Research map / Label-efficient and robust learning / OOD and anomaly detection
Towards a Unified Framework of Clustering-based Anomaly Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UniCAD, a unified probabilistic mixture model (Student-t, anomaly-aware likelihood with indicator function) that jointly learns representations, clustering and anomaly scores, plus a gravity-inspired vector-sum anomaly score, evaluated on 30 ADBench tabular datasets.
Problem
Unsupervised anomaly detection lacks a unified theory linking representation learning, clustering, and anomaly detection, so anomalies corrupt representation and clustering and their interdependencies are not exploited.
Approach
Maximize an anomaly-aware data likelihood under a Student-t mixture model in the learned latent space, with a learnable indicator that down-weights anomalies; derive the anomaly score from the likelihood and improve it with a gravity-inspired vector-sum over cluster prototypes.
Key results
- UniCAD (vector) achieves the best average AUC-ROC rank of 2.6 across 10 shown baselines on 30 tabular datasets, versus 3.7 for UniCAD (scalar) and 5.1 for the best baseline IForest.
- Ablation: replacing Student-t with Gaussian worsens average rank from 4.2 to 6.2 (17 baselines plus variants).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use vector summation instead of the scalar anomaly score (ablation) | UniCAD · 30 tabular datasets | Avg. Rank (AUC-ROC): 3.7 -> 2.6 | Helped |
| replace the Student's t distribution with a Gaussian distribution (ablation) | UniCAD · 30 tabular datasets | Avg. Rank (AUC-ROC): 4.2 -> 6.2 | Hurt |
| omit the likelihood maximization loss J(Θ, Φ) (ablation) | UniCAD · 30 tabular datasets | Avg. Rank (AUC-ROC): 4.2 -> 6.6 | Hurt |
| remove the anomaly indicator δ(xᵢ) (ablation) | UniCAD · 30 tabular datasets | Avg. Rank (AUC-ROC): 4.2 -> 5.0 | Hurt |
Surprising findings
- Scalar anomaly score (1/likelihood) is weaker than the proposed vector-sum score on most datasets, and the Gaussian mixture (DAGMM-style GMM) baseline performs poorly.
Limitations
- Evaluated only on tabular data; applicability to time series and multimodal anomaly detection is untested.
- Anomaly ratio l must be specified; optimal range tends to be lower than true anomaly ratio, and cluster number k needs careful choice.
- Shortcut-prevention via reconstruction loss is a heuristic constraint rather than a derived term.
Experimental setup
Datasets: ADBench tabular datasets (30)
Benchmarks: ADBench 30 tabular datasets
Metrics: AUC-ROC AUC-PR average rank
Baselines: SOD HBOS PCA OCSVM LOF KNN LODA IForest DAGMM ECOD COPOD DBSCAN
Builds on: DAGMM KMeans-- DCOD ADBench
Closest papers by meaning
- Dissect Black Box: Interpreting for Rule-Based Explanations in Unsupervised Anomaly Detection (NeurIPS 2024 · similarity 0.87)
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.85)
- Unsupervised Anomaly Detection Algorithms on Real-world Data: How Many Do We Need? (NeurIPS 2024 · similarity 0.85)
- AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 2025 · similarity 0.85)
- SEAD: Unsupervised Ensemble of Streaming Anomaly Detectors (ICML 2025 · similarity 0.85)
- Normal-Abnormal Guided Generalist Anomaly Detection (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.