SMOTE-Stack-XAI: An Explainable Stacked Ensemble Learning Framework Integrating Random Forest, XGBoost, SVM and Deep Neural Networks for Real-Time Credit Card Fraud Detection

Abstract

Credit cards are now the primary tool for digital transactions due to the quick expansion of e-commerce and cashless payment ecosystems. As a result, fraudulent activity has grown proportionately, resulting in significant financial losses for banks, retailers, and cardholders. On the benchmark European cardholder dataset, the authors' previous work investigated single-model Random Forest (RF) classification, a hybrid RF-SVM voting scheme, and a hybrid SVM-Logistic Regression (LR) scheme, with accuracies of 96.6%, 98.0%, and 98.2%, respectively. Despite their effectiveness, these methods are nonetheless susceptible to the dataset's high class imbalance (0.17% fraud), depend on a tiny number of base learners, and offer fraud analysts and regulators no way to evaluate their conclusions. By proposing SMOTE-Stack-XAI, a stacked ensemble framework that (i) uses the Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance, (ii) combines four heterogeneous base learners—Random Forest, XGBoost, Support Vector Machine (SVM), and a Deep Neural Network (DNN)—through a Logistic-Regression meta-learner, and (iii) integrates SHapley Additive exPlanations (SHAP) to reveal the transaction-level features that influence each fraud decision. The suggested framework is assessed using Accuracy, Precision, Recall, F1-score, and AUC on the same dataset of 284,807 European cardholder transactions. According to experimental results, SMOTE-Stack-XAI outperforms the RF, hybrid RF-SVM, and hybrid SVM-LR baselines with an accuracy of 99.3%, precision of 99.2%, recall of 99.4%, and F1-score of 99.3%. It also produces human-interpretable, feature-level explanations for each prediction, making the model appropriate for real-time, auditable fraud-detection pipelines.