Enhancing Credit Card Fraud Detection under Severe Class Imbalance using Cost-Sensitive Learning and Threshold Optimization

INTRODUCTION: Credit card fraud detection remains challenging because fraudulent transactions are rare, fraud patterns evolve over time, and precision–recall trade-offs vary under different class prevalences. While synthetic oversampling methods such as SMOTE are widely used, they may alter the observed feature distribution and complicate deployment interpretation. OBJECTIVES: This study develops a cost-sensitive and decision-threshold-calibrated fraud detection framework that preserves the original data distribution under severe class imbalance. METHODS: The framework combines cost-sensitive XGBoost and a Multi-Layer Perceptron trained with weighted Binary Cross-Entropy and Focal Loss. Hyperparameters are tuned using Optuna within a leakage-conscious validation protocol, and class imbalance is handled through scaleaware weighting rather than synthetic resampling. Decision thresholds are selected using minority-class F1 and an illustrative amount-aware cost criterion. SHAP analysis and a chronological split of the 2013 dataset are used to examine transformed-feature auditability and near-future generalization. RESULTS: On the imbalanced 2013 dataset, the optimized XGBoost model improves test-set PR-AUC from 0.7809 to 0.8815 and minority-class F1 from 0.7919 to 0.8497, with false positives decreasing from 21 to 13 on a test partition containing 98 fraud cases. On the balanced 2023 dataset, ranking performance is near-saturated and threshold selection mainly reduces false alarms. Under the illustrative assumption CFP = 1, amount-aware thresholding yields a lower simulated cost than the F1-optimized threshold, but with substantially more false positives. Chronological validation on the 2013 dataset yields lower PR-AUC than random stratified evaluation. CONCLUSION: The results suggest that fraud detection under class imbalance benefits from separating ranking optimization from decision-threshold selection. However, cost-aware and temporal findings should be interpreted as benchmark-based, deployment-motivated analyses rather than direct evidence of operational deployment readiness.