Bagging and Boosting Classification Trees to Predict Churn
Published 5/1/2006
Author: Aurélie Lemmens and Christophe Croux
View this content
Executive Summary
This study considers a well-known binary choice problem, namely, predicting customer churn behavior. In this article, the authors bring to the attention of marketers the “bagging” and “boosting” classification models, originating from the statistical machine-learning literature. Bagging (Breiman 1996) consists of sequentially estimating a binary choice model—called the “base classifier” in machine learning—from resampled versions of a given calibration sample. The obtained classifiers form a committee from which a final choice model can be derived by simple aggregation. Although bagging is simple and easy to use, more sophisticated variants also exist. “Stochastic gradient boosting” (Friedman 2002) is one of the latest developments; it includes weights in the resampling procedure.
When applied to a customer database of an anonymous U.S. wireless telecommunications company, both bagging and boosting significantly improve accuracy in forecasting churn. When predicting rare events such as churn, the Gini coefficient and the top-decile lift are more appropriate performance criteria than the error or rate. The gain in predictive performance has reached 16% for the Gini coefficient and 26% for the top-decile lift. This higher predictive performance could ultimately lead to incremental profits for companies willing to use these methods. In addition, bagging and boosting provide good diagnostic measures, the importance of variables, and partial dependence plots, which offer some face validity to the models and interesting insights into potential churn drivers.
Furthermore, the results illustrate that the use of a balanced sampling scheme is recommended and preferred to proportional sampling when attempting to predict a rare event. However, to maintain the classification error rate at a reasonable level, it is necessary to correct the predictions obtained from a balanced sample. The intercept correction constitutes an appropriate bias correction for a balanced sampling scheme.
Biography
Aurélie Lemmens is doctoral candidate at the K.U. Leuven (Belgium). She is currently working on marketing modeling techniques. She focuses on time-series analysis and choice modeling and their marketing applications. Part of her dissertation has been published in International Journal of Forecasting.
Christophe Croux is Professor of Statistics and Econometrics at K.U. Leuven (Belgium). His research interests are robust statistics, multivariate data analysis, classification, computational statistics, and applied time-series analysis. His work has been published in Biometrika, Journal of the American Statistical Association, Journal of Multivariate Analysis, and Review of Economics and Statistics, among others. He serves on the editorial board of Journal of the American Statistical Association and Computational Statistics and Data Analysis.
J Marketing Research, Volume 43, Number 2, May 2006
View Table of Contents.