Walrus Optimization-tuned Deep Belief Network for Credit Card Fraud Detection

Abstract

Detection of credit card fraud (CCFD) has become a critical research area as financial losses continue to increase annually. The traditional rule-based and conventional machine learning model struggles to address the challenges of concept drift and an imbalanced dataset. While deep belief network (DBN) algorithms can learn highly complex features, they require meticulous hyperparameter tuning, which can lead to suboptimal convergence. Hence, this study employs walrus optimization algorithm to automate the DBNs hyperparameters. Ten thousand (10,000) of the imbalance dataset containing 3000 fraudulent and 6000 non-fraudulent transactional datasets were obtained and pre-processed using imputation, min-max, and one-hot encoding techniques. The DBNs were developed as a stack of Restricted Boltzmann Machines (RBMs). The optimised DBNs (WOA-DBNs) were then developed and applied to credit card fraud detection, with data divided 60:40, 70:30, 75:25, and 80:20 (train: test), generated randomly. The implementation was performed using MATLAB 2023a. The performance of DBN-CCFD was evaluated and compared with the performance of WOA-DBN-CCFD. At the highest training ratio of 80:20, WOA-DBN-CCFD shows False Positive Rate (FPR), sensitivity, specificity, precision, F1-Score, accuracy and detection time of 8.25%, 97.81%, 91.75%, 97.93%, 97.87%, 96.60% and 26.01s as against DBN-CCFD of 12.25%, 96.81%, 87.75%, 96.93%, 96.87%, 95.00% and 33.97s respectively. This performance metric indicates that the developed WOA-DBN-CCFD shows modestly better performance in credit card fraud detection, with lower FPR and detection time, while maintaining higher values on other metrics.