Research map / Time series / Anomaly detection

Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability.

NeurIPS 2026 · Time series › Anomaly detection · approach: Time series anomaly detection evaluation · 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.

Introduces SHAD, a real-world, fully annotated benchmark of 215 high-dimensional cloud-storage telemetry series for end-to-end evaluation of time series anomaly detection, explanation, and LLM-based interpretation.

Problem

Existing time series anomaly detection benchmarks focus on detection accuracy and lack dimension-level labels and rich semantic context needed to evaluate explainability and interpretability of anomalies.

Approach

A benchmark built from controlled fault injection on a Scality RING distributed storage cluster, providing binary top-level labels, per-dimension root-cause labels, and structured XML textual annotations with per-metric documentation, used to evaluate detection (18 detectors), dimension-level attribution (NDCG), and zero-shot LLM interpretation.

Key results

• Top semi-supervised detectors (OA, USAD, AE, CNN) are statistically tied with unsupervised KMAD under VUS-PR; no method is uniformly accurate across anomaly types.
• Native detector dimension scores for OA, CNN, and KMAD perform near random for dimension localization, showing detection accuracy does not yield attribution.
• Zero-shot LLMs perform poorly: best anomaly type F1 is 21% (images), dimension localization accuracy near 1-2%, and time localization up to 12% with Claude.

Surprising findings

• Native detector attribution scores are nearly indistinguishable from random for dimension localization despite strong detection performance.
• Most LLMs predicted an anomaly for nearly all series regardless of representation, indicating bias toward anomaly presence.

Limitations

• Single storage system and one fixed baseline workload limit generality.
• LLM evaluation is zero-shot, single-call only, and limited to a few models.
• Anomalies are injected in a controlled testbed rather than observed in production.

Experimental setup

Datasets: SHAD

Benchmarks: SHAD

Metrics: VUS-PR NDCG F1-score accuracy

Baselines: OA USAD AE CNN LSTM KMAD ChatTS

Builds on: TSB-AD KMAD Catch22

Closest papers by meaning

• The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.90)
• Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.90)
• Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing (NeurIPS 2026 · similarity 0.89)
• Position: Quo Vadis, Unsupervised Time Series Anomaly Detection? (ICML 2024 · similarity 0.88)
• KAN-AD: Time Series Anomaly Detection with Kolmogorov–Arnold Networks (ICML 2025 · similarity 0.88)
• CrossAD: Time Series Anomaly Detection with Cross-scale Associations and Cross-window Modeling (NeurIPS 2025 · similarity 0.88)