集成支持向量机方法及在信用风险中的应用研究

发布时间：2018-01-17 13:40

本文关键词：集成支持向量机方法及在信用风险中的应用研究　出处：《郑州大学》2013年硕士论文　论文类型：学位论文

【摘要】：近年来,支持向量机已成为机器学习方法研究的新领域,该方法具有适合小样本、非线性、高维度等特点。而集成学习的方法也越来越受关注,它可以综合利用各个个体分类器的优点,提高整个学习器的泛化能力。目前信用风险已成为风险中一个重要的方面。为了预防信用风险的发生,人们开始对信用风险进行管理和预测,信用风险评价的方法也越来越多。本文主要采用两种集成方法建立模型,对信用风险进行评级。论文的主要研究内容包括： 1介绍了与本文研究内容相关理论和文献。首先系统阐述了信用风险评价的现状、支持向量机分类研究的现状,以及相关的理论基础。在总结前人研究的优缺点的基础上,提出了本文采用的集成支持向量机方法。 2信用风险评价指标体系的建立及建模思路。详细总结以往信用风险评价指标的选取,根据本文样本的特点,构建了用于信用风险评价的指标体系,初次共选取了反映公司七个财务方面的24个指标。同时概括介绍了本文的建模思路。 3样本选取、模型的建立以及最后实证获得的结论。选取了73对配对样本,并从原有24个指标中选取了19个指标作为研究数据。建立了基于不同集成方法的信用风险评价模型,并分别与标准的支持向量机做对比,研究结果表明,Bagging_SVM和Boosting_SVM两种集成方法的判别准确率都高于标准的支持向量机模型,并且两类模型大大提高了对风险公司的判别准确率,这在实际中具有重要的指导意义。然后,还研究了不同分类器个数对两种集成模型判别正确率的影响,结果表明,弱分类器的个数的多少对模型的判断率有一定的影响,这为以后选择性集成提供了一定的依据。最后,通过ROC图比较了三种模型的性能,研究结果表明Bagging_SVM的性能最好,Boosting_SVM居中,标准SVM最差。本文创新点有两点：第一是在配对样本的选取时提出了具体的参考标准。第二是在对集成模型的研究中,不仅考虑了不同集成方法对模型判别准确率的影响,而且研究了不同数目的弱分类器集成对模型判别准确率的影响。
[Abstract]:In recent years, support vector machine (SVM) has become a new research field of machine learning, which is suitable for small samples, nonlinear, high-dimensional and so on. It can make comprehensive use of the advantages of individual classifiers and improve the generalization ability of the whole learner. At present, credit risk has become an important aspect of risk, in order to prevent the occurrence of credit risk. People begin to manage and predict the credit risk, and there are more and more methods to evaluate the credit risk. The main contents of this thesis are as follows: Firstly, the current situation of credit risk assessment and support vector machine classification are systematically described. On the basis of summarizing the merits and demerits of previous studies, the integrated support vector machine (ISVM) method is proposed in this paper. 2 the establishment and modeling of credit risk evaluation index system. The selection of credit risk evaluation index is summarized in detail. According to the characteristics of the sample in this paper, the index system for credit risk evaluation is constructed. A total of 24 indexes reflecting seven financial aspects of the company are selected for the first time, and the modeling ideas of this paper are summarized. 3 sample selection, the establishment of the model and the final empirical conclusions. 73 pairs of paired samples were selected. From the original 24 indicators, 19 indicators are selected as the research data. The credit risk evaluation model based on different integration methods is established, and compared with the standard support vector machine, the research results show that. The accuracy of Bagging_SVM and Boosting_SVM integration methods is higher than the standard support vector machine model. And the two kinds of models greatly improve the discrimination accuracy of venture companies, which has important guiding significance in practice. Then, the effect of the number of different classifiers on the accuracy of the two integrated models is also studied. The results show that the number of weak classifiers has a certain influence on the judging rate of the model, which provides a certain basis for the future selective integration. Finally, the performance of the three models is compared by ROC diagram. The results show that the performance of Bagging_SVM is the best and the standard SVM is the worst. There are two innovations in this paper: the first is to put forward the specific reference standard in the selection of paired samples, and the second is to consider not only the influence of different integration methods on the accuracy of model discrimination in the research of integration model. Moreover, the influence of different number of weak classifiers on the accuracy of model discrimination is studied.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：F832.5;F224;TP18

【参考文献】