Research map / Time series / Anomaly detection

COMET: Codebook-based Online-adaptive Multi-scale Embedding for Time-series Anomaly Detection

NeurIPS 2026 · Time series › Anomaly detection · approach: Multivariate anomaly detection · 0 citations (OpenAlex)

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 COMET, a codebook-based unsupervised multivariate time-series anomaly detector with multi-scale patch encoding, a vector-quantized coreset scored by memory and quantization distances, and online codebook adaptation at test time via activation-based pseudo-labels and contrastive learning.

Problem

Unsupervised multivariate time-series anomaly detection struggles with anomalies spanning different temporal ranges, inter-variable interactions, and distribution shifts at inference time that make learned normal patterns drift.

Approach

Uses a VQ codebook as a compact coreset of normal patterns; pseudo-labels test patches as normal if their quantized codebook entry was activated during training (no threshold), and adapts the model only on those normal samples with contrastive learning, combined with multi-scale patch encoding and density-aware local scaling distance.

Key results

• COMET (w/o TTA) achieves the best result in 39 of 45 metrics across five benchmarks, per the paper (introduction says 39; abstract says 36 for the TTA variant).
• On PSM, COMET (w/ TTA) reaches F1(K=100) 60.26 and VUS-PR 61.20 versus the best baseline VTT at 48.35 and 52.69.
• COMET uses 567K parameters versus CATCH's 48.6M on PSM, reaching Affiliation F1 of 71.9%.

Surprising findings

• Anomaly Transformer achieves very high F1(K=0) (e.g., 97.43 on PSM) yet near-zero F1(K=100) (2.20), illustrating the distortion from point adjustment.
• Removing the quantization score improves some F1(K=0) and VUS-PR results on several datasets (e.g., SMAP F1(K=0) 85.63 vs 82.30), so the full model is not best on every metric.

Limitations

• Full model is not best on all metrics; the paper reports 39/45 (w/o TTA) and 36 or 35/45 figures inconsistently across abstract, introduction and results.
• Training time is higher than some baselines due to per-scale forward passes.
• Best-F1 threshold search is used for threshold-based metrics, which may be optimistic.

Experimental setup

Datasets: PSM SWaT SMAP MSL WADI

Benchmarks: PSM SWaT SMAP MSL WADI

Metrics: F1(K=0) F1(K=100) Affiliation F1 AUC-ROC AUC-PR R-AUC-ROC R-AUC-PR VUS-ROC VUS-PR

Baselines: CATCH D3R VTT Anomaly Transformer LSTM-AE LSTM-VAE USAD

Builds on: VQ-VAE SOFTS Anomaly Transformer CATCH M2N2

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 0.87)
• Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing (NeurIPS 2026 · similarity 0.87)
• When Model Meets New Normals: Test-Time Adaptation for Unsupervised Time-Series Anomaly Detection (AAAI 2024 · similarity 0.86)
• Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection (NeurIPS 2026 · similarity 0.86)