Feature selection based transfer ensemble model for customer churn prediction

It is difficult to get satisfactory customer churn prediction effect for the traditional model, because the class distribution of customer data is often imbalanced, and the available data in target task is little. This paper combines the transfer learning with the ensemble learning, and proposes a feature selection based transfer ensemble model (FSTE). It utilizes the customer data in both the related source domain and target domain, selects a series of feature subsets, obtains the corresponding training subsets by mapping; further, it trains a number of classifiers and gets the final customer churn prediction result by integrating the prediction results. The empirical results show that FSTE can achieve better customer churn prediction performance compared with the traditional churn prediction model, and some existing transfer learning models such as TFS, TrBagg and TrAdboost.