Reducing Fraud Detection Costs in Credit Card Transactions: An Information Fusion Approach

Objective Most companies and organizations use e-commerce to gain productivity and efficiency in their services and products in different areas such as credit cards, telecommunications, health insurance, car insurance, etc. Due to the growing volume of credit card transactions and the various methods of fraud and cheating on these cards, the demand for detecting fraud in this area is also increasing. Considering the various solutions and algorithms presented to reduce the cost of fraud detection in credit card transactions in the literature, the purpose of this research is to present a combined and optimal method to reduce the cost of fraud detection in credit card transactions for financial systems, using the fusion of heterogeneous classification and clustering algorithms at the decision-making level by using two fusion methods: Probabilistic fusion and Dempster-Shafer Evidence theory. Methods This study utilizes a transaction dataset from a Brazilian bank, covering two months from July 14 to September 12, 2004. Fraud detection performance is evaluated using a cost function derived from both a supervised learning approach, namely, a neural network, and an unsupervised method, the K-Means clustering algorithm. Drawing on established fraud detection metrics in the literature, the study adopts the cost function introduced by Gadi as the benchmark. Recognizing the high cost associated with using a single algorithm, the study implements two information fusion techniques—Dempster-Shafer evidence theory and probabilistic fusion—to reduce detection costs. Both fusion methods operate at the decision level, integrating heterogeneous outputs from the supervised and unsupervised models. Results Depending on the algorithms implemented, using only one algorithm to obtain an acceptable cost function can be very costly. While using a fusion approach can have a significant impact on cost reduction. The findings indicate that probabilistic fusion significantly outperforms the Dempster-Shafer evidence theory in minimizing the cost function. Specifically, probabilistic fusion achieves a 21.4% cost reduction compared to the artificial neural network and a 35.8% reduction relative to the K-Means algorithm. The results of this study were finally compared with a paper in which this dataset was first used with the artificial immune system (AIS) algorithm and showed a significant cost reduction. Conclusion In this study, two classification and clustering algorithms were utilized, and their fusion at the decision level was implemented to demonstrate that combined methods reduce the cost function more effectively than the use of individual algorithms. Furthermore, it was shown that probabilistic fusion yields a lower cost in detecting fraud within financial systems compared to the Dempster-Shafer Evidence theory. This finding is considered significant for banks and financial institutions aiming to develop effective fraud detection systems.