基于多属性子集选择策略的三阶段混合信用评估模型  

A three-stage hybrid credit evaluation model based on multi-attribute subset selection strategies

作　　者：张润驰 杜亚斌[1] 薛立国[1] 徐源浩 孙明明[2] ZHANG Run-chi;DU Ya-bin;XUE Li-guo;XU Yuan-hao;SUN Ming-ming(Department of Finance, School of Business, Nanjing University, Nanjing 210093, China;Lianyungang central sub-branch of the People's Bank of China, Lianyungang 222000, China)

机构地区：[1]南京大学商学院,江苏南京210093 [2]中国人民银行连云港中心支行,江苏连云港222000

出　　处：《管理工程学报》2019年第2期140-147,共8页Journal of Industrial Engineering and Engineering Management

基　　金：国家自然科学基金资助重大研究计划项目(90718008);国家自然科学基金资助重点项目(61133015);江苏省自然科学基金资助项目(2004119)

摘　　要：信用评估数据集往往存在冗余属性,现有研究一般通过属性子集选择策略进行属性筛选,但并未深入研究不同属性子集选择策略在不同信用评估模型上的适用性。本文首先实证研究了8种属性子集选择策略对7种主流模型的性能提升情况,得到了一些有意义的结论;进而设计出一种结合多个属性子集选择策略特征的三阶段混合信用评估模型——TSHCE模型。TSHCE模型在第一阶段,根据多个属性子集选择策略对各属性的重要性排序,生成属性重要性序数向量;第二阶段,根据属性重要性序数向量,以轮盘赌方法选择不同属性子集,分别训练基分类器;第三阶段,以各基分类器的分类结果组合构成再训练样本集,进一步训练连接分类器以提升模型的分类能力。实证研究表明:TSHCE模型在训练阶段能够深度挖掘样本集的可分类特征,五种性能评价指标均达到92%以上;在测试阶段,相对于最优属性子集选择策略与分类器的组合,在两组大样本数据集上分别进一步提升了1.36%和12.83%的总体分类正确率,具有优越的平衡性,同时亦适用于小样本。The credit evaluation data sets from current credit institutions always have a certain number of redundant attributes. In the current research, the attribute subset selection strategy is used to screen the attributes, but there is no in-depth study about the applicability of different attribute subset selection strategies on different data sets with different credit evaluation models. This paper firstly studied the performance of eight attribute subset selection strategies combined with seven main credit evaluation models based on two different large sample data sets. Empirical results show that attribute subset selection strategies can further enhance the classifier's performance compared to the whole attribute selection method. Also, the number of the attributes selected by the optimal strategy is not the same in different classifiers. Even in the same data set, the dimension of modeling sample data also has a different influence on different classifiers. Classifiers such as Naive Bayes has a better classification performance improvement compared to other classifiers combined with different attribute subset selection strategies, while k-NN performance is the best of the two data sets. Moreover, excluding the factor of classifiers and datasets, there exists a difference between different attribute subset selection strategies. Then we designed a three-stage hybrid credit evaluation (TSHCE) model, which combined features of multiple attribute subset selection strategies. Logically, our TSHCE model can be divided into three stages. In the first stage, this study ranks the importance of each attribute according to multiple attribute subset selection strategies, then generates the attribute importance ordinal vector, which measures the average importance ranking of each attribute in multiple attribute subset selection strategies. In the second stage, according to the attribute importance ordinal vector, this study uses the roulette method to extract subsets with different attributes and then trains several basic clas

关 键 词：属性子集选择策略 三阶段混合信用评估模型 属性重要性序数向量 信用风险 

分 类 号：F83[经济管理—金融学]