Research On Customer Credit Utilization Prediction Based On Machine Learning And Interpretable Method

After a commercial bank’s online personal credit business customers have been approved for credit,how to predict whether the customers use their credit and the number of credit utilization days and combine the interpretable method of machine learning to perform an explanatory analysis on the prediction results,analyze the key influencing factors of customer use their credit,and carry out precision marketing for credit customers according to the results of explanatory analysis.It is of great significance to improve the overall customer credit utilization rate and profitability of bank related business.Around the two problems of credit utilization prediction and credit utilization prediction of days,this paper mainly carries out two aspects of research work:Firstly,in the aspect of credit utilization prediction,this paper constructs a DNN(Deep Neural Network)interpretable credit utilization prediction model based on decision tree rule mining.Firstly,interpretable business rules are mined based on the decision tree model through the rule mining method,and the mined rules are filtered by the credit utilization hit rate index.Then,the selected rules are derived into new features through the feature derivation method and input into DNN credit utilization prediction model for prediction.Experimental comparison and analysis are carried out on four main classification task performance evaluation indicators with five mainstream machine learning models.Because the rules themselves have excellent interpretability,the interpretability of the credit utilization prediction model can be improved in terms of pre-interpretability.The prediction accuracy of the model is improved by comparative experiments.The main advantage of this method is that it takes into account the interpretability and prediction performance of the credit utilization prediction model.Secondly,in terms of the credit utilization prediction of days.this paper constructs an interpretable credit utilization prediction model of days based on ensemble learning model combined with post-hoc interpretable SHAP(Shapley Additive explanations)method.Firstly,the CatBoost model is used as the base model to construct the credit utilization prediction model of days.and the four core parameters of the model are hyperparameter optimized by TPE(Tree-structured Parzen Estimator)hyperparameter optimization algorithm.Then it compared with five mainstream machine learning models on four main regression task performance evaluation indicators.Finally,it used SHAP post-hoc interpretable method to analyze the global level,local level and feature level of the credit utilization prediction model of days,and analyzed the key factors that affected how long the customer would use their credit.The main advantages of this method are that it enhances the interpretability of the credit utilization prediction model of days from the perspective of post-interpretability.and improves the prediction accuracy of the model through the hyperparameter optimization method and comparison experiments.In addition,this paper designs and implements a customer credit utilization prediction system.This system mainly realizes the analysis of uploaded customer data.relevant data preprocessing and feature engineering operations,and predicts whether the customer will use the credit through the main machine learning classification model,and gives the key factors affecting whether the customer will use the credit through the feature importance view.It provides visual analysis ability and decision-making basis for the relevant credit business personnel of banks to carry out precise marketing of credit customers.