Substructure-Guided Graph-Level Anomaly Detection with Attention-Aware Aggregation
摘要
Graph-level anomaly detection aims at identifying anomalous individual graphs that deviate significantly from others within a graph set. Typically, previous studies mainly aggregate node or triple (e.g., ) representations to obtain graph-level representation for anomaly detection. However, graph-level anomalies usually exhibit in more complex structures, making it difficult to perceive anomalies from the perspectives of nodes or triples. Moreover, simple aggregation strategies (e.g. global averaging) smooth out the key anomalous features, resulting in the non-discriminative graph-level representation. To this end, we propose a novel method, named SubGAD, short for Substructure-guided Graph-level Anomaly Detection with attention-aware aggregation. Specifically, SubGAD designs an enhanced substructure representation learning, which leverages soft clustering to adaptively partition the graph into substructures for capturing complex anomalous structures. Furthermore, SubGAD devises an attention-aware feature aggregation strategy, which perceives the significance of features based multi-head attention mechanism. Finally, a high-quality graph-level representation is learned by SubGAD for graph-level anomaly detection. Extensive experiments on 13 real-world datasets demonstrate the superiority of SubGAD compared with the state-of-the-art methods. Our codes are available at https://github.com/scu-kdde/OAM-SubGAD-2024 .