Adaptive Hierarchical Attention for Multivariate Time Series Anomaly Detection

Abstract

In multivariate time series (MTS) anomaly detection, existing graph neural network (GNN) methods often neglect multi-level feature representations, relying solely on the final layer output and fixed thresholds, which leads to information loss and high false-positive rates. To address this issue, we propose a MTS adaptive hierarchical attention (MTS-AHA) model that integrates cross-layer features from local to global topological structures across GNN layers. Additionally, we introduce a dynamic threshold module that adaptively adjusts the threshold by fitting the tail features of the distribution. Extensive experiments on three datasets validate the superiority of our method and its components.