关键词: 信用风险评价, 机器学习模型, 可解释性, 教学式方法

Abstract: Credit risk assessment is one of the important tasks in risk prevention and control of financial institutions. In recent year, the credit risk assessment based on machine learning models have received increasing attention due to their better predictive performance. However, the machine learning models lack the interpretability, which makes it impossible for decision makers to fully trust the models and their predictive results. To solve the above problem, an improved pedagogical method is proposed. The proposed method uses the machine learning models guide to construct the credit risk assessment based on a decision tree, which can assist investors in decision-making. In our approach, to improve the approximation degree between decision tree and the correct function of machine learning model, a pseudo date set sampling method based on Weight Synthetic Minority Over-sampling Technique (Weight-SMOTE) is proposed to improve the proportion of pseudo samples, which is marked by the trusted function of machine learning model, in the pseudo data set. To achieve an effective trade-off between accuracy, interpretability and the consistency of generated decision tree with machine learning model, a new decision tree pruning method is proposed. At the same time, aiming at the limitation of fidelity evaluation indicator, a new evaluation indicator, true-fidelity, is proposed to effectively measure the approximation degree of the decision tree and the correct function of the machine learning model. In addition, three real credit risk assessment data sets are used in the experiment. The experimental results show that the improved pedagogical method can construct an interpretable credit risk assessment, which can meet the different decision-making preferences of decision makers.

Key words: credit risk assessment, machine learning model, interpretability, pedagogical method