Graph Foundation Model for Cross-granularity Anomaly Detection

Graph Foundation Models (GFMs) for Graph Anomaly Detection (GAD) have gained massive attention, playing important roles in academic and industrial circles. Existing models are mainly divided into two types: granularity-specific cross-domain models (GSCD) and cross-granularity domain-specific models (CGDS), which focus on improving cross-domain ability or cross-granularity ability. However, they fail to explore GAD under cross-granularity cross-domain settings. The aim of this study is to provide a unified solution for GFMs in cross-granularity GAD scenarios. To solve this problem, we propose a novel framework, CgGFAD, specializing in two challenges in this problem. On the one hand, to capture granularity-specific semantics and enhance cross-granularity detection ability, CgGFAD unifies the representations for different granularities and employs maximum mean discrepancy based on the Wasserstein distance to evaluate the similarity between granularities. On the other hand, to capture comprehensive granularity-specific representations and achieve cross-domain generalization, CgGFAD develops a cluster-based self-distillation and prototype generalization module, which leverages deep clustering strategies to derive latent semantic patterns and deploys prototype approximation to achieve cross-domain detection. Extensive experiments on diverse public datasets, compared with state-of-the-art baselines, demonstrate the advancement and effectiveness of CgGFAD.