Anomaly Subgraph Detection on Multiple Associated Attributed Networks

With the rapid development of artificial intelligence (AI) technology and large-scale dataset applications, anomaly subgraph detection is becoming an essential and widespread application. Most existing methods detect anomaly subgraphs depending on the observable anomalous attributes in target data. However, they may face challenges when the target data lacks such anomalous attributes. E.g., identifying suspicious gangs in the target transportation graph solely based on road traffic information, or detecting burst events in nonpublic data without explicit anomalous attributes. To address the detection of implicit anomaly subgraphs (IASs) in such target graphs, this article proposes a novel approach IAS detection (IASD) with multidimensional feature transfer. First, our approach is built upon transfer learning techniques and involves fusing features from multiple graphs. Second, we explore a feature extraction procedure based on a graph attention (GAT) network model to generate representations of anomaly features for large multidimensional attribute graphs. Third, we construct a two-layer graph by introducing an attributed graph referred to as a “source graph,” where anomalies can be easily detected, and nodes of the source graph are partially aligned with the target graph. We validate the effectiveness and robustness of IASD through experiments on five real-world datasets, and implement four anomaly subgraph detection tasks in practical applications.