acceptodds
Desk rejected at ICLR 2027

TIEGGAD: TARGET-DOMAIN INFORMATION ENHANCEMENT FOR GENERALIST GRAPH ANOMALY DETECTION

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Generalist Graph Anomaly Detection (GGAD) aims to detect anomalies in unseen target graphs using a detector trained on source graphs. However, cross-graph discrepancies may limit the detector’s cross-graph generalization ability, leaving anomaly-relevant information in the target domain insufficiently captured. We summarize this bottleneck as Target-domain Anomaly Information Deficiency (TAID). To mitigate TAID, we propose Target-domain Information Enhancement for Generalist Graph Anomaly Detection (TIEGGAD), a multi-scale structural and cross-view hypergraph enhancement framework for zero-shot GGAD. To strengthen structural modeling and complement semantic information, we construct a degree-guided multi-scale structural representation to capture structural deviations across different neighborhood scales. Furthermore, to refine the overall representation using target-domain information, we introduce a training-free cross-view hypergraph module that enables bidirectional information exchange between the semantic and structural views. Finally, parameterized vector-level integration is used to combine semantic, structural, and hypergraph-enhanced representations into a unified representation. The proposed framework requires neither target-domain labels nor retraining. Experiments on multiple real-world graph datasets demonstrate stable and competitive performance for zero-shot cross-domain node anomaly detection, while ablation and mechanism analyses further verify the effectiveness of multi-scale structural modeling and cross-view hypergraph enhancement.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph anomaly detection, generalist graph anomaly detection, zero-shot graph anomaly detection, cross-graph generalization, hypergraph learning

Related papers

Discussion (0)

Sign in to comment.