Key points are not available for this paper at this time.
Abstract Time series anomaly detection aims to identify unusual samples that significantly deviate from the majority of the data. The greatest challenge lies in endowing models with the ability to learn effective differentiation criteria. Previous methods have focused on learning intricate, sophisticated representations to overcome these difficulties. However, these approaches are often hindered by spurious correlations, which limit their potential for further improving model performance and causality.To solve the problem, we propose a novel approach, called InfoFlow, which accurately extracts and identifies pertinent details within sequences while excluding anomalous entries. This is achieved by utilizing information bottleneck constraints throughout a transformer-based model. Our primary discovery suggests that anomalous data are rare and unevenly distributed within the observed points, resulting in the phenomenon of anomaly concentration, which complicates the identification. To mitigate this issue, we employ normalizing flows to adjust the sequence distribution to a more uniform range. This approach accentuates subtle variations in the data and enhances the algorithm's sensitivity to anomalies, thereby facilitating easier detection of outlier points. Extensive experiments on real-world datasets demonstrate that our InfoFlowapproach outperforms state-of-the-art benchmarks.
Mo et al. (Thu,) studied this question.