Research map / Time series / Anomaly detection

Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection

NeurIPS 2025 · Time series › Anomaly detection · approach: Multivariate anomaly detection · 1 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 OracleAD, an unsupervised multivariate time-series anomaly detector that encodes each variable's causal embedding with per-variable LSTMs, aligns them via self-attention to a Stable Latent Structure, and scores anomalies by prediction error times SLS deviation.

Problem

Unsupervised multivariate time series anomaly detection where anomalies are rare and unlabeled, and existing complex models overstate performance and only partly detect anomalous segments while lacking root-cause interpretability.

Approach

Per-variable LSTM encoders with attention pooling produce causal embeddings trained for one-step prediction plus window reconstruction; self-attention links them into a shared latent space, and a Stable Latent Structure (epoch-averaged pairwise dissimilarity matrix) serves as a reference whose deviation, multiplied by prediction error, gives an interpretable anomaly score that localizes root-cause

Key results

• On PSM, OracleAD reaches F1 65.85 and VUS-PR 68.17, versus the best baseline F1 45.75 (SARAD) and VUS-PR 52.49 (Omni).
• On SMD (average of subsets), OracleAD reaches F1 43.03 and VUS-PR 47.52, above the best baselines by about 10.9 and 5.9 points.
• On SWaT, OracleAD achieves F1 76.50 and VUS-PR 74.16, while SARAD has the best SWaT A-ROC (85.40) and V-ROC (86.30).

Measured effects reported in the paper

Surprising findings

• Affiliation F1 is high for many models regardless of their other metric performance, which the authors attribute to a metric weakness (e.g., A.T and CATCH on SWaT and PSM).
• Several deep baselines (A.T, Patch, DLin, NLin, iTrans, Modern, DC) show nearly identical F1 values (43.45 on PSM), suggesting degenerate thresholding.

Limitations

• Authors note limitations when applied to complex or multimodal systems where globally consistent inter-variable relationships and continuous inputs may not hold.
• Deviation score may lag because structural disruption is only captured after anomalies enter the input window, and may cause false positives during recovery periods.
• Reported F1 uses optimal thresholds selected on test data, so threshold-dependent results are optimistic.

Experimental setup

Datasets: PSM SMD SWaT

Benchmarks: PSM SMD SWaT

Metrics: F1 R-F1 Aff-F1 AUC-ROC AUC-PR VUS-ROC VUS-PR

Baselines: AutoEncoder OmniAnomaly Anomaly Transformer DCdetector SARAD CATCH PatchTST TimesNet DLinear NLinear iTransformer ModernTCN

Builds on: Anomaly Transformer CATCH multi-head self-attention LSTM

Closest papers by meaning

• Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing (NeurIPS 2026 · similarity 0.90)
• Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability. (NeurIPS 2026 · similarity 0.90)
• The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.90)
• CrossAD: Time Series Anomaly Detection with Cross-scale Associations and Cross-window Modeling (NeurIPS 2025 · similarity 0.90)
• GCAD: Anomaly Detection in Multivariate Time Series from the Perspective of Granger Causality (AAAI 2025 · similarity 0.90)
• SARAD: Spatial Association-Aware Anomaly Detection and Diagnosis for Multivariate Time Series (NeurIPS 2024 · similarity 0.90)