A Comparative Assessment of the Performance of Ensemble Learning in Customer Churn Prediction

Customer churn is a main concern of most firms in all industries. The aim of customer churn prediction is detecting customers with high tendency to leave a company. Although, many modeling techniques have been used in the field of churn prediction, performance of ensemble methods has not been thoroughly investigated yet. Therefore, in this paper, we perform a comparative assessment of the performance of four popular ensemble methods, i.e., Bagging, Boosting, Stacking, and Voting based on four known base learners, i.e., C4.5 Decision Tree (DT), Artificial Neural Network (ANN), Support Vector Machine (SVM) and Reduced Incremental Pruning to Produce Error Reduction (RIPPER). Furthermore, we have investigated the effectiveness of two different sampling techniques, i.e., oversampling as a representative of basic sampling techniques and Synthetic Minority Over-sampling Technique (SMOTE) as a representative of advanced sampling techniques. Experimental results show that SMOTE doesn't increase predictive performance. In addition, the results show that the application of ensemble learning has brought a significant improvement for individual base learners in terms of three performance indicators i.e., AUC, sensitivity, and specificity. Particularly, in our experiments, Boosting resulted in the best result among all other methods. Among the four ensemble methods Boosting RIPPER and Boosting C4.5 are the two best methods. These results indicate that ensemble methods can be a best candidate for churn prediction tasks