Research map / Label-efficient and robust learning / OOD and anomaly detection
Beyond Individual Input for Deep Anomaly Detection on Tabular Data
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 NPT-AD, a reconstruction-based deep anomaly detector for tabular data that uses a Non-Parametric Transformer to reconstruct masked features using both feature-feature and sample-sample attention, scoring anomalies by reconstruction loss against the training set.
Problem
Unsupervised and semi-supervised anomaly detection on tabular data, where existing methods capture either feature-feature or sample-sample dependencies but not both.
Approach
Train an NPT on masked feature reconstruction of normal samples, and at inference condition reconstruction on the whole unmasked training set and aggregate reconstruction losses over a bank of deterministic masks to form the anomaly score, combining feature-feature (ABA) and sample-sample (ABD) attention.
Key results
- NPT-AD achieves the highest mean F1-score of 68.8 versus 67.2 for the Interpolation Contrastive method and 56.2 for the vanilla transformer, with lowest mean rank 3.1.
- NPT-AD reaches mean AUROC 89.8, the highest among compared methods (Internal Contrastive 88.8), with lowest variance (0.8).
- In the ablation, Mask-KNN (feature-only) gets F1 57.5 and transformer (sample-only) gets 57.4, while NPT-AD reaches 68.8 F1 and 89.8 AUROC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine feature-feature and sample-sample dependencies (ablation) | NPT-AD · Ablation study | F1: 57.4 -> 68.8; 57.5 -> 68.8 | Helped |
| combine feature-feature and sample-sample dependencies (ablation) | NPT-AD · Ablation study | AUROC: 83.0 -> 89.8; 84.5 -> 89.8 | Helped |
| use NPT-AD instead of the strongest deep baseline, Internal contrastive | 31 datasets · 31 datasets (mean) | F1-score: 67.2 -> 68.8 | Helped |
| use NPT-AD instead of the strongest listed AUROC baseline, Internal contrastive | 31 datasets · 31 datasets (mean) | AUROC: 88.8 -> 89.8 | Helped |
Surprising findings
- Under training-set contamination, the vanilla transformer degrades less than NPT-AD, which the authors attribute to contamination affecting NPT-AD twice (training and inference).
- NPT-AD scales poorly with number of features d because the mask bank grows combinatorially with r.
Limitations
- High computational cost in memory and time for datasets with many features, due to mask bank size and model dimension growth.
- Performance degrades with training-set contamination, with F1 near 50% at 10% contamination.
- Does not outperform on all datasets (e.g., Speech, Seismic comparisons mixed; some datasets lose to KNN or COPOD).
Experimental setup
Datasets: Arrhythmia Thyroid ODDS datasets (28) Fraud Campaign Backdoor synthetic contamination dataset
Benchmarks: 31 tabular anomaly detection datasets (ODDS, Arrhythmia, Thyroid, Fraud, Campaign, Backdoor)
Metrics: F1-score AUROC mean rank mean std
Baselines: GOAD DROCC NeuTraL-AD Interpolation Contrastive learning Transformer Isolation Forest KNN RRCF COPOD PIDForest
Builds on: Non-Parametric Transformers neural-transformation contrastive anomaly detection (Interpolation Contrastive) NeuTraL-AD GOAD DROCC
Closest papers by meaning
- Towards One-for-All Anomaly Detection for Tabular Data (ICML 2026 · similarity 0.90)
- Tuning-Free One-Class Discriminant Learning for Tabular Anomaly Detection (ICML 2026 · similarity 0.89)
- Disentangling Tabular Data Towards Better One-Class Anomaly Detection (AAAI 2025 · similarity 0.89)
- Unsupervised Anomaly Detection for Tabular Data Using Deep Noise Evaluation (AAAI 2025 · similarity 0.89)
- Tab-Shapley: Identifying Top-k Tabular Data Quality Insights (AAAI 2025 · similarity 0.86)
- Scalable, Explainable and Provably Robust Anomaly Detection with One-Step Flow Matching (NeurIPS 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.