Research On Personal Credit Risk Evaluation Of Credit Card Based On Data Mining

In recent years,with the continuous development of China's economy and the continuous improvement of the income level of residents,the purchasing power of Chinese people is more and more strong.As one of the main methods of payment and settlement,the credit card has been widely popularized.Along with it,the credit card risk is escalating.It is of great significance for banks to carry out credit evaluation for credit card users to reduce credit risk and reduce losses.In the review of the domestic and foreign present researches for credit card risk assessment,the concept and characteristics of credit card risk,credit card risk types.Then,the advantages and disadvantages of different methods are analyzed.Aiming at the problem of credit evaluation for credit card application,an attribute reduction model based on random forest is proposed to avoid the slowness of modeling process caused by redundant attributes.In addition,in the process of dealing with a credit card application,most of the data is unbalanced.Therefore,according to previous data processing methods,the SMOTE-UNDER method is proposed to balance the data,and experiments are carried out on some of the UCI data sets,the experimental results show that the proposed method can effectively deal with unbalanced data classification problem.On this basis,a risk evaluation model based on GA-SVM is proposed.In this model,two important parameters of SVM are optimized by genetic algorithm,which can effectively improve the prediction accuracy.Aiming at the problem of cardholder risk classification,based on the analysis of the common credit risk types and the manifestations in the transaction behavior,the credit risk level segmentation index system based on the behavior of the cardholder is establised.As customer segments are generally done by clustering,based on the advantages and disadvantages of several traditional clustering algorithms,an improved objective clustering algorithm is set up to analyze the risk level of different categories of customers.On the basis of OCA clustering method proposed by academician of Ukraine Academy of Sciences,using K-means as the main means of realization,bringing into density based clustering idea,the Improved-OCA clustering method can not be limited by spherical clustering and the optimal number of clusters can be determined automatically.Using this method to cluster the cardholders based on their transaction behaviors,the cardholders are divided into 3 categories.Analyzing all kinds of features,the cardholders are labeled as low risk&medium value class,medium risk&high value class and high risk&low value class,so that banks can take different measures to prevent the risk of different groups of customers.Aiming at the phenomenon of overdue repayment in the process of the repayment of the cardholder,the two stage forecasting model based on BP neural network is proposed.Study on the prediction of the overdue problem mostly focused on whether the cardholders will overdue repayment,but overdue repayment is not equal to bad debts.Only customers overdue longer than a certain period of time is the real "bad" customers.Therefore,on the basis of the first stage classification of whether the cardholder will overdue,the cardholder overdue time is forecast.Only customers classified as overdue customers in the first stage,and then are forecast to be overdue beyond the acceptable period in the second stage,are risky customers who banks should pay more attention to.Considering that the BP neural network prediction results are affected by the initial weight and the results are unstable,some improvements are made to the traditional BP neural network.Using Adaboost method to training a number of weak classifiers and weak predictors,a strong classifier and a strong predictor are established,with which the first stage classification of whether customers will overdue,and the second stage prediction of customers'overdue periods are finished.Finally,a numerical experiment is carried out on the repayment of the cardholders of a bank,and the results show that the two stage of the model has high prediction accuracy.