license: mit
author: Christopher Alberto Hamonangan
datasets:
- BhinnekaAI/TSB-Fraud-Tabular
- BhinnekaAI/MNT-Log-Traffic
- BhinnekaAI/CLB-Patient-Screening
- BhinnekaAI/IMC-Sensor-Telemetry
- BhinnekaAI/ESCL-Logistics-Audit
tags:
- Machine-Learning
- Unsupervised
- Algorithm
- AI
Homeostatic Distribution Score (HDS)
An Unsupervised Multi-Context Anomaly Detection Framework
Author : Christopher Alberto Hamonangan
HDS (Homeostatic Distribution Score) is an experimental unsupervised anomaly detection framework designed to identify samples that deviate from a learned reference distribution through multi-context structural analysis.
The framework combines multiscale spatial distribution analysis, linear relational inconsistency, hierarchical neighborhood analysis, adaptive token fusion, and episodic memory retrieval into a unified anomaly scoring mechanism.
Key Features
- Multiscale Distribution Analysis: Evaluates neighborhood sparsity using a hybrid Euclidean–Mahalanobis approach.
- Linear Network: Measures global projection residuals and local relational deviations.
- Tree Network: Uses KD-Tree neighborhood distances to assess spatial inconsistency.
- Adaptive Token Fusion: Combines structural signals through sample-dependent weighting.
- Episodic Memory: Retrieves similar reference samples to incorporate contextual information into anomaly scoring.
Model Variants
| Variant | Local Relational Engine | Design Goal |
|---|---|---|
| HDS-Deep | Local SVD-based analysis | Capture local neighborhood structure and relational deviations |
| HDS-Fast | Neighborhood mean-deviation | Reduce the computational cost of local relational analysis |
Both variants produce continuous anomaly scores for ranking potentially anomalous samples. Their relative performance depends on the dataset and evaluation conditions.
Benchmark Evaluation
HDS-Deep and HDS-Fast are evaluated against conventional anomaly detection methods across five tabular anomaly detection domains.
Evaluation Metrics
- ROC-AUC: Measures the ability to rank anomalous samples above normal samples across classification thresholds.
- PR-AUC: Measures the precision–recall trade-off and is particularly informative for imbalanced datasets.
Higher scores indicate better ranking performance.
Benchmark Datasets
The following links provide access to the datasets used in the evaluation. Replace each placeholder with the corresponding public dataset repository.
| Dataset | Repository |
|---|---|
| Credit Card Fraud | TSB-Fraud-Tabular |
| Network Intrusion | MNT-Log-Traffic |
| Patient Screening | CLB-Patient-Screening |
| IoT Sensor Anomaly | IMC-Sensor-Telemetry |
| Logistics Audit | ESCL-Logistics-Audit |
1. Credit Card Fraud Simulator
| Method | ROC-AUC | PR-AUC |
|---|---|---|
| HDS-Deep | 0.885 | 0.659 |
| HDS-Fast | 0.856 | 0.606 |
| Isolation Forest | 0.568 | 0.125 |
| HBOS | 0.487 | 0.101 |
| KNN-Distance | 0.762 | 0.286 |
| LOF | 0.773 | 0.304 |
| One-Class SVM | 0.616 | 0.173 |
| Robust Covariance | 0.372 | 0.076 |
2. Network Intrusion Logs
| Method | ROC-AUC | PR-AUC |
|---|---|---|
| HDS-Deep | 0.660 | 0.152 |
| HDS-Fast | 0.666 | 0.153 |
| Isolation Forest | 0.989 | 0.835 |
| HBOS | 1.000 | 0.990 |
| KNN-Distance | 0.824 | 0.226 |
| LOF | 0.337 | 0.069 |
| One-Class SVM | 0.849 | 0.259 |
| Robust Covariance | 0.379 | 0.074 |
3. Patient Biomarker Screening
| Method | ROC-AUC | PR-AUC |
|---|---|---|
| HDS-Deep | 0.902 | 0.756 |
| HDS-Fast | 0.903 | 0.763 |
| Isolation Forest | 0.742 | 0.312 |
| HBOS | 0.666 | 0.193 |
| KNN-Distance | 0.852 | 0.549 |
| LOF | 0.831 | 0.353 |
| One-Class SVM | 0.784 | 0.400 |
| Robust Covariance | 0.889 | 0.734 |
4. IoT Sensor Anomaly
| Method | ROC-AUC | PR-AUC |
|---|---|---|
| HDS-Deep | 0.805 | 0.370 |
| HDS-Fast | 0.822 | 0.398 |
| Isolation Forest | 0.946 | 0.736 |
| HBOS | 0.541 | 0.116 |
| KNN-Distance | 0.998 | 0.984 |
| LOF | 0.398 | 0.076 |
| One-Class SVM | 0.680 | 0.240 |
| Robust Covariance | 0.740 | 0.162 |
5. Supply Chain Inventory Audit
| Method | ROC-AUC | PR-AUC |
|---|---|---|
| HDS-Deep | 0.468 | 0.092 |
| HDS-Fast | 0.494 | 0.101 |
| Isolation Forest | 0.553 | 0.126 |
| HBOS | 0.552 | 0.132 |
| KNN-Distance | 0.549 | 0.135 |
| LOF | 0.541 | 0.124 |
| One-Class SVM | 0.551 | 0.133 |
| Robust Covariance | 0.547 | 0.134 |
HDS-Deep vs. HDS-Fast: Direct Comparison
The following table compares the two HDS variants across all five benchmark categories.
