Feature construction for fraudulent credit card cash-out detection

摘要

As a type of credit card fraud behavior, fraudulent cash-out causes banks and companies to incur huge losses. Furthermore, such organizations lack effective methods to detect fraudulent cash-out. To detect fraudulent cash-out accurately, we construct four feature sets based on domain knowledge from industry experts, tips from fraudsters, related media reports, and the relevant literature. In addition, we construct features using a functional data analysis algorithm to capture the time-dependent behavioral patterns of cardholders. We also construct a benchmark feature set based on the traditional approach of Whitrow's strategy. We compare these feature sets using a real data set comprising real transactions of 25,000 credit cards with three machine learning methods: eXtreme gradient boosting, random forest, and support vector machine. The results reveal that our proposed features, which consider both snapshot and dynamic behavioral patterns of cardholders, achieve considerably superior performance to that of Whitrow's strategy. The precisions at the top 5%, 10%, 15%, and 20% levels are improved by magnitudes of 0.049, 0.081, 0.053, and 0.046, respectively.