Mitigating unbalanced and overlapped classes in credit card fraud data with enhanced stacking classifiers system

Abstract

Background: Credit cards remain the preferred payment method by many people nowadays. If not handled carefully, people may face severe consequences such as credit card frauds. Credit card frauds involve the illegal use of credit cards without the owner’s knowledge. Credit card fraud was estimated to exceed a $35.5 billion loss globally in 2020, and results in direct or indirect financial loss to the owners. Hence, a detection system capable of analysing and identifying fraudulent behaviour in credit card activities is highly desirable. Credit card data are not easy to handle due to their inherited problems: (i) unbalanced class distributions and (ii) overlapping classes. General learning algorithms may not be able to address and handle the problems well. Methods: This study addresses these problems using an Enhanced Stacking Classifiers System (ESCS) that comprises two sequential levels. The first level is an excellent classifier for detecting normal credit card transactions (the majority class), while the second level contains stacking classifiers that distinguish credit card frauds (the minority class). The ESCS can improve the fraud detection via the second level, which contains sensitive classifiers to identify the misclassified fraud transactions as normal transactions from the first classifier. The meta-classifier then combines the decisions of the base classifiers from the levels to produce the final detections. Results: We evaluated the ESCS using the benchmark credit card fraud dataset (CCFD) that exhibits the two problems. The highest true positive rate (TPR) for detecting credit card frauds was 0.8841, which outperformed the single classifiers, bagging, boosting, and other researchers’ works. Conclusions: This study proves that the ESCS, with an additional level added to the stacking classifiers, can improve fraud detection on credit card data.