Cost-Sensitive Semi-Supervised Ensemble Model for Customer Churn Prediction

Abstract

Customer churn prediction is a vital part of customer relationship management. In real churn prediction modeling, the process of assigning labels to customers is costly and time-consuming, and thus there are more unlabeled than labeled samples. To solve this problem, this study combines semi-supervised learning, cost-sensitive learning method Metacost with ensemble method random subspace to propose a semi-supervised ensemble model based on Metacost. This model includes the following three stages: (1) use the Metacost method to modify the class label of the initial labeled training set to obtain a new training set; (2) label the unlabeled samples selectively; (3) train several basic classifiers with the final training set and use them to classify the samples in the test set. The empirical results for two customer churn prediction data sets demonstrate that the proposed model raises customer churn prediction performance compared with commonly used supervised ensemble and semi-supervised models.