Default discrimination of credit card: Feature combination selection based on improved FDAF-score

Abstract

Default discrimination of credit card refers to the phenomenon in which banks discriminate against certain groups of customers based on their credit status. The challenges faced by banks and other financial institutions in evaluating borrowers and making lending decisions are inseparable from the discrimination of customers' credit status. Feature selection is a critical step in default discrimination, and the selection of features directly affects the results of default discrimination. Therefore, the identification of a group of optimal features to maximize the ability of default identification is the core problem tackled in this research. The innovation of this paper is that it extends the single feature selection model based on F-score and Fisher discriminant analysis (FDAF-score) to the feature combination selection. Taking the maximum FDAF-score of the combination of the whole features as the standard, an optimal feature combination with the maximum discrimination ability is deduced. It avoids the disadvantage that single feature selection cannot ensure that the combination of selected features still has the maximum default discrimination ability. In this paper, we find that the improved feature combination selection model based on FDAF-score was better than the single feature selection FDAF-score model. The results were confirmed in four credit data sets (binary classification) of Australia, Germany, Taiwan and Japan, and six comparative data sets (five of which are multi-classification data sets). When the feature combination selection method is combined with seven common classification models, except for image and German credit data sets, the classification accuracy of most models is higher than the FDAF-score single feature selection method.