Topology-Preserved Information Bottleneck for Multiview Anomaly Detection.

Abstract

Anomaly detection (AD) techniques are widely used in various fields. Existing techniques primarily focus on learning a normal region from single-view data, which may be not suitable for multiview data that provides more comprehensive information from multiple perspectives. Therefore, AD techniques designed for multiview data are necessary. Straightforwardly, one can concatenate the features learned from multiple single-view data into a joint representation to conduct AD. However, this may overlook the inevitable overlaps between views, potentially masking view-specific information due to the repetitive calculations of these overlaps. Among the various possible methods, one way to address this is to compress redundant information while maintaining comprehensive information across views. Following this way, in this article, we leverage the principle of information bottleneck (IB) to extract concise and comprehensive representations for multiview data. But it is problematic to directly use these representations for AD, since the multiview fusion process may disturb the intrinsic structure of the original data. That is, samples distributed at the edges/center of the original normal data distribution are mapped closer to the center/edges. This might cause abnormal samples (close to the normal data at the edges) to be incorrectly mapped into the normal region during inference. In the AD scenario, the absence of abnormal training samples makes it unfeasible to preserve this structure using supervised information. In this article, we design a topology-preserved regularization that unsupervisedly constrains the latent representations to preserve the original data's intrinsic structure, to improve the AD performance. Overall, we propose a topology-preserved multiview information bottleneck (TMVIB) feature extraction method to extract concise, comprehensive, and topology-preserved latent representations from multiview data. Interestingly, we find that the TMVIB feature extraction method itself can be viewed as a regularized anomaly detector, allowing it to output anomaly scores directly. Experiments on synthetic and real-world multiview datasets demonstrate the effectiveness of the proposed TMVIB.