Application of Active Learning for Churn Prediction with Class Imbalance

Abstract

Churn prediction is a major focus that all the companies need to concern. Many studies have shown that class imbalance has a significant impact on churn prediction, but there is still no consensus on which technique is the best to cope with this issue. Recently, active learning has proved to be effective for imbalance learning. We try to apply it to churn prediction in this paper. In order to verify its effectiveness, we carry out experiments on six real-world data sets from the telecommunication industry and compare its performance with the other three benchmark resampling methods. Besides using the AUC to measure the accuracy of classification, we take the profit-based measure --- expected maximum profit (EMP) into account which measures the real cost and benefit produced by the models. The experimental results show that active learning is a good choice in dealing with the class imbalance problem in churn prediction.