基于相似样本归并的大样本混合信用评估模型

A hybrid large sample credit evaluation model based on combining similar samples

当前面向大样本设计的信用评估模型,大多没有深入探究大样本的分布特征,只是简单地将传统评估方法应用在大样本上.首先提出了用于描述大样本分布特征的相关属性集、边界向量等若干概念及定义,并证明了其主要性质.之后在两个大样本数据集的基础上,研究了样本在相似性方面的分布特征,最后设计了一种大样本混合信用评估模型——HLSCE模型.HLSCE模型认为在大样本数据集中,样本的同一属性在不同局部区域内,对分类性能的贡献是不同的.具体地,HLSCE模型根据各样本与边界向量的相似性差异,结合生物启发式算法,将样本归并划分为若干子集并分别在其上训练基分类器.实证研究表明,HLSCE模型的分类精度相比于现有的代表性信用评估模型更高,同时也具有更为优越的平衡性与稳定性.

Most of the current credit evaluation models designed for large samples have no in-depth consideration of the distribution characteristics of large samples,and just simply apply traditional evaluation methods to large samples.This paper firstly proposes the concept and definition of the related attribute set,boundary vector and so on to describe the distribution characteristics of large samples and proves their main attributes.Then the characteristics of sample distribution are studied in the aspect of similarity based on two large sample data sets.Finally,a hybrid large sample credit evaluation model：HLSCE model is designed.The key idea of HLSCE model is that in large sample data sets,the contribution of the same attribute of samples in different local areas are different to classification performance.Specifically,HLSCE model divides,with biological heuristic algorithm,the whole data set into several subsets according to the similarity between samples and boundary vectors,and then trains the basic classifiers respectively on each subsets.The empirical study shows that compared with the existing representative credit evaluation models,our HLSCE model has a higher classification accuracy,as well as a better balance and stability.