Tailored Bipartite Graph Anomaly Detection via Fixed-Attention Network

Shaowen Tang, Raymond K. Wong

Abstract: In recent, Graph Neural Networks (GNNs) have become increasingly popular in the field of graph anomaly detection due to the capability of learning expressive representations. Nevertheless, only few studies pay attention to bipartite graphs in practice, where the traditional GNNs show limited performance on anomaly detection due to the invalidation of homophily assumption. To address this issue, in this paper, we propose a novel end-to-end framework tailored for bipartite graph anomaly detection namely BiG-FAN, revealing the intrinsic heterophily between nodes and leveraging such information to achieve judicious message aggregation process via a fixed-attention network. In addition, an attribute-based prediction module is included to enhance the performance by exploiting semantic features. Empirical experiment results on real-world datasets demonstrates that the proposed framework outperforms the state-of-the-art baselines on bipartite graph anomaly detection.