Random Subspace Method for Improving Performance of Credit Cardholder Classification

The main task of credit card risk management for commercial banks and credit card retailers is analyzing credit cardholders’ behavior and predicting bankruptcy. In this paper, we investigate dimension reduction techniques for improving performance of the high-dimensional credit cardholder data classification. We choose MCLP as base classifier. We theoretically analyze the characteristics of PCA, filter, wrapper, and RSM. Then, our experimental research focuses on RSM. Due to the specialities of dimension reduction and ensemble of RSM, we experimentally compare its performance with that of PCA, single classifier and Bagging on the same training set and test set. The results show that RSM can highly improve classification performance of the credit cardholder data set. From the idea of ensemble, RSM demonstrates its superiority over Bagging. From the angle of dimension reduction, RSM shows its predominance over single classifier, and the same advantage as PCA. Finally, we explain our results from the aspects of MCLP algorithm, RSM and data set.