A graph attention-based model for multivariate time series anomaly detection

Abstract

Multivariate time series anomaly detection has significant application value in fields such as financial risk control, industrial manufacturing and smart transport. However, traditional models often overlook inter-variable dependencies, making it difficult to capture anomalous patterns in complex systems. To address this, this paper introduces MTAD-GAT, a novel model based on graph attention mechanisms. By constructing graph structures to model temporal dependencies and feature correlations among variables, and combining forecasting and reconstruction approaches, the model improves both the accuracy and robustness of anomaly detection.

This paper verifies the model performance on three public datasets: SMD, MSL, and SMAP. Through experiments and investigations, we found that the F1 values of the MTAD_GAT model on these three datasets were all higher than those of the traditional methods. To further analyze the effectiveness of the model's internal architecture, this thesis conducted multiple ablation studies to assess the contribution of each component. Results shows that: The dual GAT modules significantly enhance the ability of MTAD_GAT to capture temporal-feature dependencies; The integration of prediction and reconstruction branches improves robustness and anomaly coverage; The upgraded GATv2 performs better than GATv1.