Multi-view anomaly detection via hybrid instance-neighborhood aligning and cross-view reasoning

Multi-view anomaly detection aims to identify anomalous instances whose patterns are disparate across different views, and existing works usually project the multi-view data into a common subspace for abnormal instance identification. Nevertheless, these methods often fail to explicitly excavate the inter-view dependency and discrepancy among the multi-view data, which are of crucial importance to detect inconsistent patterns across different views interactively. To address this problem, we propose an efficient multi-view anomaly detection method via instance-neighborhood aligning and cross-view reasoning, which can well parse the inter-view dependency and discrepancy to detect various kinds of anomalous multi-view instances. To be specific, we first utilize the view-specific encoder to project the original data into the latent feature space, in which a novel instance-neighborhood aligning scheme is seamlessly embedded to preserve the consistent neighborhood structures of multiple views and maximize the consistency for the semantically relevant instances, which indirectly enhances the inter-view dependencies. Meanwhile, a cross-view reasoning module is efficiently designed to explore the inter-view dependencies and discrepancies, which can explicitly boost the inter-view correlations and differences to reason the inconsistent view patterns. Through the joint exploitation of the view-specific reconstruction loss, instance-neighborhood aligning loss, and cross-view reasoning loss, different kinds of anomalous multi-view instances can be well detected more reliably. Extensive experiments evaluated on benchmark datasets, quantitatively and qualitatively, verify the advantages of the proposed multi-view anomaly detection framework and show its substantial improvements over the state of the arts. The code is available at: https://github.com/tl-git320/INA-CR.