Rule anchored graph learning improves synthetic identity fraud detection through multi relation structure rather than score level fusion

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Synthetic Identity Fraud (SIF), in which fraudsters combine real and fabricated information to create untraceable identities, is a fast-growing and costly problem in payment systems. Prior work treats graph neural networks as an assumed improvement over transparent rule-based identity-linkage scoring, but rarely tests where, if anywhere, that improvement originates. We build a heterogeneous Client-SSN-Email-Phone graph from a PaySim-derived identity linkage dataset (2,433 profiles, 433 fraudulent) and compare rule-based, non-graph machine learning, non-deep propagation, and deep graph models across five identity-noise levels and four random seeds. A rule-anchored hybrid combining the rule score with a Heterogeneous Graph Transformer (HGT) and support vector machine (Rule + HGT+SVM) achieves the best average Precision-Recall Area Under the Curve (PR-AUC) of 0.835, versus 0.777 for Rule-only, with a positive difference across all 20 seed-noise combinations tested. Because rule and graph scores rank clients almost identically, we test whether two more sophisticated fusion mechanisms, an adaptive per-client weight and a tie-breaking scheme, can extract further gains; although reordering tied rule scores can in principle affect rank-based metrics, neither mechanism produced consistent gains over the simple fixed fusion, since 99.5% of clients share their rule score with another client. A relation-level ablation instead shows the graph’s value is structural: removing SSN, Email, or Phone individually from the graph consistently reduces average performance across all tested seed-noise configurations, and a density-matched control confirms this is not merely a graph-sparsity artifact (nominal p ≤ 0.0007 for all three in the descriptive 20-pair comparison). What matters is having some relations represented completely rather than all relations represented but thinned. These results show where graph learning’s contribution actually comes from in SIF detection: not from how a fused score is weighted, but from the completeness of the identity relations the encoder can access. Transparent rule-based scoring remains the necessary backbone of a defensible detection pipeline.