Unsupervised Graph Anomaly Detection on Directed Attribute Network
Haoyang Li, Siwei Wang, Xinwang Liu, Xinbiao Gan · 2024
Attribute networks are commonly used graph structures in complex application domains. With the increasing prominence of security issues, the detection of anomalies in attributed networks has become a widely studied topic. In recent years, anomaly detection methods based on graph deep learning have become increasingly popular. However, current research primarily focuses on undirected attributed networks. These methods fail to fully utilize the directional information of edges when facing directed networks which are widely present in reality, resulting in suboptimal performance. In order to tackle this problem, we propose a new self-supervised framework for anomaly detection in directed attributed networks, referred to as UDGAD. Our framework includes a novel bidirectional subgraph constructing algorithm that distinguishes the direction of edges, effectively leveraging the complex topological structure of directed networks. Furthermore, we propose a self-supervised model based on graph autoencoder and graph neural networks, which explores anomalies from the perspectives of attribute reconstruction and neighborhood matching. Finally, we evaluate anomaly degree of all nodes by multi-round computational evaluation. Experimental results demonstrate that our method outperforms traditional anomaly detection methods on directed attributed networks with ground truth labels.