Research map / Label-efficient and robust learning / OOD and anomaly detection
Unsupervised Anomaly Detection Algorithms on Real-world Data: How Many Do We Need?
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.
Benchmarks 32 unsupervised anomaly detection algorithms on 52 real-world tabular datasets without hyperparameter tuning, finding kth-NN strongest overall and a local/global dataset split favoring kNN and EIF respectively.
Problem
Practitioners lack guidance on which unsupervised anomaly detection algorithm to use on real-world multivariate tabular data, and prior comparisons were small in dataset count or algorithm coverage and rarely distinguished local from global anomalies on real data.
Approach
A large-scale, optimization-free comparison averaging AUC over sensible hyperparameter grids, followed by clustering of algorithm-dataset performance that reveals distinct local and global dataset groups with different best algorithms.
Key results
- kth-NN significantly outperforms 14 of 15 other algorithms at p=0.05 in the overall comparison (mean AUC 0.769), with EIF top on mean AUC (0.773).
- On the local datasets kNN and ABOD are top performers, significantly outperforming 16 other methods at p=0.05.
- On the global datasets EIF has the highest mean AUC (0.824), outperforming 13 of 15 methods at p=0.05.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use kth-NN rather than the other evaluated algorithms | Unsupervised anomaly detection · All datasets | mean AUC: 0.769 | Helped |
| use kNN for local anomaly detection | Unsupervised anomaly detection · Local datasets | mean AUC: 0.740 | Helped |
| use EIF for global anomaly detection | Unsupervised anomaly detection · Global datasets | mean AUC: 0.824 | Helped |
| use local methods such as LOF instead of global methods | Unsupervised anomaly detection · Local datasets | mean AUC: LOF: 0.713; EIF: 0.614 | Mixed |
| use local methods such as LOF instead of global methods | Unsupervised anomaly detection · Global datasets | mean AUC: LOF: 0.603; EIF: 0.824 | Hurt |
| use CBLOF rather than the highest-performing algorithms | Unsupervised anomaly detection · All datasets | mean AUC: 0.519 vs 0.773 | Hurt |
| use EIF rather than kNN for global anomaly detection | Unsupervised anomaly detection · Global datasets | mean AUC: 0.774 -> 0.824 | Helped |
| use kNN rather than EIF for local anomaly detection | Unsupervised anomaly detection · Local datasets | mean AUC: 0.614 -> 0.740 | Helped |
Surprising findings
- Original CBLOF is by far the worst performer, significantly outperformed by 21 of 22 algorithms at p=0.10 level in the overall comparison.
- Local methods such as LOF, COF and ensemble-LOF rank among the worst overall but among the best on the local datasets, a Simpson's paradox-like reversal.
Limitations
- Neural network methods were not well tuned for tabular data and perform poorly, limiting conclusions about deep methods.
- Datasets are mostly from ODDS and near-duplicates were removed; the collection is not constructed systematically and may be biased.
- Only real-valued tabular data considered; no categorical, image, text or time-series data.
Experimental setup
Datasets: ODDS ADBench ELKI datasets Goldstein comparison datasets GAAL datasets extended autoencoder datasets EOAD yeast6
Benchmarks: 52 multivariate tabular anomaly detection datasets
Metrics: ROC-AUC percentage of maximum AUC Iman-Davenport statistic Nemenyi post-hoc test
Baselines: kNN LOF Isolation Forest COPOD ECOD DeepSVDD CBLOF OCSVM
Builds on: PyOD ODDS ADBench
Closest papers by meaning
- Towards a Unified Framework of Clustering-based Anomaly Detection (ICML 2025 · similarity 0.85)
- Unifying Unsupervised Graph-Level Anomaly Detection and Out-of-Distribution Detection: A Benchmark (ICLR 2025 · similarity 0.84)
- We Need to Rethink Benchmarking in Anomaly Detection (NeurIPS 2026 · similarity 0.83)
- Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.83)
- SEAD: Unsupervised Ensemble of Streaming Anomaly Detectors (ICML 2025 · similarity 0.83)
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.83)
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.