New multi-view graph detector beats LSTM-AE by 5.1 points on anomaly detection

A new arXiv paper by Wei et al. presents a comprehensive benchmark of multivariate time-series (MTS) anomaly detection, testing 10 family-representative detectors (statistical, reconstruction, association, frequency, and transformer-based) on five datasets (SMD, MSL, SMAP, PSM, MSDS) with unified protocols. The benchmark yields three key findings: no single-bias baseline dominates across all datasets; absolute perturbation VUS-ROC is more informative than retention ratios; and MSDS behaves as an event-dense deployment workload rather than a sparse point-anomaly benchmark. All methods shared the same windowing, scoring, hardware, and metric protocols, with robustness tests using three random seeds.

The study introduces an adaptive detector family that combines a NOTEARS-constrained directed channel-graph view with optional patch-attention and temporal-association views. This multi-view approach achieves the best macro-average VUS-ROC (0.675, +5.1 points over second-best LSTM-AE), ranks first overall, and reaches top-3 on every dataset. Its robustness gains are especially notable: under noise, channel dropout, and time-shift perturbations, it obtains the strongest absolute VUS-ROC. The authors release MSDS preprocessing, configurations, scripts, and seed-level metric dumps.