Machine learning study demonstrates enhanced fraud detection and adversarial robustness across transaction graphs, highlighting the value of causal message filtering over correlation.
Key Points
Develop and evaluate the Causal Flow Graph Network (CFGN), a model designed to overcome adversarial camouflage and class imbalance by removing non-causal confounded connections during message passing.
Designed CFGN with a causal gate approximating Pearl's backdoor adjustment criterion combined with structural attention via a learned adaptive mixture coefficient (lambda).
Evaluated performance across three diverse fraud detection benchmarks: the Elliptic Bitcoin network, YelpChi anti-spam reviews, and Amazon behavioral fraud, testing under standard, adversarial edge injection, and temporal drift conditions.
CFGN achieved a cross-validated AUC-PR of 0.6567 and AUC-ROC of 0.8868 on the Elliptic Bitcoin dataset, surpassing seven baseline and state-of-the-art graph neural networks.
The causal mixture coefficient dominated specifically in adversarially camouflaged domains (Bitcoin lambda: 0.71 ± 0.02; YelpChi lambda: 0.73 ± 0.04) versus heuristic labels (Amazon lambda: 0.39 ± 0.03).
Under 0.4 ± 0.0 adversarial edge injection on YelpChi, CFGN maintained 75.0% average precision compared to 50.0% for GAT and 53.0% for GCN, and experienced only an 8.6% drop under temporal concept drift compared to a 53.6% drop for GCN.