Improving Customer Churn Detection Through Balanced Ensemble Learning

Abstract

Predicting customer churn represents a major challenge for telecommunication providers, driven by fierce market competition and frequent customer switching that can significantly threaten long-term revenue stability. Failure to accurately identify customers with high churn potential often leads to ineffective retention strategies. This study examines the effectiveness of integrating data balancing techniques with ensemble learning models to enhance churn prediction performance on imbalanced datasets. A quantitative experimental method is applied using a publicly available telecommunications dataset. The preprocessing phase focuses on handling incomplete records, transforming categorical attributes into numeric representations, and scaling feature values to improve data quality. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) is applied exclusively to the training data. The study evaluates three classifiers, including Logistic Regression as a baseline and two ensemble methods, Extreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LightGBM). Model performance is examined using several evaluation metrics such as accuracy, precision, recall, F1-score, and the Area Under the ROC Curve (AUC). The results reveal that ensemble learning approaches outperform Logistic Regression, particularly with respect to recall and AUC performance. LightGBM achieves the best overall performance and demonstrates stable predictive capability across all evaluation measures. Feature importance analysis reveals that customer tenure and billing-related attributes, including monthly charges and total charges, are dominant factors influencing churn behavior. These results demonstrate that integrating data balancing techniques with ensemble learning methods offers a robust and effective solution for supporting proactive customer retention initiatives in the telecommunications sector.