Research map / Label-efficient and robust learning / OOD and anomaly detection
Unsupervised Anomaly Detection for Tabular Data Using Deep Noise Evaluation
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 a noise-evaluation method for tabular unsupervised anomaly detection that trains an MLP/ResMLP to output the magnitude of injected diverse noise (zero for clean normal data), using the max of the output as the anomaly score, with theoretical bounds and experiments on 47 ADBench and 25 OCC datasets.
Problem
Unsupervised anomaly detection on tabular data, especially hard anomalies that lie close to the normal data, without access to any real anomalies during training.
Approach
Generate highly diverse noise-augmented copies of normal samples (mixed noise types and levels assigned per feature group) and train a network to regress zero on clean data and the absolute noise magnitude on noisy data; the aggregated predicted noise magnitude serves as the anomaly score, with no extra perturbator or hyperparameter in the objective.
Key results
- Ours-ResMLP reaches 92.68% mean AUC and Ours-MLP 92.27% on the 25 OCC datasets, with mean ranks 2.04 and 1.68 against 12 baselines.
- Abstract reports 92.27% AUC and 1.68 average ranking across more than 60 benchmark datasets.
- Noise levels around 1.0 are reported as optimal; Gaussian, Rayleigh and Uniform noise give stable ~92% AUC, while Salt&Pepper and Bernoulli perform poorly.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use noise evaluation instead of KNN | OCC anomaly detection · 25 tabular datasets | mean AUC (%) | Helped |
| use Gaussian noise level around 1.0 (ablation) | noise evaluation method · OCC | AUC | Helped |
| use noise levels that are too small or too high (ablation) | noise evaluation method · OCC | AUC | Hurt |
| replace Gaussian noise with Salt & Pepper or Bernoulli noise (ablation) | noise evaluation method · OCC | AUC | Hurt |
| use Gaussian, Rayleigh, or Uniform noise (ablation) | noise evaluation method · OCC | AUC: 92\% AUC | No clear effect |
| use maximum aggregation instead of minimum or mean (ablation) | noise evaluation model output · KDD-CUP99 | performance | Helped |
Surprising findings
- Large noise levels were found to hurt performance, and salt-and-pepper and Bernoulli noise performed poorly despite being natural discrete perturbations.
- Reported gains over the strongest baselines are small (about 1% AUC), with the MLP variant recommended for faster inference.
Limitations
- Focused only on tabular data; extension to images requires pretrained feature encoders and is untested.
- Theoretical guarantees rely on assumptions (Lipschitz-type constant c, bounds on g(h(x)) for normal data) that are hard to verify in practice.
- Performance is sensitive to the noise level, and the paper notes that the hard-anomaly divergence cannot be measured in practice.
Experimental setup
Datasets: ADBench (47 datasets) 25 OCC tabular datasets from UCI and prior works KDD-CUP99
Benchmarks: ADBench 47 tabular datasets (UAD) 25 tabular datasets (OCC)
Metrics: AUC F1 mean rank p-value (paired t-test)
Baselines: IForest COPOD AE KNN LOF DeepSVDD AnoGAN ECOD SCAD NeuTraLAD PLAD DPAD
Builds on: Deep SVDD PLAD denoising score matching ADBench PyOD
Closest papers by meaning
- Beyond Individual Input for Deep Anomaly Detection on Tabular Data (ICML 2024 · similarity 0.89)
- Disentangling Tabular Data Towards Better One-Class Anomaly Detection (AAAI 2025 · similarity 0.85)
- Dense Projection for Anomaly Detection (AAAI 2024 · similarity 0.85)
- Normal-Abnormal Guided Generalist Anomaly Detection (NeurIPS 2025 · similarity 0.85)
- Unsupervised Anomaly Detection in The Presence of Missing Values (NeurIPS 2024 · similarity 0.84)
- Contamination-Resilient Anomaly Detection via Adversarial Learning on Partially-Observed Normal and Anomalous Data (ICML 2024 · 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.