Graph-enhanced anomaly detection framework in multivariate time series using Graph Attention and Enhanced Generative Adversarial Networks

Abstract

The complexity and scale of distributed systems in cloud computing present significant challenges for effective time series anomaly detection, which aims to identify unusual patterns in time series data that deviate from expected behavior. Traditional anomaly detection technologies in this domain suffer from high false positive rates. This challenge arises from the difficulty of balancing high recall rates with the reduction of false positives, which are both essential for ensuring operational integrity and user satisfaction in cloud environments. To address these challenges, this paper presents the Efficient Hybrid Graph Attention Mechanism and Enhanced Generative Adversarial Network (EH-GAM-EGAN), an innovative unsupervised model tailored for multivariate time series anomaly detection in cloud computing networks. First, EH-GAM-EGAN utilizes a graph attention mechanism combined with Long Short-Term Memory networks to effectively capture and analyze complex node relationships, thereby improving the understanding of data interdependencies. Second, it integrates an enhanced generative adversarial network, which precisely computes reconstruction and discrimination errors. This approach facilitates a thorough analysis of anomalies by examining reconstruction, discrimination, and prediction errors, resulting in significantly improved detection accuracy and model reliability. Extensive experiments on four publicly available cloud computing datasets empirically validated the effectiveness of EH-GAM-EGAN. The results show that EH-GAM-EGAN achieved average improvements of 17.93%, 17.88%, and 21.46% in precision, recall, and F1 scores, respectively.