A Cost-Sensitive Machine Learning Framework for Optimizing Financial Loss Mitigation in Credit Card Fraud Detection

Abstract

Credit card fraud is the most common type of cybercrime due to the shift in digital economy which caused billions of dollars in losses annually, the deployment of standard machine learning algorithm is disturbed by a heavily imbalanced dataset. in this study, an analysis between four machine learning algorithms is conducted which are Logistic Regression, Random Forest, Artificial Neural Network (MLP) and XGBoost, to mitigate the data imbalance problem, this study used a Cost Sensitive Learning Framework. which consists of 284,807 transactions, using 10-fold cross-validation, the framework integrates post-hoc probability calibration via Platt Scaling and a 99-step training search grid to dynamically optimize decision thresholds (τ*) against operational friction penalties (1.5% interchange fee loss plus a $0.10 administrative fee per False Positive), Model evaluations prioritize Area Under the Precision-Recall Curve (AUPRC), inference throughput latency, and net financial savings. Experimental results demonstrate a clear operational trade-off between fraud loss mitigation, customer friction penalties, and inference speed. Tuned Random Forest achieved the leading economic protection, securing an AUPRC of 84.86%±4.09% and a maximum Net Financial Savings of $45,277.21 (75.30% financial efficiency) by effectively suppressing False Positive penalties ($95.88). Hardware latency benchmarking reveals that tree-based ensembles incur higher traversal overhead, with Random Forest recording the highest per-sample latency of 1.473 μs (678,759 transactions/sec), whereas Logistic Regression (0.209 μs), ANN (0.955 μs), and XGBoost (0.601 μs) demonstrated superior throughput. Nevertheless, all evaluated candidate models comfortably exceed 600,000 transactions/sec, proving that all proposed architectures-including Tuned Random Forest-comfortably satisfy real-time high-throughput banking Service Level Agreements (<100 ms).