Addressing Class Imbalance in Credit Card Fraud Detection: A Hybrid Deep Learning Approach

The rise of credit card fraud is a global concern, demanding reliable detection methods that can overcome challenges with imbalanced datasets and limited exploration of hybrid modeling approaches. This study introduces a hybrid deep learning architecture combining Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) layers alongside SMOTE-TOMEK preprocessing to address imbalanced data issues in credit card fraud detection. The research analyzes a substantial dataset containing both legitimate and fraudulent transactions, evaluating the performance of GRU, LSTM, and the novel Hybrid model through comprehensive data exploration, preprocessing, and feature selection. Performance evaluation uses metrics including accuracy, precision, recall, F1 Score, AUROC, and AUPRC. The experimental results demonstrate the effectiveness of deep learning architectures, with AUROC values of 0.974551 for LSTM, 0.958174 for GRU, and 0.976205 for the Hybrid model. The Hybrid model showed particularly promising results with a precision of 0.9121 and AUPRC of 0.886068, outperforming the individual models. These findings indicate that combining complementary deep learning architectures enhances fraud detection by leveraging their respective strengths in capturing both long-term dependencies and transaction patterns. These insights offer valuable guidance to financial institutions in implementing effective fraud detection systems while emphasizing the importance of continuous improvement of deep learning algorithms to address evolving cyber threats.