Diffusion Graph Model for Time Series Anomaly Detection via Anomaly-aware Graph Sparsification and Augmentation

Unsupervised methods, particularly reconstruction-based methods have become the dominant approach for multivariate time series anomaly detection (TSAD), which distinguish between normal and abnormal series based on the magnitude of the reconstruction error. However, in this process, the heterophilic connections (normal łeftrightarrow abnormal datapoints) often lead the model to simultaneously capture the distributions of both normal and abnormal data, impeding effective anomaly detection based on reconstruction error. To address this challenge, we introduce a novel diffusion graph model framework, dubbed øurmethod, which jointly models spatial-temporal correlations and explicitly severs heterophilic connections for improved reconstruction. Specifically, øurmethod first transforms multivariate time series into a spatial-temporal joint graph and utilizes graph diffusion to progressively denoise and learn anomaly-free node representations. Through the tailored anomaly-aware graph sparsification and contrastive augmentation, øurmethod effectively captures anomaly patterns and eliminates anomaly-related heterophily correlations on the spatio-temporal joint graph. Extensive experiments on TSAD datasets demonstrate that øurmethod achieves state-of-the-art performance, showcasing the expressiveness of our framework.