Role-GAD: Role-Aware Graph Anomaly Detection via Dual-Codebook Vector Quantization

Graph anomaly detection (GAD) is fundamental to identifying fraudulent and abnormal behaviors in financial and social networks. Unlike node classification, where classes often form relatively coherent distributions, we empirically show that graph anomalies arise from heterogeneous mechanisms and remain highly fragmented, whereas normal data exhibit fewer recurring semantic and structural patterns. These findings motivate a prototype-based representation principle: recurring normal patterns should be summarized through a compact set of semantic and relational roles, while heterogeneous anomaly mechanisms remain distinguishable through rare role assignments and atypical role interactions. Based on this principle, we propose \model, a role-aware framework that abstracts node semantics and edge interactions into dual discrete role codebooks and performs role-aware message passing over a role-induced graph augmented with global role-triplet statistics. Extensive experiments on five real-world benchmarks demonstrate that Role-GAD achieves the best average performance while converging up to $10\times$ faster than the strongest accuracy-oriented baseline. Our code and data are available at https://anonymous.4open.science/r/role-gad-5932.