基于样本依赖代价矩阵的小微企业信用评估方法

Credit Scoring of Small and Micro Enterprises Based on Sample-Dependent Cost Matrix

摘要:

针对小微企业信用历史数据规模较小，而且类别不平衡问题较为严重，提出基于样本依赖代价矩阵的Smote XGboost?Bayes Minimum Risk (SXG?BMR)模型，对整体样本进行低倍率过采样，以弱化类别不平衡问题，降低模型过拟合的风险；模型将集成学习模型与最小风险贝叶斯决策相结合，以实现代价敏感。同时，模型中引入了样本依赖的代价矩阵，该代价矩阵不仅与类别有关，而且与样本自身属性有关，可以更为准确地表征代价。使用标准信用数据集和上海市小微企业信用数据集，进行多种算法的对比分析，结果表明，该模型性能优良。

关键词:信用评估;样本依赖;最小风险贝叶斯;XGBoost模型;代价敏感学习

Abstract:

Because the credit history data of small and micro enterprises are small and the problem of class imbalance is more serious, this paper proposes a Smote XGboost-Bayes Minimum Risk (SXG-BMR) model based on the sample-dependent cost matrix. The whole sample is oversampled at a low rate to weaken the problem of class imbalance and reduce the risk of model overfitting. The model combines the integrated learning model with the minimum risk Bayes decision to realize the cost sensitivity. At the same time, this paper introduces the sample-dependent cost matrix into the model. The cost matrix is related not only to the category, but also to the attributes of the sample.Therefore ,it can characterize the cost more accurately. In the empirical study,this paper uses a standard credit dataset and a real credit dataset of small and micro enterprises in Shanghai. Besides,it compares and analzes of various algorithms. The results show that the SXG-BMR model proposed in this paper has a good performance.

Key words:credit scoring;sample-dependent;minimum Bayes risk;XGBoost model;cost sensitive learning