Anomaly Detection via Graph Attention Networks-Augmented Mask Autoregressive Flow for Multivariate Time Series

Anomaly detection in multivariate time series (MTS) has been applied to various areas. Recent studies for detecting anomalies in high-dimensional data have yielded promising results. However, these methods are incapable of explicitly dealing with the complex contextual information that exists between features. In this paper, we present a novel unsupervised anomaly detection framework for MTS. We model the complex relationships of MTS using graph attention networks from the perspectives of time and features, respectively. Furthermore, our framework employs masked autoregressive flow for density estimation, which is then treated as an anomaly score, to identify anomalies. Extensive experiments show that our model outperforms baseline approaches in terms of accuracy on three publicly available datasets and accurately captures temporal and inter-feature relationships.