DuoGAT: Dual Time-oriented Graph Attention Networks for Accurate, Efficient and Explainable Anomaly Detection on Time-series

Recently, Graph Neural Networks (GNNs) have achieved state-of-the-art performance on the multivariate time-series anomaly detection task by learning relationships between variables (sensors). However, they show limitations in capturing temporal dependencies due to lack of sufficient consideration on the characteristics of time to their graph structure. Several studies constructed a time-oriented graph, where each node represents a timestamp within a certain sliding window, to model temporal dependencies, but they failed to learn the trend of changes in time-series. This paper proposes Dual time-oriented Graph ATtention networks (DuoGAT) that resolves the aforementioned problems. Unlike previous work that uses the simple complete undirected structure for time-oriented graphs, our work models directed graphs with weighted edges that only connect from prior events to posterior events, and the edges that connect nearby events are given higher weights. In addition, another time-oriented graph is used to model time series stationary via differencing, which especially focuses on capturing the series of changes. Empirically, our method outperformed the existing state-of-the-art work with the highest F1-score for the four real-world dataset while maintaining low training cost. We also proposed a novel explanation method for anomaly detection using DuoGAT, which provides time-oriented reasoning via hierarchically tracking time points critical in a specific anomaly detection. Our code is available at: https://github.com/ByeongtaePark/DuoGAT