基于差异性度量的多分类器融合个人信用评估研究

发布时间：2018-01-09 20:15

本文关键词：基于差异性度量的多分类器融合个人信用评估研究　出处：《哈尔滨工业大学》2014年硕士论文　论文类型：学位论文

【摘要】：随着我国商业银行风险管理的要求进一步加强，个人消费信贷业务中的个人信用评估体系不断优化提升以保证商业银行风险管理更高的要求。在理论上各种单一模型不断被优化，无论是从数据集的收集、最优特征提取还是从单一模型的顶层设计不断的完善，都已经到达了一定的高度，尤其是在大数据的呼声越来越高的背景下，数据挖掘技术日益成熟，个人信用评估问题也逐步成为其中的一类十分具体的要求解决的实际问题，并且从大量的技术中选择恰当的理论来切合个人信用评估问题的实际核心点成为日益迫切需要解决的问题，虽然大量的个人信用评估模型被构建，但其中能够实现绝对优势的模型仍未被发现，尤其是理论上的模型如何能够应用于实际仍是一大问题。随着单一模型的不断优化准确率不断提升及多分类器融合系统的发展，本研究在此基础上进行了更进一步的研究以期能够实现评估模型准确率和稳健性的进一步提升。本研究对基于差异性度量多分类器融合的个人信用评估问题进行了系统的研究，本研究是在个人信用评估单一的模型数量规模不断增大，且单一模型不断优化的基础上进行，同时，也伴随着多分类器融合问题逐步趋于成熟的背景而开展，基于这两点趋势发展，本文提出了基于差异性度量进行多分类器融合以提高个人信用评估模型的准确率和稳健性，规避商业银行风险管理当中个人信贷违约给银行带来的损失。本研究首先对差异性度量证据理论融合的个人信用评估模型进行设计，从证据理论融合模型构建，差异性度量方法讨论及具体思路分析几个方面进行研究；其次，，本文根据前文思路进行了基于差异性度量的单一分类器的证据理论融合，构建了5种具有代表性的单一个人信用模型并利用4种差异性度量方法以对模型之间的互补性进行考察，选取其中差异性最大的组合进行证据理论融合，由结果看出，未全部发挥出双方模型的优势，之后本研究利用准确率和稳健性对差异性度量的证据理论融合模型做进一步优化，结果表明这种综合的优化实现了模型准确率和稳健性的提升；再次，为实现单一模型的优化并排除样本结构对于准确率和差异性之间相关关系问题的考察，本文构建了基于集成的个人信用评估模型，并对集成模型进行了差异性度量和基于证据理论的融合，最后对集成模型的准确率和差异性之间的相关关系进行了实证分析，并期望实证结果能够为后续模型优化提供理论基础和相应的指导。
[Abstract]:With the requirement of risk management of commercial banks in our country further strengthened. Personal credit evaluation system in personal consumer credit business is constantly optimized to ensure higher requirements of commercial bank risk management. In theory, a variety of single models are constantly optimized, whether from the collection of data sets. Optimal feature extraction or continuous improvement from the top design of a single model has reached a certain height, especially in the context of big data's growing voice, data mining technology has become increasingly mature. Personal credit evaluation has gradually become one of the very specific practical problems to be solved. And choosing the appropriate theory from a large number of technologies to meet the actual core of personal credit evaluation has become an increasingly urgent problem, although a large number of personal credit evaluation models have been constructed. However, the model that can achieve absolute superiority has not been found. In particular, how the theoretical model can be applied in practice is still a big problem. With the continuous improvement of the accuracy of single model optimization and the development of multi-classifier fusion system. On this basis, further research is carried out to improve the accuracy and robustness of the evaluation model. This study systematically studies the problem of personal credit evaluation based on diversity measurement multi-classifier fusion. In this study, the number of individual credit assessment models is increasing. And the single model is continuously optimized on the basis of, at the same time, along with the multi-classifier fusion problem gradually mature background, based on these two trends to develop. In order to improve the accuracy and robustness of personal credit evaluation model, this paper proposes a multi-classifier fusion based on diversity measurement. In order to avoid the loss caused by individual credit default in the risk management of commercial banks, this study first designs the personal credit evaluation model which integrates the evidence theory of difference measurement, and constructs the model of evidence theory fusion. The difference measure method discussion and the concrete train of thought analysis several aspects carries on the research; Secondly, this paper carries on the evidence theory fusion of the single classifier based on the difference measure according to the previous ideas. In this paper, five representative single person credit models are constructed, and four different measures are used to investigate the complementarities between the models, and the most different combinations are selected for evidence theory fusion. From the results, not all of the advantages of the two models, after this study using accuracy and robustness to measure the evidence theory of diversity fusion model to further optimize. The results show that the model accuracy and robustness can be improved by the synthesis optimization. Thirdly, in order to optimize the single model and eliminate the sample structure for the accuracy and diversity of the correlation between the problem, this paper builds an integrated personal credit evaluation model. Finally, the correlation between the accuracy and the difference of the integration model is analyzed empirically. The empirical results are expected to provide theoretical basis and corresponding guidance for subsequent model optimization.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：F224;F830.5

【参考文献】