Effect of Automated Hyper-Parameter Tuning on Deep Belief Network for Credit Card Fraud Detection

Credit card fraud detection (CCFD) is a critical facet in the financial industry as digital payments are becoming increasingly popular. While traditional rule based CCFD often fails due to concept drift and class imbalance challenges, Deep Belief Networks (DBN) can automatically learn complex hidden features but is computationally intensive and requires critical optimal hyperparameters tuning, which can be tuned with walrus optimization algorithm (WOA). This research therefore developed an automated hyperparameter tuning using walrus optimization algorithm to optimize deep belief network model (WT-DBN) for credit card fraud detection system. The research was carried out with dataset of a total of 10,000 real-world credit card transaction records. These data underwent pre-processing, after which a Walrus Optimization Algorithm was developed and applied to enhance the model detection capability by selecting the optimal hyperparameters. Finally, the WT-DBNs model was applied to credit card fraud detection, and implemented in MATLAB R2023a. At the highest training ratio of 80% for training and 20% for testing, Walrus tuned-DBN achieved a FPR, Sensitivity, Specificity, Precision, F1-Score, accuracy and detection time of 9.50, 97.5%, 90.5%, 97.62%, 96.86%, 96.1% and 25.91s, respectively as compared to 12.17 96.83%, 87.83%, 96.95%, 95.98%, 95.03% and 33.95s respectively for standard DBN. The best fitness value of Walrus tuned-DBN was 0.05883. The comparative results of the standard (DBN) and WT-DBN, across different data divisions shows that Walrus Optimized tuned-DBN(WT-DBN) enhances better convergence and improved classification performance compared to the conventional DBN.