| In recent years,driven by the expansion of consumer spending and the increasing willingness to use credit for consumption,China’s consumer loans have grown significantly.Data shows that as of the end of December 2022,the balance of RMB consumer loans in China reached 56036.1 billion yuan.Due to the large number of customers,using manual customer review for customer default screening is inefficient.Therefore,it is crucial for the banking industry to establish a model with high accuracy to achieve automatic screening of customer default situations.Because user data is not easily available,bank customer data given in the ccf big data and computing intelligence contest is selected as the target domain,and the network public customer data set is the source domain for feature based transfer learning.Descriptive statistical analysis was conducted on important variables in the dataset,and missing value processing,data encoding,and data standardization preprocessing operations were performed on the data.Analyzing features reveals that the dataset contains a total of 35 features,and excessive features can reduce the spatial and temporal complexity of calculations and even affect the accuracy of prediction results.Therefore,feature screening is necessary.Through correlation coefficient method,variance filtering method,mutual information method,hypothesis testing method,Lasso CV method,and random forest method,the characteristics were screened,and interest,class,and employee were selected＿type 、 work＿Year and other 25 features were included in subsequent modeling analysis.Through summary statistics,it was found that the ratio of normal users to defaulting users in the dataset is approximately 4:1,which belongs to an unbalanced dataset.If direct modeling will lead to excessive attention to most samples and neglect of a few samples,reducing the accuracy of the model,so first divide the data set according to the ratio of 7:3.For the data in the training set,use SMOTE oversampling,Border Line SMOTE oversampling,and SMOTE Tomek mixed sampling methods to process unbalanced data.In the modeling phase,logistic regression,random forest,GBDT,XGBoost models are established for the sampled balance training set.In order to minimize the impact of sampling randomness on indicators,each model is established five times and the average evaluation indicator is calculated as the final measurement indicator of the model.In view of the problems existing in a single model,the model was improved and compared.It was proved that the logical regression,random forest and XGBoost integrated model based on SMOTE Tomek mixed sampling was an ideal model for predicting personal loan default.The accuracy of the model prediction was 81.2%,and the recall rate was 85%,indicating that the model could predict defaulting users well.Based on the calculated feature importance values,the most influential feature for customer default situations is employee＿type、class、recircle＿b、debt＿loan＿Ratio,banks can focus on these four variables when conducting customer screening. |
