网络借贷业务个人信用评价方法研究

发布时间：2018-01-16 05:21

本文关键词：网络借贷业务个人信用评价方法研究　出处：《合肥工业大学》2016年博士论文　论文类型：学位论文

【摘要】：网络借贷信用评价能够有效地缓解借贷双方间的信息不对称性,降低违约风险和交易成本。然而,网络借贷业务中借款人的财务信息难以获取和验证,给传统的基于财务信息的信用评价方法带来巨大的困难。事实上,网络环境下,借款人的信用相关数据不仅包括财务信息,也包括非财务信息。这些非财务信息广泛分布在不同的网络平台中,具有体量大、价值密度低和质量参差不齐等特点,给网络借贷信用评价研究带来新的困难。为此,本文在综述信用评价相关理论与方法的基础上,结合网络借贷业务的特点,从信用评价的数据预处理、信用特征选择和信用评价模型构建三个方面,对网络借贷业务的信用评价问题展开研究。研究的主要内容和相关结论如下。(1)网络借贷个人信用评价数据预处理方法网络借贷业务信用评价数据的质量参差不齐,缺失值和异常值现象严重。在缺失值处理方面,针对多重填补法难以对包含类别变量的数据集进行缺失值填补的问题,提出一种分类多重填补法。该方法利用类别变量与连续变量数学特征间的关系,根据类别变量信息对连续变量的相关数据特征进行估计,提高缺失值填补效果。在异常值处理方面,针对单一信用特征的异常值处理问题,提出一种基于KNN的异常值纠偏方法,该方法能够利用近邻样本的相关信用特征对异常值进行纠正。针对密度分布不均匀空间中的异常样本检测的距离阈值难以确定问题,提出一种基于DBSCAN和相对密度的异常样本检测方法。该方法首先利用DBSCAN算法将密度分布不均匀空间分成若干个密度分布均匀的类,然后在每个类中运用相对密度方法确定异常样本。最后,在拍拍贷平台上进行信用数据预处理实验,结果表明经过预处理的数据能够显著增强信用评价模型的性能。(2)网络借贷个人信用特征选择方法网络借贷业务信用信息的体量大,价值密度低。需要结合相关理论与方法,对信用特征进行定性初选和定量筛选。在信用特征定性初选阶段,结合信用所具有的资本性和社会资本理论,从结构维度、关系维度和认知维度三个方面,分析网络借贷业务中借款人的社会资本,并结合借款人的个人信息、借款历史信息和身份验证信息等,研究融合社会资本的信用特征定性初选方法。在信用特征的定量筛选阶段,考虑到信用特征的变量类型多样且与信用状态变量间的关系复杂,提出一种基于综合定量分析的信用特征筛选方法,该方法运用相关分析、卡方统计量分析、信息增益分析和支持向量回归分析等定量分析方法,筛选与信用状态变量具有线性和非线性关系的定类与定距信用特征。最后,在拍拍贷平台上的实验结果表明,提出的信用特征初选和筛选方法能够全面获取多变量类型和多关系类型的信用特征。(3)网络借贷个人信用评价模型现有的信用评价模型在网络借贷环境下的应用效果不佳,需要对现有模型进行改进并结合网络借贷环境下信用表现出的相关特性,设计新的信用评价模式和模型。现有的Adaboost集成学习模型仅根据误分类率调整基分类器的样本权重,忽略了分歧度和误分类成本等因素对于样本权重的影响,造成集成后的模型精确度下降。为此,提出一种基于分歧度与误分代价的Adaboost信用评价模型,该模型能够对分类困难样本和误分代价高的样本进行有针对性的学习,提高信用评价结果的准确性。在拍拍贷平台上的实验结果表明,基于分歧度和误分代价的Adaboost信用评价模型的性能显著优于传统的Adaboost模型。此外,网络借贷环境下,根据信用表现出的全息性,设计一种Peer-to-Peer协同信用分析机制,获取并集成评价对象在多个网络平台上的相关信用特征,建立基于协同分析模式的跨业务的信用评价模型,从而对借款人的信用做出更加全面的评价。在相关的网络借贷平台、电子商务平台和社会网络平台上的实验结果表明,基于协同分析的跨业务信用评价模型能够有效地提升信用评价结果的有效性。
[Abstract]:The evaluation of the network credit lending can effectively alleviate the information asymmetry between the two sides to reduce the default risk and transaction cost. However, the network lending business in the financial information is difficult to obtain and verify, to the traditional credit evaluation method based on financial information has brought great difficulties. In fact, under the network environment, the borrower credit related data not only include financial information, including non financial information. These non financial information are widely distributed in different network platform, has the characteristics of large amount of low value density and uneven quality, bring new difficulties to the evaluation of the network lending credit. Therefore, based on the review of credit evaluation theory and method. Combined with the characteristics of the network lending business, from data preprocessing of credit evaluation, the construction of three aspects of credit feature selection and credit evaluation model of network Research on the credit evaluation problem. Lending business the main research contents and conclusions are as follows. (1) the network lending personal credit evaluation data preprocessing method of network lending business credit evaluation data quality uneven, missing values and outliers seriously. Processing value in the absence, for the multiple imputation method is difficult to contain type variables the data sets fill problem of missing values, this paper proposed a classification of multiple imputation. The method of categorical variables and continuous variables, the mathematical relation between features, estimated according to the data characteristics of continuous variables and class variables, improve the effect of missing values missing. In treatment of abnormal, the abnormal value for a single credit characteristic to deal with the problem, this paper proposes a novel value correction method based on KNN, this method can correct the abnormal value of the credit characteristics of neighbor samples Is the distance threshold. According to the abnormal samples of uneven density distribution in space detection is difficult to determine the problem, this paper proposes a novel sample DBSCAN and relative