Self-attention-based graph transformation learning for anomaly detection in multivariate time series

Abstract

Multivariate time series anomaly detection has widely applications in many fields such as finance, power, and industry. Recently, Graph Neural Network (GNN) have achieved great success in this task due to their powerful ability of modeling multivariate relationships. However, most existing methods employ shallow networks with only two layers, resulting in restricted node information transfer range and limited sensing field. In this paper, we propose a self-attention based graph transformation learning (AT-GTL) method to solve this problem. AT-GTL uses a global self-attention graph pooling (GATP) module to aggregate all node features to obtain global features. Then, a graph transformation learning pipeline is constructed based on neural transformation learning, and a triplet contrastive loss (TCL) is constructed to optimize the global feature extraction networks using potential features from multi-viewpoints. Extensive experiments on three real-world datasets demonstrate that our method can effectively aggregate global graph features and detect anomalies, providing a new transformation learning solution for multivariate time series anomaly detection.