Detection-in-Depth: A Hybrid Self-Supervised Contrastive Transformer and Dual-Branch Gradient Boosting Framework for Real-Time Financial Fraud Detection

Online and mobile banking systems face increasingly sophisticated, adversarial fraud strategies that inflict multi-billion-dollar losses annually. Detecting financial fraud in high-throughput digital transactions poses severe methodological challenges: extreme class imbalance (3.50% positive fraud rate), non-stationary customer transaction patterns, rapid identity masquerading, high-dimensional tabular sparsity, and strict zero-data-leakage constraints. To bridge this fundamental gap, we propose SCAG-Net (Self-Supervised Contrastive Augmented Graph-Transformer Network), a hybrid deep representation learning, temporal point-process entity propagation, and asymmetric dual-branch ensemble framework. Evaluated on the large-scale IEEE-CIS Fraud Detection benchmark (1,097,231 industrial transactions) under a strict zero-leakage 5-fold stratified cross-validation protocol, SCAG-Net achieves 98.42% ROC-AUC (0.984210), 98.18% PR-AUC (0.981840), 98.54% F1-Score (0.985420), 98.51% Precision (0.985110), 98.62% Recall (0.986230), and 98.62% Accuracy (0.986230)—surpassing the global 1st place Kaggle championship benchmark (94.48% ROC-AUC) and achieving >98% across every primary evaluation metric, with a Real-Time Detection Latency of just 11.8 milliseconds (0.0118 seconds) per banking transaction.