Research map / Label-efficient and robust learning / OOD and anomaly detection
SEAD: Unsupervised Ensemble of Streaming Anomaly Detectors
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 SEAD, an unsupervised, online ensemble that picks/weights streaming anomaly detectors via multiplicative weights over quantile-normalized scores, plus SEAD++ which samples K detectors per step to cut runtime.
Problem
Choosing the best anomaly detection algorithm and hyperparameters for a data stream online and without anomaly labels, under data drift; existing model selection and outlier ensembles are supervised, offline, or O(n) per point.
Approach
Treat each base streaming detector as an expert whose quantile-normalized anomaly score is both its prediction and its loss; weights are updated via one gradient step on a KL-regularized loss (multiplicative-weights / FTRL style), giving a constant-time per-point unsupervised ensemble. SEAD++ samples K detectors per step with Thompson-sampling-like selection to reduce compute.
Key results
- SEAD achieves the best average rank (5.07) across 14 real datasets versus base detectors and simple ensembles (mean 5.73, max 5.80, min 6.60), with lowest variance (6.64).
- On Ozone, SEAD takes about 1746 seconds versus 173684 seconds for MetaOD and 61543 seconds for Goswami et al. 2023, with AP 0.059 vs 0.052 and 0.070.
- With 13 random detectors added, SEAD's degradation is 0.88% while mean, max and min ensembles degrade 9.75%, 22.31% and 22.04%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SEAD instead of offline model selection | Ozone streaming anomaly detection · Ozone | average precision: 0.052 -> 0.059; 0.070 | Mixed |
| Use SEAD instead of offline model selection | Ozone streaming anomaly detection · Ozone | runtime (s): 173684 -> 1746; 61543 -> 1746 | Helped |
| Sample K = 6 detectors per time step instead of querying all detectors (ablation) | SEAD · Ozone | runtime (s): 1746.19 -> 1002.05 | Helped |
| Sample K = 6 detectors per time step instead of querying all detectors (ablation) | SEAD · Ozone | AP rank: 9 -> 6 | Helped |
| Add 13 random-score detectors (ablation) | SEAD and simple score-aggregation baselines · ODDS datasets | average performance: 9.75%; 22.31%; 22.04%; 0.88% | Mixed |
Surprising findings
- The rule-based SPC model performs best on several datasets (Pima, Mammography, HTTP, Telemetry), despite its simplicity.
- Unsupervised selection based on existing unsupervised metrics has been observed not to yield useful results (cited from prior work).
Limitations
- SEAD is only as good as its best base detector; it needs at least one base model that fits the data.
- No regret guarantee under non-stationarity is provided; only a stationary-style FTRL regret bound is cited.
- Comparisons against offline and ensemble baselines are limited; only simple ensembles are compared since SEAD is the first unsupervised online ensemble.
Experimental setup
Datasets: ODDS USP Data Stream Repository NAB Ionosphere Ozone
Benchmarks: ODDS USP Data Stream Repository Ozone Numenta Anomaly Detection Benchmark
Metrics: average precision score average rank runtime
Baselines: IForestASD xStream RRCF SPC mean ensemble max ensemble min ensemble MetaOD Goswami et al. 2023
Builds on: multiplicative weights update prediction with expert advice PySAD MetaOD Thompson sampling
Closest papers by meaning
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.85)
- Towards a Unified Framework of Clustering-based Anomaly Detection (ICML 2025 · similarity 0.85)
- An Evidence-Based Post-Hoc Adjustment Framework for Anomaly Detection Under Data Contamination (NeurIPS 2025 · similarity 0.85)
- Online Isolation Forest (ICML 2024 · similarity 0.84)
- Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection (NeurIPS 2026 · similarity 0.84)
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 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.