DEVELOPMENT OF DEEP BELIEF NETWORK MODEL FOR CREDIT CARD FRAUD DETECTION

Abstract

Credit card fraud detection (CCFD) has attracted various critical research interest owning to increase in fraudulent activities in financial transactions. Machine learning and rule-based system, often struggle to solve the inherent challenges of imbalanced transactional datasets, concept Drift, real-time detection and Feature Engineering. Hence, this study developed credit card fraud detection system by employing a light weight deep belief network (DBN) model. The DBN which can learn hidden pattern of imbalanced dataset was formulated with 2 hidden layer, 244 nuerons and 0.00921 learning rate. The architecture was formed using stacked layers of restricted boltzman machine while weight and biases were updated using contrastive divergence technique. The model performance on CCFD was accompanied on 10,000 transactional records with imbalanced ratio of 98:2 legitimate to fraudulent datasets. The result shows an average of 96.48% sensitivity, 96.36% precision, 95.66% F1-score and 94.96 accuracy for DBN-CCFD as compared to 95% sensitivity, 88% precision, 91% F1-score and 94% accuracy of DBNex. This study realized that a light weight architectural through careful selection of BDNs hyperparameters can result in excellent performance, providing a new direction for fraud detection system design that prioritizes efficiency without sacrificing accuracy.