A Dynamic Transfer Ensemble Model for Customer Churn Prediction

Abstract

It is difficult to get satisfactory churn prediction results by traditional models, because the available customer samples in target domain are usually few and the class distribution of customer data is imbalanced. This study proposes a group method of data handling (GMDH) based dynamic transfer ensemble (GDTE) model for churn pre-diction. It first transfers the data in related source domains to the target domain by transfer learning technique, and then adopts resampling technique to balance the class distribution of the training data. Finally, it trains a series of base classifiers and dynamically selects a proper classifier ensemble for each test sample by GMDH. The experimental results in two datasets show that the performance of GDTE is better than that of one traditional churn prediction strategy, as well as three transfer learning strategies.