Performance Analysis of Cost-Sensitive Oversampling Mechanisms for Credit Card Fraud Detection

Credit card fraud detection presents a persistent challenge due to severe class imbalance and the high costs associated with undetected fraudulent transactions. While oversampling techniques such as Synthetic Minority Oversampling Technique (SMOTE) and Adaptive Synthetic Sampling Approach for Imbalanced Learning (ADASYN) have improved minority class detection, most methods are cost-agnostic and do not account for the financial impact of individual misclassifications. This paper proposes a comprehensive framework for cost-sensitive oversampling, wherein the probability of generating synthetic minority samples is proportional to the transaction amount, thereby prioritizing high-impact fraud cases. We systematically evaluate standard and cost-sensitive variants of several popular oversampling methods—including SMOTE, ADASYN, Borderline-SMOTE, SMOTE-ENN, and SMOTE-Tomek—across five widely used machine learning classifiers Random Forest, XGBoost, LightGBM, multi-layer perceptron, and logistic regression. Experiments conducted on a real-world credit card dataset demonstrate that cost-sensitive oversampling substantially increases total cost savings and reduces the average cost per misclassification, with especially marked improvements for neural and linear models. The findings offer practical guidance for deploying cost-sensitive solutions in imbalanced, high-stakes financial applications.