Enhancing XGBoost Performance for Customer Churn Prediction using SMOTE-ENN Hybrid Resampling

Customer churn (when clients stop using a company's services) is a major threat to profitability in the telecommunications sector, where annual churn rates can hit 20% to 40%. Consequently, predicting churn accurately is vital for customer retention. This study develops a highly effective machine learning model to predict churn using the XGBoost algorithm. Based on the IBM Telco Customer Churn dataset (7,043 records), the data initially showed a significant class imbalance (73.5% non-churners vs. 26.5% churners). To resolve this, the researchers applied the SMOTE-ENN hybrid resampling method to balance the distribution. Afterward, GridSearchCV was utilized to systematically find the most optimal hyperparameter configuration. The optimized XGBoost model demonstrated exceptional performance, achieving 96.02% accuracy, 94.39% precision, 90.34% recall, a 92.32% F1-score, and an AU-PRC of 97.64%. By successfully combining XGBoost with SMOTE-ENN and GridSearchCV, this study outperforms previous research, providing a robust, highly accurate predictive model that companies can practically implement to identify at-risk customers.