Multi-View Group Anomaly Detection

Multi-view anomaly detection is a challenging issue due to diverse data generation mechanisms and inconsistent cluster structures of different views. Existing methods of point anomaly detection are ineffective for scenarios where individual instances are normal, but their collective behavior as a group is abnormal. In this paper, we formalize this group anomaly detection issue, and propose a novel non-parametric bayesian model, named Multi-view Group Anomaly Detection (MGAD). By representing the multi-view data with different latent group and topic structures, MGAD first discovers the distribution of groups or topics in each view, then detects group anomalies effectively. In order to solve the proposed model, we conduct the collapsed Gibbs sampling algorithm for model inference. We evaluate our model on both synthetic and real-world datasets with different anomaly settings. The experimental results demonstrate the effectiveness of the proposed approach on detecting multi-view group anomalies.