Graph Neural Network-Based Anomaly Detection in Multivariate Time Series

Given high-dimensional time series data (e.g., sensor data), how can we detect anomalous events, such as system faults and attacks? More challengingly, do this in a way that captures complex inter-sensor relationships, detects explains anomalies which deviate from these relationships? Recently, deep learning approaches have enabled improvements anomaly detection datasets; however, existing methods not explicitly learn the structure of relationships between variables, or use them to predict expected behavior series. Our approach combines with graph neural networks, additionally using attention weights provide explainability for detected anomalies. Experiments on two real-world datasets ground truth show our method more accurately than baseline approaches, correlations sensors, allows users deduce root cause anomaly.