RECAP: Residual Community Adhesion Patterns for Generalist Graph Anomaly Detection

Generalist graph anomaly detection aims to identify anomalous nodes in unseen graphs without target-specific retraining. Existing approaches often rely on source-graph labels, target prompts, or coarse normal/anomalous prototypes, which may be unreliable when labels are unavailable and both normal and anomalous patterns are diverse. This paper proposes **RECAP**, a framework based on **RE**sidual **C**ommunity **A**dhesion **P**atterns for label-free cross-graph detection. Rather than transferring a binary anomaly boundary, the method learns a residual-pattern mapping by aligning soft community assignments with similarity graphs constructed from high-order ego-neighbor residual representations. During inference, predicted assignments induce community prototypes, and anomalies are ranked by prototype adhesion deviation with an auxiliary neighborhood community-context inconsistency term. This formulation captures diverse residual patterns while preserving a decomposed score that attributes anomalies to prototype deviation, local community-context mismatch, or both. Cross-dataset transfer experiments show that, without source labels, target prompts, or target fine-tuning, RECAP achieves competitive anomaly-ranking performance under multiple source–target splits. Beyond ranking accuracy, the framework provides model-intrinsic diagnostics through community prototypes, decomposed adhesion/context scores, and node-level reports, offering structured behavioral evidence beyond a scalar anomaly score.