Bagging and boosting classification trees to predict churn

In this paper, the bagging and boosting techniques are proposed as performing tools for churn prediction. We apply these algorithms on a customer database of an anonymous U.S. wireless telecom company. Bagging is easy to put in practice and, as well as boosting, leads to a significant increase of the classification performance when applied to the customer database. Furthermore, we compare bagged and boosted classifiers respectively computed from a balanced versus a proportional sample to predict a rare event (here, churn), and propose a simple correction method for classifiers constructed from balanced training samples