Multisource Fraud Intelligence Using Graph Neural Networks for Merchant, Account, and Transaction Risk Assessment

Abstract

Fraud intelligence in payment platforms is inherently relational: a transaction belongs to an account, reaches a merchant, and inherits risk from prior behavior on both sides. Yet graph models can underperform strong tabular learners when fraud labels are scarce or when new accounts appear after training. This paper evaluates that trade-off with a leakage-controlled heterogeneous graph experiment on a 10,000-transaction public BankSim extract containing 6,021 customer accounts, 899 merchants, 16 categories, and 78 frauds. Transactions are ordered by simulation step; steps 1–110 train the models, steps 111–145 select thresholds and graph-fusion weight, and steps 146–180 form a locked future test with 2,000 transactions and 16 frauds. Causal account, merchant, and pair histories are computed before each transaction. A two-layer inductive GraphSAGE encoder is trained on the account–merchant graph and combined with a causal XGBoost score. Standalone GraphSAGE is unstable and weak (test 0.0746), whereas transaction-only XGBoost reaches 0.4251. Validation-selected graph fusion improves to 0.4500 and F1 to 0.4103 while detecting 8/16 frauds with 15 false alerts at the frozen operating threshold. Aggregated risk reaches 0.4502 for accounts and 0.4654 for merchants. The results support graph representations as a complementary relational signal, not a guaranteed replacement for transaction models.