UAST-HGNN: Accounting-Semantic Temporal Heterogeneous Graph Learning for Fraud-Ring Detection
ID:29
Submission ID:185 View Protection:ATTENDEE
Updated Time:2026-10-10 19:49:51
Hits:225
Online
Abstract
Fraud-ring detection in enterprise ledgers requires integrating relational and temporal evidence without discarding accounting meaning. UAST-HGNN is therefore formulated as a shared heterogeneous graph architecture in which accounting-semantic journal states, ER-conditioned relation messages, temporal transitions, and ring summaries are fused into one calibrated decision path with member- and edge-level explanations. Evaluation is conducted on HeteroLedger under label-blind candidate discovery, fixed train/dev/calibration partitions, five-seed repetition, and a held-out 25,000-candidate test set. Under this protocol, PR-AUC and ROC-AUC are 0.8735 ± 0.0108 and 0.9194 ± 0.0101, respectively. The highest baseline mean PR-AUC is 0.7743 ± 0.0177 for TGAT. At the 20% ordered prefix, 0.7045 ± 0.0170 PR-AUC is retained, indicating that useful evidence is accumulated early in the observed journal sequence. The supported interpretation of the shared score is further delimited by accounting and ER interventions together with deletion-based fidelity tests.
Keywords
Financial fraud detection,heterogeneous graphs,Temporal graphs,Accounting semantics,Explainable graph learning
Submission Author
Anh Khoa Nguyen
Posts and Telecommunications Institute of Technology
Minh Tuan Nguyen
Posts and Telecommunications Institute of Technology
Minh Tam Nguyen
Posts and Telecommunications Institute of Technology
Comment submit