Uncertainty-Aware Credit Card Fraud Detection Using a Fuzzy Graph Attention Network with Weak-Signal Edge Preservation

Abstract

Coordinated card fraud can leave only weak, partial traces between transactions, which crisp graph detectors discard. This study evaluates a Fuzzy Graph Attention Network (Fuzzy-GAT) in which Gaussian edge-membership weights modulate attention, so that weak-signal edges are preserved rather than pruned. Because the public European Credit Card Fraud benchmark contains no account, device or IP identifiers, the graph links each transaction to its ten nearest neighbours in feature space, and no relational entities were synthesised. After removing 1,081 duplicate records (283,726 transactions; 473 frauds), every model was evaluated over five stratified 70/10/20 splits, at both the default and a validation-optimised decision threshold. The Fuzzy-GAT reached an F1-score of 79.3 ± 2.9% and a ROC-AUC of 0.969 ± 0.012. It matched Logistic Regression (79.4 ± 1.8%) and a crisp Graph Convolutional Network (79.9 ± 3.0%), but Random Forest (82.1 ± 3.3%; paired t-test p = 0.014) and XGBoost (82.8 ± 2.4%; p = 0.002) outperformed it. In a three-seed ablation, fuzzy weighting gave no benefit, and preserving weak-signal edges changed F1 by −0.1 percentage points; pruning weak edges at thresholds from 0.05 to 0.80 left F1 unchanged. In a model-agnostic stress test built from perturbed copies of test frauds, no model resisted camouflage: at moderate strength the Fuzzy-GAT detected 47.2% of cases, against 56.4% for Logistic Regression. On this benchmark, fuzzy edge weighting added computational cost without measurable benefit; whether it helps where genuine relational metadata exist remains open.