Investigating the impact of undersampling and bagging: an empirical investigation for customer attrition modeling

Abstract

Given the growing interest in using AI and analytics to support CRM decision making, we discuss why undersampling and bagging are popular prediction techniques in customer churn prediction (CCP). The former helps in tackling the class imbalance problem and the latter improves model stability. However, extant CCP literature is unclear on the impact of undersampling on model stability and predictive performance, while bagging has difficulties in handling the class imbalance problem. Therefore, we extend existing CCP research to benchmark underbagging, which combines undersampling and bagging. Having both prediction techniques combined we recuperate customer data that would have been lost in undersampling by using them in multiple bags and passing an undersampled, more balanced training set to the classifier. In an extensive experiment including 11 real-life CCP datasets, underbagging is benchmarked against its constituents and other popular CCP classifiers in terms of predictive performance, profit and operational efficiency. Our results indicate that underbagging is a valid and reliable alternative framework for CCP prediction.