density detection method based on using the method of DBSCAN algorithm with non-uniform density space into a plurality of uniform density distribution, and then use the method to determine the relative density of abnormal samples in each class. Finally, experiments of credit data preprocessing in a pat on the loan platform. The results show that the preprocessed data can significantly enhance the performance of credit evaluation model. (2) lending network personal credit feature selection method of network lending business credit information of the large, low value density. According to the related theory and method the credit characteristics, qualitative and quantitative screening. In the primary credit characteristics of qualitative theory of primary stage, with the combination of credit capital and social capital, From the three aspects of structure dimension, relational dimension and cognitive dimension, network analysis of the borrower lending business in social capital, and combined with the borrower's personal information, borrowing history information and authentication information, the credit characteristics of primary qualitative study on fusion method of social capital. In the quantitative characteristics of credit screening stage, taking into account the characteristics of credit variables various types of credit and the relationship with the state variables is complex, a screening method of credit characteristics of comprehensive quantitative analysis based on the method of using correlation analysis, chi square statistic analysis method to analyze the information gain analysis and support vector regression analysis, quantitative screening and credit of state variables is linear and nonlinear relationship with from the characteristics of credit. Finally, in a pat on the loan on the platform. The experimental results show that the proposed method can select the characteristics of credit primaries and full access to more than The characteristics of credit variable types and multiple relation types. (3) the application effect of network lending personal credit evaluation model of credit model in the existing network lending environment is poor, need to be improved and combined with the characteristics of the network lending environment credit showed on the existing model, design a new credit evaluation model and Adaboost model. The existing ensemble learning model only based on sample weight classification error rate adjustment of base classifiers, ignoring the influence of divergence and misclassification cost factors for sample weight, decrease the model accuracy after integration. Therefore, this paper proposes a model for divergence and misclassification cost Adaboost credit evaluation based on the model to classification difficult samples and high misclassification cost for targeted learning, improve the accuracy of credit evaluation results. In a pat on the loan on the platform. The experimental results show that the base The Adaboost model in credit evaluation of Adaboost divergence and misclassification cost model significantly outperforms the traditional. In addition, the network lending environment, according to the holographic credit the credit analysis mechanism of collaborative design of a Peer-to-Peer, acquisition and integration evaluation objects in multiple network platform related credit characteristics, the establishment of credit evaluation cross business collaborative analysis model based on the model, make the evaluation more comprehensive and credit of the borrower. The related network lending platform, e-commerce platform and social network platform. The experimental results indicate that the evaluation of the effectiveness of cross business credit cooperative analysis model can effectively improve the credit evaluation based on the results.

【学位授予单位】：合肥工业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：F832.4;F724.6

【参考文献】