| Credit card is an important part of the modern financial industry,which can reduce social transaction costs and promote economic development to a certain extent.After years of rapid expansion,the growth rate of card issuance has slowed down significantly.The management of existing customers will become the main direction of banks’ future development.Existing customer management will become the main source of future growth.Therefore,it is very important to manage existing customer relationships and to identify customers who are prone to churn in a timely manner.To maintain the existing customer relationship,it is necessary to analyze the customer churn situation.Generally speaking,the churn of credit card customers is only a small part of all customers,so the credit card customer churn data set presents unbalanced characteristics.In previous studies,the prediction of credit card customer churn mostly focused on the classification model,and the classification effect was improved by optimizing the model.The treatment of unbalanced characteristics only used one or two resampling methods to change the data distribution.Few researchers build models from both the data preprocessing level and the classification algorithm level,and compare the pros and cons of the models.This paper uses the credit card customer churn data set from the LEAPS platform to conduct research and analysis.According to the unbalanced characteristics of the data set,a model is established from the two levels of data preprocessing and classification algorithm,and the resampling method on the bank customer churn data set is deeply analyzed.The degree of fit and applicability of the classifier model,and the integration of resampling methods into the ensemble learning model to explore the best modeling method for the data set.In terms of modeling,starting from the data preprocessing level,six sampling methods,namely random under-sampling,ENN,SMOTE,Borderline-SMOTE,SMOTE+ENN,SMOTE+Tomek Links,are selected to combine SVM and decision tree model to analyze the data.Experiments show that the effect of the model after resampling is better than that without sampling;among the six resampling methods,the SMOTE+ENN method has the highest recall rate,AUC value and G-mean value,and the effect is better;the decision tree classification effect is better than the SVM model.Then,starting from the classification algorithm level,the Bagging-based Balanced Random Forest(BRF)algorithm,the SMOTEBagging algorithm,the Boosting-based RUSBoost algorithm and the Double Ensemble-based Easy Ensemble algorithm are selected as classifiers for modeling,and the prediction effect and classification performance of each model are compared.Experiments show that among the four ensemble learning methods,the BRF algorithm performs better,and its performance is obviously better than the SVM and the decision tree.Finally,the feature importance ranking is calculated based on the BRF model,and the main factors affecting customer churn are mined.The algorithm selected in this paper has a good performance on the customer churn data set,and has certain practical significance for solving the imbalance problem of the credit card customer churn data set. |
