支持向量机在个人信用评估中的应用

【作者】 孙瑾；

【导师】 许青松；

【作者基本信息】 中南大学 ， 统计学， 2008， 硕士

【摘要】 个人信用评估是商业银行风险管理的重要组成部分,国外银行界对于信用评估的研究已经有50多年的历史,发展出了包括统计方法和非统计方法两大类许多种方法。支持向量机(SVM)是近年来在统计学习理论的基础上发展起来的一种新的机器学习方法,它具有很强的泛化能力。本文的研究中心就是支持向量机在个人信用评估中的应用,引入遗传算法作为筛选属性变量和调节参数的优化算法,建立基于遗传算法和支持向量机的个人信用评估模型。最后将支持向量机作为AdaBoost算法的基础学习器,建立AdaBoost-SVM模型,应用到个人信用评估,实证分析表明,此种模型较之单一的支持向量机更有效。本文所做的主要工作为:1、考虑到模型的输入变量和模型的参数之间存在着相互依赖性,本文引入遗传算法将属性变量选择和参数调节两方面的工作同时进行,同步优化,使支持向量分类器性能达到最优。2、提出了动态AdaBoost支持向量机模型。传统的AdaBoost算法在整个Boosting过程中使用同一个学习器,这样做的话会造成有的支持向量机过强,有的过弱,而最终Boosting效果欠佳,因此,我们在每一次Boosting过程中都通过调节参数使支持向量机精度仅略高于随机猜测,得到一个动态的AdaBoost支持向量机模型。实证分析表明,该模型优于普通的支持向量机。更多还原

【Abstract】 Personal credit scoring is an important part of commercial banks’ risk management. In the last 50 years, many credit scoring methods have been developed by foreign banks. Support Vector Machine(SVM) is a new machine learning method developed in recent years on the foundation of statistical learning theory. The focus of this thesis is to apply SVM on Crediting Scoring. In this paper, Genetic Algorithm was used to choose the optimal input feature subset and set the best kernel parameters simultaneously, establishing a credit scoring model named GA-SVM. In addition, SVM was applied as the basic learning machine of AdaBoost algorithm, establishing another credit scoring model named AdaBoost-SVM. Experimental results have shown that AdaBoost-SVM is better than GA-SVM, which is better than the usual SVM.The main job of this paper are following:1、The traditional methods of credit scoring prefer to do feature selection and parameters optimization independently. The correlation between them is not considered, prohibiting the global optimal results. This paper tries to combine feature selection with parameter optimation based on genetic algorithm during SVM modeling.2、Dynamic Boosting has been coupled with SVM to established a AdaBoost-SVM. Traditional AdaBoost prefers to use the identical learning machine during the boosting process. In this paper, we design a parameter adjusting strategy to get different and moderately accurate SVM component classifier for boosting. And good results have been obtained on benchmark data sets.更多还原