Research On Reject Inference In Personal Credit Evaluation

With the rapid development of China’s economy,people’s consumption concept has gradually changed,and the share of consumer credit activities and personal loans is gradually increasing.In 2020,the scale of personal consumption loans in China will exceed 15 trillion yuan,a year-on-year increase of 12.14%.At the same time,thanks to the continuous development of computer technology,China’s financial industry is transforming to digitalization and informationization,and financial lending has broken the traditional geographical restrictions,and more people have access to credit.However,due to the large number of borrowers in personal lending business,and each person’s reputation,income and asset level are different,the personal credit information collected by financial institutions mainly comes from the initiative of borrowers,which leads to certain information asymmetry,and the credit risk problem in financial lending is still outstanding.At the same time,China’s credit system construction started late.Although it is developing rapidly,it is still relatively backward,and the ability of preventing borrowers’ credit risks is weak,which restricts the development of personal credit business of financial institutions in China.China’s mainstream credit scoring methods mainly include drawing lessons from FICO credit evaluation model or credit scoring methods based on statistical analysis.Related research on rejection reasoning is essentially based on the direct application of traditional statistical methods or existing machine learning methods without extensive optimization or improvement of specific models or questions.Compared with traditional statistical methods,machine learning technology is more suitable for the characteristics of personal credit big data,which is also the development trend of the industry in the future.Therefore,establishing a scientific and effective model for personal credit scoring is of very practical importance.,so as to reasonably quantify and control the credit risk and correctly identify the borrower’s default degree,and it is also of theoretical value to enrich the research of personal credit evaluation in China’s financial institutions.Firstly,this paper starts with the concept of credit risk,and analyzes the present situation and causes of credit risk of lending by financial institutions in China,which lays a realistic and theoretical foundation for building a reasonable personal credit evaluation model.Secondly,this paper adopts the rejection inference technology in personal credit evaluation,and adds unlabeled samples to the model construction,which can solve the sample selection bias.At present,there are some statistical and machine learning methods to infer the status of rejected samples.In this paper,a new rejection inference method based on joint distribution adaptation and cost-sensitive semi-supervised support vector machine is proposed.This method aims at the common problem of different distribution between accepted sample set and rejected sample set in credit evaluation,and uses joint distribution adaptation to make the distribution of rejected sample set close to that of accepted sample set.By extracting the common features of accepted sample set and rejected sample set,the distribution difference between them is reduced,and the utilization efficiency of rejected samples is greatly improved.Thirdly,this paper adds the cost-sensitive algorithm to the construction of the model,aiming at solving the problem that the cost of misclassified defaulters in credit evaluation is far greater than that of correctly classified people who pay on time,and reducing the total cost of credit evaluation.Finally,according to the borrower’s credit data selected in this paper,the data are explained and analyzed from four aspects:the borrower’s basic information,job information,asset status and loan information.At the same time,combined with the data characteristics,this paper not only selects the traditional credit scoring model evaluation indicators such as accuracy rate,but also selects the evaluation indicators such as F1 value and G-mean of dealing with unbalanced data and giving different classification results different costs,thus forming a complete set of personal credit evaluation indicators system.Through three experiments on China’s credit data set,the proposed method is compared and verified,and the results verify the effectiveness of the proposed rejection inference method based on joint distribution adaptation and cost-sensitive semi-supervised support vector machine.Combined with the conclusion of this paper and the situation of personal credit business of financial institutions in China,targeted policy suggestions are put forward.