XGBoost Ensemble Model for Churn Prediction in Telecom: A Machine Learning Framework

Abstract

Background: Customer churn remains a critical challenge in the telecommunications industry, where saturated markets and high acquisition costs demand advanced predictive models. Churn prediction plays an important role for retention-oriented companies, however, imbalance and complex behavioral patterns make traditional techniques inefficient.

Methods: Four machine learning algorithms namely, XGBoost, LightGBM, CatBoost, and Voting Ensemble were selected according to such criteria as accuracy, precision, recall, F1-score, and AUC-ROC. Furthermore, the methods of confusion matrix, ROC curve, correlation, and SHAP are utilized to analyze results.

Results: CatBoost algorithm achieved the highest values for all criteria (accuracy: 76.05%, F1: 60.77%), while Voting Ensemble obtained the highest AUC-ROC (82.31%). The model demonstrates balance (recall: 74.3%, precision: 67.2%) with slight inclination towards positive class. ROC analysis (AUC = 0.82.31) suggests the model's high predictive abilities. The most significant variables were contract, tenure, and support calls with noticeable non-linear effects.

Conclusion: From the research, it is evident that the machine learning algorithms based on boosting and ensembles perform well in predicting churn in data sets with an imbalance problem. In such algorithms, there is proper modeling of interaction among features without compromising the sensitivity to precision trade-off ratio.

Implications/Novelty: The current research aims at combining powerful ensembles with interpretable machine learning methods to ensure accuracy and interpretability. The identification of the factors that cause churn along with the relationships between them will help develop better approaches to address the problem.