HAN-CAD: hierarchical attention network for context anomaly detection in multivariate time series

Anomaly Detection in multivariate time series (MTS) plays an important role in many real-world Web services such as the Web traffic monitoring system. With abundant MTS data, exploiting the relationships among different variables, i.e., inter-variable relationships, is crucial for detecting anomalies. Recent studies have made substantial efforts to promote relationship learning from graph neural network. However, existing methods mostly neglect the distinctive characteristics of inter-variable relationships under different contexts, i.e., dynamics of inter-variable relationships. Therefore, we propose a “Hierarchical Attention Networks for Context Anomaly Detection” (HAN-CAD) model to fully exploit the inter-variable relationships and their dynamics. More concretely, we model each time series segment (context sequence) as a graph, where variables in the sequence are nodes and edges denote correlation patterns among variables. Then, the first graph attention layer is built on this graph to obtain the variable representation, which captures the relationships among different variables. Thereafter, the second attention layer outputs the sequence representation by integrating inter-variable relationships within the current context sequence. Finally, anomalies can be detected based on the reconstruction model, i.e., AutoEncoder. Extensive experiments on real-world datasets demonstrate that the proposed method can effectively detect anomalies in MTS and outperforms recent state-of-the-art methods.