Handling class imbalance in customer churn prediction

Customer churn is often a rare event in service industries, but of great interest and great value. Until recently, however, class imbalance has not received much attention in the context of data mining [Weiss, G. M. (2004). Mining with rarity: A unifying framework. SIGKDD Explorations, 6 (1), 7-19]. In this study, we investigate how we can better handle class imbalance in churn prediction. Using more appropriate evaluation metrics (AUC, lift), we investigated the increase in performance of sampling (both random and advanced under-sampling) and two specific modelling techniques (gradient boosting and weighted random forests) compared to some standard modelling techniques. AUC and lift prove to be good evaluation metrics. AUC does not depend on a threshold, and is therefore a better overall evaluation metric compared to accuracy. Lift is very much related to accuracy, but has the advantage of being well used in marketing practice [Ling, C., & Li, C. (1998). Data mining for direct marketing problems and solutions. In Proceedings of the fourth international conference on knowledge discovery and data mining (KDD-98). New York, NY: AAAI Press]. Results show that under-sampling can lead to improved prediction accuracy, especially when evaluated with AUC. Unlike Ling and Li [Ling, C., & Li, C. (1998). Data mining for direct marketing problems and solutions. In Proceedings of the fourth international conference on knowledge discovery and data mining (KDD-98). New York, NY: AAAI Press], we find that there is no need to under-sample so that there are as many churners in your training set as non churners. Results show no increase in predictive performance when using the advanced sampling technique CUBE in this study. This is in line with findings of Japkowicz [Japkowicz, N. (2000). The class imbalance problem: significance and strategies. In Proceedings of the 2000 international conference on artificial intelligence (IC-AI'2000): Special track on inductive learning, Las Vegas, Nevada], who noted that using sophisticated sampling techniques did not give any clear advantage. Weighted random forests, as a cost-sensitive learner, performs significantly better compared to random forests, and is therefore advised. It should, however always be compared to logistic regression. Boosting is a very robust classifier, but never outperforms any other technique. (C) 2008 Elsevier Ltd. All rights reserved.