Multi-mode Graph Attention-Based Anomaly Detection on Attributed Networks

Attribute network anomaly detection is extensively utilized in areas such as finance, network security, and the Internet of Things. Prevailing anomaly detection methods encounter challenges in following aspects: capturing sparsity, nonlinearity, and ensuring the uniqueness of anomalies. To address these issues, this paper introduces an autoencoder framework built upon multi-mode graph attention networks, which models attribute networks using graph attention networks to capture sparsity and nonlinearity. Secondly, the network encoder utilizes multi-mode graph attention to learn node embeddings, which contain feature information from different neighborhoods of nodes to ensure the uniqueness of anomalies. Subsequently, decoder leverages the acquired node embeddings to rebuild both the topology and node attributes of the attribute network. Finally, anomaly detection is conducted by evaluating the reconstruction errors of attributes and structures. The experimental results on three attribute network datasets demonstrate the framework’s effectiveness.