Sparse multi-criteria optimization classifier for credit risk evaluation

Over the past few decades, many classifier methods are suggested for credit risk evaluation. With ever-increasing amounts of data, for multi-criteria optimization classifier (MCOC) and other traditional classification methods, owing to the correlation among different features in data these classifiers often give the poor predictive performance. Thus, some dimensionality reduction techniques are firstly used to find important features; then, these classifier models are built on the reduced data set. However, because feature selection and classification are carried out in different feature spaces, the purpose of increasing predictive accuracy and interpretability is difficult to achieve truly. It is therefore important to research the new classifier methods with simultaneous classification and feature selection so as to improve the predictive accuracy and obtain the interpretable results. In this paper, we propose a novel sparse multi-criteria optimization classifier (SMCOC) based on one-norm regularization, linear and nonlinear programming, respectively, and construct the corresponding algorithm. The experimental results of credit risk evaluation and the comparison with linear and quadratic MCOCs, logistic regression and support vector machines have shown that the proposed SMCOC can enhance the separation of different credit applicants, the efficiency of credit scoring, the interpretability of risk evaluation model and the generalization power of risk prediction for new credit applicants.