Prediction Of Bank Credit Card Customer Churn Based On Stacking Model Fusion

With the rapid development of information science and technology,the competition between commercial banks is increasing.On the one hand,In the Internet era with information openness,consumers can quickly obtain public information about financial products and services and optimize purchase strategies,which may cause some of the customers of a bank to flow to other banks.On the other hand,the personalized needs of consumers may also lead to the customer loss of some banks.In order to reduce customer loss as much as possible,we can make use of the bank's customer information for predicting potential customers loss so that bank managers can improve customer service in the follow-up work.In this thesis,data mining and data modeling are used to analyze the past customer data information collected by the bank,and a predicting method of bank credit card customer loss based on the fusion of SMOTE over-sampling and Stacking model is put forward:(1)Preliminary exploratory analysis on the original dataset was conducted(mainly including one-hot encoding of classification variables,standardization of numerical variables,data visualization and other operations),providing a breakthrough point for bank managers to find strategies to maintain customers from an intuitive perspective.(2)SMOTE over-sampling method was used to balance the original data.Random forest(RF),XGBoost,LightGBM and CatBoost models were gradually established with the balance data.Comparing the evaluation parameters of each single model,it found that fl-score of each model is above 96%.It shows that each single model has a good identification effect of customer loss.Specially,catBoost model is the best.(3)A Stacking model was builded,with random forest(RF),XGBoost,LightGBM,CatBoost as the first-layer base learners and Logistic regression model as the second-layer learner.Grid search and cross validation were used for model training.The Recall,AUC,F1-Score and other evaluation parameters of Stacking model are higher than those of other single algorithm models.Therefore,it is believed that the bank credit card model predicting customer loss based on Stacking model can identify potential customers loss well.