Extreme gradient boosting trees with efficient Bayesian optimization for profit-driven customer churn prediction

Customer retention campaigns increasingly rely on predictive analytics to identify potential churners in a customer base. Traditionally, customer churn prediction was dependent on binary classifiers, which are often optimized for accuracy-based performance measures. However, there is a growing consensus that this approach may not always fulfill the critical business objective of profit maximization, as it overlooks the costs of misclassification and the benefits of accurate classification. This study adopts extreme gradient boosting trees to predict profit-driven customer churn. The class weights and other hyperparameters of these trees are optimized using Bayesian methods based on the profit maximization criterion. Empirical analyses are conducted using real datasets obtained from service providers in multiple markets. The empirical results demonstrate that the proposed model yields significantly higher profits than the benchmark models. Bayesian optimization and adjustment of class weights contributed to enhanced model profitability. Furthermore, when optimizing multiple hyperparameters, the computational cost of model optimization is significantly reduced compared with an exhaustive grid search. Additionally, we demonstrate the robustness of the proposed model through a sensitivity analysis employing Bayesian optimization. Using the proposed model, marketing managers can design targeted marketing plans to retain customer groups with a higher likelihood of churning.