Asymptotic Consistent Graph Structure Learning for Multivariate Time-Series Anomaly Detection
Huaxin Pang, Shikui Wei, Youru Li, Ting Liu, Huaqi Zhang, Ying Qin, Yao Zhao
Abstract
Capturing complex inter-variable relationships is crucial for anomaly detection for multivariate time series (MTS) data. In recent years, graph neural networks (GNNs) have been introduced to explicitly model complex inter-variable relationships from global static or local dynamic views, improving the performance of anomaly detection tasks significantly. However, these approaches usually ignore exploring distinct interaction patterns within short context windows or fail to capture unbiased inter-variable relationships over longer time windows. To address this limitation, we propose a novel Asymptotic Consistent Graph Structure Learning (ACGSL) framework for multivariate time series anomaly detection. Specifically, a Sequence Aggregation Module (SeAM) together with a denoising filter is developed to learn the unbiased representation for each temporal variable more effectively. Furthermore, a Feature-Accumulation Graph Construct Module (FA-GCM) enhanced by asymptotic consistent graph optimization loss is proposed to construct stable interaction graphs over adaptive time windows. We conduct experiments on five benchmarks and achieve remarkable performance enhancement in anomaly detection, even acquiring a maximum gain of 3.64% over the second-best baseline. Furthermore, ACGSL is able to explicitly give stable inter-variable interacted graphs over arbitrary local normal or anomalous states. Extensive experiments and ablation studies demonstrate the effectiveness and robustness of our proposed ACGSL in anomaly detection.