Credit Card Fraud Detection: Data Balancing Techniques and Model Interpretability

Credit card fraud detection is a crucial task in fi-nancial institutions to safeguard against fraudulent transactions. This study investigates the performance of Random Forest and Gradient Boosting models on a Nigerian credit card fraud detection dataset. Key features were selected using Chi-Square and Analysis of Variance (ANOVA) tests, and the models were evaluated across four data balancing techniques: no balancing (imbalanced), under-sampling, over-sampling, and hybrid sampling. The primary focus was on recall, as misclassifying fraudulent transactions can have significant financial implications. Gradient Boosting consistently outperformed Random Forest, achieving higher Recall, F1-Score, and AUC-ROC values across all experiments, particularly with over and hybrid sampling. To enhance interpretability, SHapley Additive exPlanations (SHAP) were applied to the Gradient Boosting model. SHAP analysis revealed that the amount of the transaction and average income expenditure were the most influential features in predicting fraud. Local and global explanations highlighted how specific feature values impacted the model’s decisions, while dependency plots uncovered interaction effects between key features. These findings contribute to the advancement of fraud detection methodologies, offering insights into effective strategies for mitigating financial risks associated with fraudulent transactions.