基于MARS-SVM的信用卡信用评估模型研究

发布时间：2018-04-15 12:47

本文选题：信用评估 + 多元自适应样条回归　；参考：《浙江工商大学》2012年硕士论文

【摘要】：信用卡对于个人、企业甚至国家的重要性不言而喻。目前,信用卡面临的主要难题就是信用风险,在普遍缺乏信用管理机制的中国该问题显得尤为突出。如何有效的控制风险,最大化的取得收益,成为信用卡发布机构面临的一个重大难题。因此,从理论层面到实践层面上,信用评估理论有了其广阔的发挥空间。信用卡信用评估的目的在于利用现有的客户属性包括社会属性和自然属性将信用卡申请者分为两类：对于能够较好的履行还款义务的申请者划分为“好客户”,同意颁发信用卡。对于可能出现拖欠或拒绝还款的申请者划分为“坏客户”,拒绝通过信用卡申请。早期的信用评估主要依赖于经验式的定性分析,缺乏效率且极易受到操作人员的主观影响。为此,众多的专家学者试着设计合适的信用评估模型以用于定量的处理信用风险问题。判别分析和logistic回归是最常用的(参数)统计分析方法。随着计算机技术的发展,以数据驱动作为核心思想的机器学习理论越来越受欢迎,决策树、神经网络等在信用评估问题中取得了较大的成功。设计一个合适的信用评估模型是本文的主要研究内容,为此,文章首先介绍了信用评估模型的存在意义。其次在文献综述部分详细的给出了建立信用评估模型的各个步骤以及目前的研究状况。然后通过对常见信用模型的梳理指出了其存在的优缺点。最后本文提出了MARS-SVM模型,充分利用MARS全局处理变量,并能对变量重要性进行排序的优点弥补了SVM不能进行特征选择的缺陷,从而得到了具有较高预测能力的混合模型：MARS是现代回归分析方法,对数据分布要求不高,通过逐步向前引入变量,逐步向后删除不重要变量的方式建立回归模型。所以MARS对变量的重要性排序具有全局最优性。SVM利用格点搜索法并采用交叉验证的方式确定惩罚参数和核函数参数。因此虽然其算法特性避免了“维度灾难”,但是过多的预测变量会影响其工作效率,而MARS的变量筛选正是其合理的补充。SVM的核心部分是核函数的选择,Rbf核函数具有普适性,易操作性的优点最受欢迎。但同时Rbf可能会导致特征空间样本信息损失所以在Rbf核函数基础上提出了KOBF核函数。本文将同时使用KOBF和Rbf作为SVM的核函数以对比分类效果。为了验证模型MARS-SVM的预测能力,本文做了对比实验。利用logistic回归、分类决策树、神经网络对同一样本数据集做了分类处理。结果显示,MARS-SVM模型具有较好的预测能力。
[Abstract]:The importance of credit cards to individuals, businesses and even the state is self-evident.At present, credit risk is the main problem faced by credit card, especially in China, where credit management mechanism is generally lacking.How to effectively control the risk and maximize the income has become a major problem faced by credit card issuers.Therefore, from the theoretical level to the practical level, the theory of credit evaluation has its broad scope.The purpose of credit card credit evaluation is to use existing customer attributes, including social attributes and natural attributes, to divide credit card applicants into two categories: those who are better able to meet their repayment obligations are classified as "good customers."Agree to issue a credit card.Applicants who may default or reject payments are classified as "bad customers" who reject credit card applications.The early credit evaluation mainly depends on the empirical qualitative analysis, which is inefficient and easily influenced by the operator.Therefore, many experts and scholars try to design appropriate credit evaluation model to deal with credit risk quantitatively.Discriminant analysis and logistic regression are the most commonly used statistical analysis methods.With the development of computer technology, the theory of machine learning with data-driven as its core is becoming more and more popular. Decision tree and neural network have achieved great success in credit evaluation.Designing a suitable credit evaluation model is the main research content of this paper. Therefore, this paper first introduces the significance of the credit evaluation model.Secondly, in the part of literature review, the steps of establishing credit evaluation model and the current research situation are given in detail.Then, the advantages and disadvantages of the common credit models are pointed out.Finally, this paper proposes a MARS-SVM model, which makes full use of MARS to deal with variables globally, and can sort the importance of variables to make up for the defect that SVM can not select features.It is concluded that the hybrid model: Mars with high predictive ability is a modern regression analysis method with low requirement for data distribution. The regression model is established by introducing variables forward step by step and deleting unimportant variables step by step.Therefore, MARS has global optimality for the importance of variables. SVM uses lattice search method to determine the penalty parameters and kernel function parameters.Therefore, although its algorithm features avoid "dimensionality disaster", too many prediction variables will affect its work efficiency, and the selection of kernel function is the core part of MARS, which is a reasonable supplement to .SVM.The advantages of ease of operation are most popular.But at the same time, Rbf may lead to the loss of sample information in the feature space, so the KOBF kernel function is proposed based on the Rbf kernel function.In this paper, both KOBF and Rbf are used as kernel functions of SVM to compare classification effects.In order to verify the prediction ability of the model MARS-SVM, a comparative experiment is made in this paper.Logistic regression, classification decision tree and neural network are used to classify the same sample data set.The results show that the MARS-SVM model has better prediction ability.
【学位授予单位】：浙江工商大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：F832.2;F224

【参考文献】