assumption is that causal (structural) relationships are invariant in\nall domains. Then, it is tempting to learn a predictor for label $Y$ that\ndepends only on its causal parents. However, many real-world problems are\n\"anti-causal\" in the sense that $Y$ is a cause of the covariates $X$ -- in this\ncase, $Y$ has no causal parents and the naive causal invariance is useless. In\nthis paper, we study representation learning under a particular notion of\ndomain shift that both respects causal invariance and that naturally handles\nthe \"anti-causal\" structure. We show how to leverage the shared causal\nstructure of the domains to learn a representation that both admits an\ninvariant predictor and that also allows fast adaptation in new domains. The\nkey is to translate causal assumptions into learning principles that\ndisentangle \"invariant\" and \"non-stable\" features. Experiments on both\nsynthetic and real-world data demonstrate the effectiveness of the proposed\nlearning algorithm. Code is available at https://github.com/ybjiaang/ACTIR.\n", "date": "2022-07-04", "categories": []} | |
| {"id": "2206.11646", "title": "Invariant Causal Mechanisms through Distribution Matching", "abstract": " Learning representations that capture the underlying data generating process\nis a key problem for data efficient and robust use of neural networks. One key\nproperty for robustness which the learned representation should capture and\nwhich recently received a lot of attention is described by the notion of\ninvariance. In this work we provide a causal perspective and new algorithm for\nlearning invariant representations. Empirically we show that this algorithm\nworks well on a diverse set of tasks and in particular we observe\nstate-of-the-art performance on domain generalization, where we are able to\nsignificantly boost the score of existing models.\n", "date": "2022-06-23", "categories": []} | |
| {"id": "2206.11131", "title": "Variational Causal Dynamics: Discovering Modular World Models from\n Interventions", "abstract": " Latent world models allow agents to reason about complex environments with\nhigh-dimensional observations. However, adapting to new environments and\neffectively leveraging previous knowledge remain significant challenges. We\npresent variational causal dynamics (VCD), a structured world model that\nexploits the invariance of causal mechanisms across environments to achieve\nfast and modular adaptation. By causally factorising a transition model, VCD is\nable to identify reusable components across different environments. This is\nachieved by combining causal discovery and variational inference to learn a\nlatent representation and transition model jointly in an unsupervised manner.\nSpecifically, we optimise the evidence lower bound jointly over a\nrepresentation model and a transition model structured as a causal graphical\nmodel. In evaluations on simulated environments with state and image\nobservations, we show that VCD is able to successfully identify causal\nvariables, and to discover consistent causal structures across different\nenvironments. Moreover, given a small number of observations in a previously\nunseen, intervened environment, VCD is able to identify the sparse changes in\nthe dynamics and to adapt efficiently. In doing so, VCD significantly extends\nthe capabilities of the current state-of-the-art in latent world models while\nalso comparing favourably in terms of prediction accuracy.\n", "date": "2022-06-22", "categories": []} | |
| {"id": "2206.01101", "title": "Weakly Supervised Representation Learning with Sparse Perturbations", "abstract": " The theory of representation learning aims to build methods that provably\ninvert the data generating process with minimal domain knowledge or any source\nof supervision. Most prior approaches require strong distributional assumptions\non the latent variables and weak supervision (auxiliary information such as\ntimestamps) to provide provable identification guarantees. In this work, we\nshow that if one has weak supervision from observations generated by sparse\nperturbations of the latent variables--e.g. images in a reinforcement learning\nenvironment where actions move individual sprites--identification is achievable\nunder unknown continuous latent distributions. We show that if the\nperturbations are applied only on mutually exclusive blocks of latents, we\nidentify the latents up to those blocks. We also show that if these\nperturbation blocks overlap, we identify latents up to the smallest blocks\nshared across perturbations. Consequently, if there are blocks that intersect\nin one latent variable only, then such latents are identified up to permutation\nand scaling. We propose a natural estimation procedure based on this theory and\nillustrate it on low-dimensional synthetic and image-based experiments.\n", "date": "2022-06-02", "categories": []} | |
| {"id": "2606.24160", "title": "An Introduction to Causal Reinforcement Learning", "abstract": "Causal inference provides a set of principles and tools that allow one to combine data and knowledge about an environment to reason with questions of counterfactual nature, i.e., what would have happened had reality been different, even when no data of this unrealized reality is currently available. Reinforcement learning provides methods to learn a policy that optimizes a specific measure (e.g., reward, regret) when the agent is deployed in an environment and pursues an exploratory, trial-and-error approach. These two disciplines have evolved independently and with virtually no interaction between them. We note that they operate over different aspects of the same building block, counterfactual relations, which makes them umbilically connected. Based on these observations, novel learning opportunities arise when this connection is explicitly acknowledged and mathematized. To realize this potential, we note that any environment where the RL agent is deployed can be decomposed as a collection of autonomous mechanisms with different causal invariances, parsimoniously modeled as a structural causal model; any standard RL setting implicitly encodes such a model. This formalization allows us to put under a unifying treatment different modes of learning, including online, off-policy, and causal calculus learning, which appear unrelated in the literature. However, these modalities are not exhaustive: we introduce several natural and pervasive classes of learning settings that entail novel dimensions of analysis. Specifically, we introduce and discuss through causal lenses generalized policy learning, where to intervene, imitation learning, and counterfactual learning. These tasks lead to a broader view of counterfactual learning and suggest great potential for studying causal inference and reinforcement learning side by side, which we call causal reinforcement learning (CRL).", "date": "2026-06-23", "categories": []} | |
| {"id": "2606.28228", "title": "Disentangling Continuous-Time Latent Dynamics: Identifiability of Latent SDEs via Diffusion Shifts", "abstract": "Causal representation learning for time series has developed strong identifiability results in discrete-time latent causal models, but identifiability in continuous-time latent stochastic differential equation (SDE) models remains largely open. We address this gap using environment-induced shifts in diffusion covariance. We study additive-noise latent SDEs observed through an unknown nonlinear diffeomorphism, with shared drift but environment-specific diffusion covariance. We show that two diagonal diffusion regimes with pairwise distinct coordinate-wise variance ratios identify the latent coordinates up to permutation and scaling, without any sparsity assumption on the drift. We first prove this result for linear Ornstein--Uhlenbeck systems and then extend it to general additive-noise latent SDEs. Under mild smoothness, the instantaneous drift-Jacobian causal graph is identifiable up to the same permutation. We propose a two-stage estimator for latent disentanglement and optional graph recovery; experiments on synthetic systems confirm the predicted identifiability boundary, and an application to Hardanger Bridge monitoring data illustrates the approach on real sensor trajectories.", "date": "2026-06-26", "categories": []} | |