POST: Prior-Observation Adversarial Learning for MTS Anomaly Detection

A new framework called POST (Prior-Observation Adversarial Learning of Spatio-Temporal Associations) is proposed for multivariate time series anomaly detection. It addresses the spatial over-generalization problem in existing methods that use Graph Neural Networks with sequence models. POST unifies spatio-temporal modeling through adversarial learning, alternately learning adjacency matrices as structural priors and modeling association discrepancies between prior and data-driven observations. This improves detection recall and enables channel-level anomaly localization. The paper is available on arXiv with ID 2605.18128.