Research map / Time series / Anomaly detection

Causality-Aware Contrastive Learning for Robust Multivariate Time-Series Anomaly Detection

ICML 2025 · 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 CAROTS, a causality-aware contrastive learning pipeline for multivariate time-series anomaly detection that uses causality-preserving and causality-disturbing augmentors plus a similarity-filtered one-class contrastive loss, combined with a forecasting-based causal discovery error score.

Problem

Unsupervised multivariate time-series anomaly detection often ignores inter-variable causal relationships, leading to false alarms when normal causal-preserving variations look anomalous and missed anomalies that break causal structure.

Approach

Use a forecasting-based causal discoverer (CUTS+) to generate causality-preserving positives and causality-disturbing negatives, train a contrastive encoder with a similarity-filtered one-class contrastive loss that progressively admits more diverse positives, and ensemble the latent-centroid distance with the forecasting error.

Key results

• On SWaT, CAROTS reaches AUROC 0.852, AUPRC 0.764 and F1 0.791, outperforming all listed baselines (best baseline CTAD AUROC 0.820).
• On Lorenz96, CAROTS achieves AUROC 0.998 for point global and 0.975 for point contextual anomalies, with average 0.909 versus 0.700 for the best baseline TimesNet.
• Ablation on SWaT shows removing the similarity filter drops AUROC from 0.852 to 0.819 and removing the CL score drops it to 0.814.

Measured effects reported in the paper

Surprising findings

• CAROTS without the forecasting-error score (CAROTS dagger) scores higher than full CAROTS on WADI AUROC (0.622 vs 0.502) and several other datasets, so the ensemble does not uniformly help.
• On VAR, CAROTS scores lower than TimesNet, USAD and the reconstruction baselines on the PG and CT anomaly types, with average AUROC 0.805 versus 0.810 for USAD.

Limitations

• Assumes causal relationships in normal data are time-invariant; non-stationary causal shifts are not handled.
• Does not explicitly model confounding variables.
• Does not model exogenous variables.

Experimental setup

Datasets: SWaT WADI PSM SMD_2-1 SMD_3-7 MSL_P-14 MSL_P-15 Lorenz96 VAR

Benchmarks: SWaT WADI PSM SMD_2-1 SMD_3-7 MSL_P-14 MSL_P-15 Lorenz96 VAR

Metrics: AUROC AUPRC F1

Baselines: AnomalyTransformer TimesNet USAD SimCLR SSD CSI CTAD CARLA

Builds on: CUTS+ CTAD SimCLR CARLA Granger causality

Closest papers by meaning

• Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.89)
• GCAD: Anomaly Detection in Multivariate Time Series from the Perspective of Granger Causality (AAAI 2025 · similarity 0.88)
• Explaining Time Series via Contrastive and Locally Sparse Perturbations (ICLR 2024 · similarity 0.86)
• Graph-Aware Contrasting for Multivariate Time-Series Classification (AAAI 2024 · similarity 0.86)
• CauDiTS: Causal Disentangled Domain Adaptation of Multivariate Time Series (ICML 2024 · similarity 0.86)
• Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection (NeurIPS 2026 · similarity 0.85)