Decoupled focal loss and bayesian temperature fusion XGBoost for imbalanced credit card fraud detection

Abstract

Accurate credit card fraud detection is crucial for financial system security. Under class imbalance, existing models struggle with the precision–recall trade-off: traditional Focal Loss relies on static parameters that cannot adapt to sample difficulty, conventional fusion strategies are parameter-sensitive and prone to extreme weight assignments, and a single model often fails to capture both complex nonlinear patterns and robust structural features. This study proposes the FL-XGB Fusion framework, which adopts a decoupled warm-up scheduling strategy for Focal Loss—an adaptive class balancing coefficient α targeting the empirical fraud rate, and a progressively increasing focusing parameter γ preventing premature suppression of hard examples—together with a Bayesian temperature-optimized adaptive fusion network that quantifies prediction uncertainty via information entropy, dynamically allocates DNN and XGBoost weights, and employs TPE sampling to search for the optimal temperature per fold. Experiments on three heterogeneous datasets (IEEE-CIS, CreditCard, and PaySim) under five-fold stratified cross-validation show that the framework achieves the highest F1-Score among eight baseline methods on all three datasets, improving over standalone XGBoost by 1.51–1.81 pp, while remaining competitive in AUC-ROC. Ablation studies with nine variants and non-parametric tests quantify the contribution of each core component, indicating that the gains, though modest in magnitude, are consistent across the three datasets with different imbalance ratios.