Feature-selection-based dynamic transfer ensemble model for customer churn prediction

Customer churn prediction is one of the key steps to maximize the value of customers for an enterprise. It is difficult to get satisfactory prediction effect by traditional models constructed on the assumption that the training and test data are subject to the same distribution, because the customers usually come from different districts and may be subject to different distributions in reality. This study proposes a feature-selection-based dynamic transfer ensemble (FSDTE) model that aims to introduce transfer learning theory for utilizing the customer data in both the target and related source domains. The model mainly conducts a two-layer feature selection. In the first layer, an initial feature subset is selected by GMDH-type neural network only in the target domain. In the second layer, several appropriate patterns from the source domain to target training set are selected, and some features with higher mutual information between them and the class variable are combined with the initial subset to construct a new feature subset. The selection in the second layer is repeated several times to generate a series of new feature subsets, and then, we train a base classifier in each one. Finally, a best base classifier is selected dynamically for each test pattern. The experimental results in two customer churn prediction datasets show that FSDTE can achieve better performance compared with the traditional churn prediction strategies, as well as three existing transfer learning strategies.