Context-Aware Anomaly Detection in Attributed Networks

Abstract

Anomaly detection in attributed networks has received increasing attention due to its broad applications in various high-impact domains. Compared to traditional anomaly detection, the main challenge of this task lies in how to integrate the network structure and node attributes to spot anomalies. However, existing methods attempt to integrate two kinds of information into a fixed representation and neglect the contextual information. Specifically, a fixed feature vector is directly adopted to evaluate its abnormality without considering the node’s diverse roles when interacting with different neighbors. In this paper, we propose a novel Context-Aware Anomaly Detection (CAAD) framework in attributed networks. CAAD derives context-aware embeddings for each node pair with a mutual attention mechanism. The embeddings extracted by feature interactions can concentrate on the most relevant attributes of network structures. Numerous context information provides us with multiple perspectives to understand the structure connection and detect local anomaly structure. Moreover, we develop an anomaly gated mechanism to assign global anomalous scores to node pairs. The anomalous scores are learnable and applied to reduce the adverse effect of anomalies during the training process. By jointly optimizing network embeddings and anomaly gated mechanism, our model can spot anomalies in local and global collaborations. Experiments on various real-world network datasets indicate that the proposed model achieves state-of-the-art results.