A Fusion Approach for Credit Card Fraud Detection: A Stacked Ensemble Framework

Credit card fraud has become a pressing issue with the increase in digital transactions, leading to significant financial losses. To address this, our study introduces a stacking ensemble framework for fraud detection using deep learning models, LSTM-RNN, GRU, and CNN as base models, with Logistic Regression as the meta-model. We tackle the challenge of data imbalance by applying sampling techniques, including SMOTE, ADASYN, random undersampling, NearMiss, and condensed nearest neighbor. Our findings reveal that oversampling methods, specifically SMOTE and ADASYN, considerably improve model performance, enhancing the system's ability to detect fraudulent transactions. This demonstrates the necessity of oversampling strategies to build effective and reliable fraud detection models, especially when dealing with highly imbalanced datasets. Our research underscores the importance of employing advanced ensemble methods and strategic sampling to improve fraud detection accuracy and mitigate financial risk.