| Dataset | Deep ROC-AUC | Fast ROC-AUC | Deep PR-AUC | Fast PR-AUC |
|---|---|---|---|---|
| Credit Card Fraud | 0.885 | 0.856 | 0.659 | 0.606 |
| Network Intrusion | 0.660 | 0.666 | 0.152 | 0.153 |
| Patient Biomarker | 0.902 | 0.903 | 0.756 | 0.763 |
| IoT Sensor | 0.805 | 0.822 | 0.370 | 0.398 |
| Supply Chain | 0.468 | 0.494 | 0.092 | 0.101 |
| Macro Average | 0.744 | 0.748 | 0.406 | 0.404 |
Head-to-Head Summary
- ROC-AUC: HDS-Fast achieves a macro average of 0.748, compared with 0.744 for HDS-Deep.
- PR-AUC: HDS-Deep achieves a macro average of 0.406, compared with 0.404 for HDS-Fast.
- Credit Card Fraud: HDS-Deep performs better on both metrics.
- Other four datasets: HDS-Fast records slightly higher ROC-AUC and PR-AUC scores.
The aggregate differences are small. Neither variant demonstrates universal superiority across all datasets. Runtime and memory measurements are needed to determine whether HDS-Fast provides a meaningful computational advantage.
Evaluation Protocol
The reported benchmark uses train–test evaluation for each dataset.
The primary evaluation metrics are ROC-AUC and PR-AUC. HDS-Deep and HDS-Fast are compared with Isolation Forest, HBOS, KNN-Distance, LOF, One-Class SVM, and Robust Covariance.
For reproducible evaluation, the accompanying experiment code should specify:
- Dataset versions and preprocessing procedures.
- Train–test split strategy and random seeds.
- Feature selection and scaling procedures.
- Anomaly label definitions and score direction.
- Implementation details and hyperparameters for each baseline.
All preprocessing and model fitting should be performed without using test labels. Test labels should be reserved for final metric calculation.
Intended Use
HDS is intended for research and experimentation in:
- Unsupervised anomaly detection.
- Multicontext structural scoring.
- Distribution-based anomaly ranking.
- Comparative evaluation of tabular anomaly detection methods.
HDS is an experimental research framework. Its performance varies across datasets, and the reported results should not be interpreted as evidence of universal superiority over established methods.
Technical Summary
| Property | Description |
|---|---|
| Framework | Homeostatic Distribution Score (HDS) |
| Task | Unsupervised anomaly detection |
| Input | Tabular feature vectors |
| Output | Continuous anomaly ranking scores |
| Variants | HDS-Deep, HDS-Fast |
| Primary metrics | ROC-AUC, PR-AUC |
| Benchmark domains | Five |
| Baseline methods | Isolation Forest, HBOS, KNN-Distance, LOF, One-Class SVM, Robust Covariance |
Project Links
- Source Code: HDS GitHub
- Datasets: See the Benchmark Datasets section above.
Research Status
HDS is an experimental framework for multi-context unsupervised anomaly detection. Further testing across random seeds, dataset variations, and controlled computational benchmarks can strengthen its reproducibility and robustness.
Contributions, feedback, and independent evaluations are welcome.