Study On The Personal Credit Assessment Methods In The Context Of Online Lending Businesses

Credit assessment in online lending businesses can effectively remit the information asymmetry among traders, reducing the default risk and transaction cost. However,under the online lending context, it is hard to obtain and verify the financial information of borrowers, which bring enormous difficulties to traditional financial information-based credit assessment methods. In fact, in network environment, the creditworthiness-related data of borrowers includes not only financial information but also non-financial information. These non-financial information distribute in various network platforms,and have the attributes of big volume, low profit density, and uneven quality, taking new difficulty for credit assessment of online lending businesses. Therefore, in this thesis,we review the relevant theory and methodology of credit assessment, and consider the features of online lending business. Based on these, we study the problem of personal credit assessment in network context from three aspects: the preprocessing of creditworthiness-related data, the feature selection of creditworthiness, and the credit assessment model construction. The specific research contents of this thesis are shown as following.(1) Data preprocessing methods in online lending businessesThe quality of credit data in online lending businesses is uneven. The missing value and outliers are extremely common. In the aspect of missing value processing, due to the multiple imputation method cannot deal with missing values in the data set which contain category variables, we propose a classification multiple imputation method.This method can improve the effect of missing value imputation by using the relationships between category variables and continuous variables on the parameter estimation of mathematical characteristics of variables. In the aspect of outliers processing, to deal with outliers on a single credit feature, we propose a KNN-based outliers rectification method. This method can correct the value of outliers through related credit feature of K-nearest neighbors. Because the distance threshold of abnormal samples is hard to confirm under non-uniform density distribution space, we propose a DBSCAN and relative density-based abnormal sample detection method. This method use DBSCAN algorithm to divide the non-uniform density distribution space into several uniform density space; and then use relative density metric to identify abnormal samples. We conduct experimental study regarding credit data preprocessing in PPDai platform. The results show that the processed data can significant improve the effects of credit assessment model.(2) Feature selection methods of creditworthiness in online lending businessesThe credit data in online lending business has the attributes of big volume and low profit density. Therefore, we need to select credit feature based on relevant theory and methodology. In the stage of qualitative primary selection of credit feature, for the social capital nature of creditworthiness, we analyze the social capital of borrowers in online lending businesses from the structural dimension, relational dimension, and cognitional dimension based on social capital theory. Then, we propose a credit feature selection method that fusion of social capital information, personal information, credit history information, and identity information of borrowers. In the stage of quantitative final selection of credit feature, because the variable types of credit features are vary and the relationships between credit features and credit target variable are different, we propose a comprehensive quantitative analysis-based credit feature selection method.This method can select the category and continuous credit features which have linear and non-linear relationships with credit target variable by using many kinds of quantitative analysis method, such as correlation analyses, Chi-square statistic analyses,information-gain analyses, and support vector regression analyses. The experimental results on PPDai platforms show that our method can comprehensive obtain the related credit features.(3 ) Credit assessment model in online lending businessesThe application effect of existing credit assessment model is not satisfying in online lending businesses. We need to improve the existing model and consider the attributes of creditworthiness to design new credit assessment pattern and model. Existing Adaboost ensemble learning model updates the sample weighs in basic classifiers just according to the rate of misclassification, which ignores the effect of disagreement and misclassfication cost on sample weights, leading to the decline of accuracy of ensemble model. Therefore, we propose a disagreement and misclassfication cost based Adaboost credit assessment model. This model can improve the accuracy of Adaboost model by focusing on the samples that are difficult to classify and have high misclassfication cost.The experimental results on PPDai platform show that the effects of our model are significant better than that of traditional Adaboost model. In addition, considering to the holographic attribute of creditworthiness,we design a Peer-to-Peer collaborative credit analysis mechanism, acquiring and integrating the relevant credit features of evaluation object in various network platforms, and constructing a cross-business credit assessment model based on collaborative analysis. Based on this, we construct a comprehensive evaluation on the creditworthiness of borrowers in online lending businesses platforms.We verify the validity of our cross-business credit assessment model through experimental study on relevant online lending platform, e-commerce platform and social network platform.