Abstract:

Customer Churn is now becoming a significant problem in the banking sector. It is necessary to seek solutions to predict the rate of customer churn in banks; however, the dataset for customer churn prediction in banks is imbalanced. In this paper, Random Forest (RF) based on two popular resampling techniques, named SMOTE and ADASYN, are used to obtain a banking customer churn prediction model. A wide range of metrics, including Accuracy, Recall, Precision, Specificity, F1 score, Mathews correlation coefficient, and ROC-AUC, are used to comprehensively evaluate the prediction model. Through the experimental results, the values of Accuracy and ROC-AUC of the RF model based on SMOTE and ADASYN indicate positive results. Moreover, this paper also shows feature importance in the dataset based on the RF algorithm.