Capturing Collaborative Structure in Graph Anomaly Detection by MoE-Gated Neural Common Neighbor

Abstract

In real-world graphs, anomalies usually exhibit collaborative structure, manifested as densely connected sub-graphs. Capturing multi-hop relationships is essential to identify and characterize the connections among these anomalous nodes. Existing graph anomaly detection (GAD) methods fail to capture such structures due to two core challenges: 1) standard recursive neighbor aggregation paradigm causes over-squashing of high-order neighbor information, which is vital for detecting collaborative suspicious behaviors; 2) different collaboration patterns depend on specific hop counts, with irrelevant hops introducing substantial noise. In this paper, we propose MoE-Gated Common Neighbor Attention (MCNA), an auxiliary parallel branch of the Graph Neural Network based (GNN-based) GAD backbones that can effectively distinguish collaborative structure to enhance the backbone's performance. MCNA explicitly encodes neural common neighbor information to capture the arbitrary k-hop collaborative structure among anomalies. To balance the information and noise from multi-hop neighbors, an MoE mechanism is employed to choose the valuable hops in terms of structural information. The structural collaborative patterns captured by MCNA and the attribute-based patterns captured by backbones are finally fused with a gating mechanism. Extensive experiments on 10 real-world datasets and 8 different backbones demonstrate that MCNA achieves average improvements of 5.36% on AUROC and 7.76% on AUPRC, showcasing the comprehensive universality of our proposed method.