Research map / Time series / Anomaly detection

Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection

NeurIPS 2026 · Time series › Anomaly detection · approach: Time series anomaly detection 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 RAMSAD, a training-free retrieval-based model selector that picks time-series anomaly detectors by averaging detector VUS-PR profiles of the top-k most similar knowledge-base series, optionally using TSFM embeddings aligned with dataset descriptions.

Problem

Selecting the best anomaly detector for a heterogeneous time series, especially under out-of-distribution (domain shift) conditions where trained meta-learning selectors degrade.

Approach

Treat detector selection as similarity search: retrieve top-k similar series from a labeled knowledge base (via segment-level TSFM embeddings aggregated to series level) and pick the detector with the highest average VUS-PR over those neighbors, with an optional offline multi-positive contrastive alignment to dataset text descriptions.

Key results

• RAMSAD achieves average VUS-PR 0.74 in-distribution, versus 0.69 for the best trained selector SATzilla.
• In OOD, RAMSAD with N=10 ensemble exceeds OE by 2% in VUS-PR while being about 4.7x faster.
• Retrieval in plain Chronos-2 TSFM space reaches 0.75 ID and 0.59 OOD, outperforming DTW (0.69 ID, 0.44 OOD).

Surprising findings

• A frozen off-the-shelf TSFM retrieval space already outperforms all trained selectors in ID.
• Performance is driven by representation-space proximity rather than shared domain membership, per the leave-one-domain-out analysis.

Limitations

• Limited to static univariate time series; multivariate, concept drift, and streaming are out of scope.
• Requires a labeled knowledge base with precomputed detector performance profiles for every candidate detector.
• Performance drops under OOD (0.74 to 0.59 ID-to-OOD for N=1).

Experimental setup

Datasets: TSB-AD

Benchmarks: TSB-AD

Metrics: VUS-PR average rank (critical difference)

Baselines: SATzilla OE UADB AutoTSAD AutoOD-A AutoOD-C MetaOD TSADAMS ISAC ARGOSMART

Builds on: ARGOSMART ISAC Chronos-2 TSB-AD

Closest papers by meaning

• Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.89)
• The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.88)
• Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability. (NeurIPS 2026 · similarity 0.88)
• Position: Quo Vadis, Unsupervised Time Series Anomaly Detection? (ICML 2024 · similarity 0.87)
• CrossAD: Time Series Anomaly Detection with Cross-scale Associations and Cross-window Modeling (NeurIPS 2025 · similarity 0.87)
• When Model Meets New Normals: Test-Time Adaptation for Unsupervised Time-Series Anomaly Detection (AAAI 2024 · similarity 0.87)