Multivariate Time Series Anomaly Detection in a Regularization Perspective

Real-world scenarios such as Internet, industrial equipment and finance field generate a large number of multivariate time series all the time which are important for describing the operational state of a system. Therefore, anomaly detection on the multivariate time series has become a hot topic today. How to utilize regularization to eliminate overfitting is an important issue since it inhibits the representative power of existing models. In this paper, a reconstruction model called Autoregressive Graph Adversarial Network (ARGAN) is proposed. First, we develop a latent space reconstruction strategy to guarantee ARGAN's representative ability for the key features. Then, the autoregressive regularization using temporal dependency is proposed to inhibit overfitting. Finally, a regularized annealing strategy is designed to balance reconstruction and regularization. The proposed model can achieve better performance on four real-world datasets compared with other six algorithms.