Research on personal credit scoring model based on multi-source data

Abstract

In the Internet financial personal credit loan business, it is necessary to construct a credit scoring model for users, and the problems of unbalanced user categories, high data dimensions and sparse features make it difficult to model the credit situation of users. This paper adopts the idea of grouping modeling. It proposes an improved BIV value feature screening method and a weighted average model based on Logistic Regression, Random Forest and Catboost, which provides a set of solutions for user modeling in this scenario. The grouping modeling idea pre-groups the customers and reduces the feature sparsity problem. The improved BIV value shows the influence of each feature on the results and points out the mutation threshold. The oversampling method alleviates the category imbalance problem. AUC is used as the model result evaluation index, and the results show that the classification effect of the model is good. The results show that customers with a long history of credit history and a history of good credit behavior have lower credit risk.