LAF: A Local-Affinity-Based Adaptive Filter for Unsupervised Graph Anomaly Detection

Graph neural networks (GNNs) have achieved remarkable success in unsupervised graph anomaly detection (GAD) tasks. The existing GNN-based unsupervised GAD models typically employ self-supervised learning to capture the intrinsic low-dimensional representations of data, thereby adapting to the inductive bias of GNNs toward homophily. However, they often ignore the anomaly-discriminative property of nodes, termed the one-class homophily property, i.e., normal nodes tend to have strong affinity with each other, while the homophily in anomalous nodes is significantly weaker than that in normal nodes. In this paper, motivated by the one-class homophily, we propose a novel local-affinity-based adaptive graph filter (LAF) to address label imbalance challenge in unsupervised GAD tasks. Our method generates subgraphs by removing heterophilic edges from the raw graph, thereby strengthening the isolation of anomalous nodes and enhancing high-frequency information of the graph structure. Subsequently, we apply an adaptive graph filter to each subgraph to dynamically integrate low-frequency and high-frequency information, so as to better model the anomalous features. Instead of minimizing the commonly used data reconstruction errors, our method optimizes the model by maximizing the local node affinity. The final result is obtained by averaging the anomalous scores from multiple models trained on different subgraphs, during the inference phase. Experimental results on six real-world GAD datasets show that LAF outperforms the existing baseline algorithms in unsupervised GAD tasks.