A balanced transfer learning model for customer churn prediction

Abstract

Class imbalance presents significant challenges to customer churn prediction. Most of previous research efforts addressing class imbalance focus on the usage of in-domain information. In fact, due to the development of information technology, customer data of related domains may be gathered. These data come from different time-periods, districts or product categories. In this paper, we develop a new churn prediction model based on transfer learning model, which uses customer data from related domains to address the issue of data imbalance. The new model is applied to a real-world churn prediction problem in the bank industry. The results show that the new model provides better performance than traditional method such as resampling and cost-sensitive learning in dealing with class balance.