Research map / Label-efficient and robust learning / OOD and anomaly detection
Disentangling Tabular Data Towards Better One-Class 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 Disent-AD, which disentangles two non-overlapping correlated attribute subsets (CorrSets) from normal tabular data via two-head self-attention with a cosine disentangling loss and a reconstruction loss, for one-class tabular anomaly detection.
Problem
One-class anomaly detection on tabular data requires learning the intrinsic correlation among attributes of normal samples, but the prior learnable-mask method (MCM) can collapse to near-uniform masks that mask nothing, weakening correlation learning.
Approach
Instead of learnable soft masks, use a two-head self-attention to implicitly disentangle two non-overlapping attribute subsets (CorrSets) from normal samples, enforce their independence with a cosine-similarity disentangling loss, and reconstruct the full input from each subset; the sum of the two reconstruction errors is the anomaly score.
Key results
- Averaged over 20 tabular datasets, the method reaches 78.7% AUC-PR and 91.7% AUC-ROC, versus 72.6% and 89.6% for the next best method MCM (reported as a 6.1% and 2.1% improvement).
- Achieves the best mean rank on both metrics (2.0 for AUC-PR, 1.7 for AUC-ROC out of 12), and is best on 13 of 20 datasets for both metrics.
- Large gains on Census (39.7 vs 21.3 AUC-PR) and Glass (58.9 vs 15.3 AUC-PR for the best baseline MCM).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use two distinct subsets with two-head self-attention instead of one attention head and one subset (ablation) | Disent-AD · Thyroid | AUC-PR: 72.6$^{\pm1.6}$ -> 89.4$^{\pm0.7}$ | Helped |
| Use two-head self-attention to extract two subsets instead of deriving both from one attention head (ablation) | Disent-AD · Thyroid | AUC-PR: 78.8$^{\pm4.0}$ -> 89.4$^{\pm0.7}$ | Helped |
| Add the disentangling loss to make the two subsets distinct (ablation) | Disent-AD · Thyroid | AUC-PR: 72.4$^{\pm1.2}$ -> 89.4$^{\pm0.7}$ | Helped |
| Replace the disentangling loss with MCM's diversity loss (ablation) | Disent-AD · Thyroid | AUC-PR: 89.4 -> 88.3 | Hurt |
| Replace the disentangling loss with MCM's diversity loss (ablation) | Disent-AD · Satellite | AUC-PR: 89.1 -> 88.9 | Hurt |
| Use patch-splitting preprocessing with shape 3×M/2 rather than no preprocessing (ablation) | Disent-AD · Satellite | AUC-PR: 75.7$^{\pm0.2}$ -> 89.1$^{\pm0.0}$ | Helped |
| Shuffle the original attribute order before preprocessing (ablation) | Disent-AD · Satellite | AUC-PR: 89.1$^{\pm0.0}$ -> 88.5$^{\pm0.5}$ | Hurt |
| Increase training epochs and latent channel count to 200 and 512 (ablation) | Disent-AD · Satellite | AUC-PR: 88.3$^{\pm0.1}$ -> 89.1$^{\pm0.0}$ | Helped |
Surprising findings
- Using three subsets led to a near-uniform third attention map, so performance decreased as the subset number increased beyond two.
- Simply reusing the disentangling loss idea with MCM's diversity loss instead of the cosine disentangling loss gave lower results (e.g., 88.3 vs 89.4 AUC-PR on Thyroid).
Limitations
- Method is designed specifically for tabular data and does not directly apply to image or point cloud data.
- Performance on some datasets (e.g., Campaign, Mammography, Cardiotocography comparisons) is not best; MCM is better on several datasets.
- Patch-splitting preprocessing is needed for some datasets to facilitate disentanglement, and it lacks a principled selection rule.
Experimental setup
Datasets: Arrhythmia Breastw Cardio Census Campaign Cardiotocography Fraud Glass Ionosphere Mammography NSL-KDD Optdigits
Benchmarks: Arrhythmia Breastw Cardio Census Campaign Cardiotocography Fraud Glass Ionosphere Mammography NSL-KDD Optdigits
Metrics: AUC-PR AUC-ROC mean rank
Baselines: IForest LOF OCSVM ECOD DAGMM DeepSVDD GOAD NeuTraLAD ICL SLAD MCM
Builds on: MCM DeepOD
Closest papers by meaning
- Beyond Individual Input for Deep Anomaly Detection on Tabular Data (ICML 2024 · similarity 0.89)
- Unsupervised Anomaly Detection for Tabular Data Using Deep Noise Evaluation (AAAI 2025 · similarity 0.85)
- Tab-Shapley: Identifying Top-k Tabular Data Quality Insights (AAAI 2025 · similarity 0.85)
- Beyond Normal References: Discriminative Few-Shot Anomaly Detection (NeurIPS 2026 · similarity 0.85)
- ReMasker: Imputing Tabular Data with Masked Autoencoding (ICLR 2024 · similarity 0.84)
- CACTI: Leveraging Copy Masking and Contextual Information to Improve Tabular Data Imputation (ICML 2025 · similarity 0.84)
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.