Credit Card Fraud Detection with ADASYN Oversampling and SHAP-based Interpretability: A Comparative Ensemble Approach

Credit card fraud continues to be a significant threat to financial systems, exacerbated by the highly imbalanced nature of transaction datasets and the opaque decision-making of complex machine learning models. This paper proposes a hybrid fraud detection framework that integrates Adaptive Synthetic (ADASYN) oversampling to address class imbalance and SHAP (SHapley Additive exPlanations) to enhance model interpretability. Five machine learning classifiers Logistic Regression, Random Forest, XGBoost, LightGBM, and Multilayer Perceptron—are evaluated on the widely used Kaggle credit card fraud dataset. ADASYN significantly improves the minority class representation in the training set, enabling models to achieve higher fraud recall without overwhelming false positives. Among the models tested, Random Forest delivered the best trade-off between precision (85.7%) and recall (79.6%), achieving an F1-score of 82.5% and ROC-AUC of 0.9633. SHAP analysis provided granular insight into feature contributions, transforming black-box predictions into transparent and auditable decisions. Comparative analysis with eight state-of-the-art studies demonstrates that while recent approaches often report near-perfect results, the proposed model strikes a balance between predictive performance, computational efficiency, and interpretability qualities essential for practical deployment in financial fraud detection systems based on benchmark transactional data. The study highlights that integrating ADASYN with ensemble learning and SHAP can create a robust, explainable, and scalable fraud detection system suitable for deployment in dynamic financial environments.