Cost-sensitive payment card fraud detection based on dynamic random forest and k-nearest neighbors

Payment card fraud leads to heavy annual financial losses around the world, thus giving rise to the need for improvements to the fraud detection systems used by banks and financial institutions. In the academe, as well, payment card fraud detection has become an important research topic in recent years. With these considerations in mind, we developed a method that involves two stages of detecting fraudulent payment card transactions. The extraction of suitable transactional features is one of the key issues in constructing an effective fraud detection model. In this method, additional transaction features are derived from primary transactional data. A better understanding of cardholders' spending behaviors is created by these features. After which the first stage of detection is initiated. A cardholder's spending behaviors vary over time so that new behavior of a cardholder is closer to his/her recent behaviors. Accordingly, a new similarity measure is established on the basis of transaction time in this stage. This measure assigns greater weight to recent transactions. In the second stage, the dynamic random forest algorithm is employed for the first time in initial detection, and the minimum risk model is applied in cost-sensitive detection. We tested the proposed method on a real transactional dataset obtained from a private bank. The results showed that the recent behavior of cardholders exerts a considerable effect on decision-making regarding the evaluation of transactions as fraudulent or legitimate. The findings also indicated that using both primary and derived transactional features increases the F-measure. Finally, an average 23% increase in prevention of damage (PoD) is achieved with the proposed cost-sensitive approach. (C) 2018 Elsevier Ltd. All rights reserved.