False Positives in Credit Card Fraud Detection: Measurement and Mitigation

Abstract

Credit Card Fraud Detection is a classification problem where different types of classification errors cause different costs. Previous works quantified the financial impact of data-driven fraud detection classifiers using a cost-matrix based evaluation approach, however, none of them considered the significant financial impact of false declines. Analysts reported that fraud prediction in e-commerce still has to deal with false positive rates of 30-70%, and many cardholders reduce card usage after being wrongly declined. In our paper, we propose a new method for assessing the cost of false declines and evaluate several state-of-the-art fraud detection classifiers using this method. Further, we investigate the effectiveness of ensemble learning as previous work supposed that a combination of diverse, individual classifiers can improve performance. Our results show that cost-based evaluation yields valuable insights for practitioners and that our ensemble learning strategy indeed cuts fraud cost by almost 30%.