Enhancing Telecom Customer Churn Prediction Using Voting Ensemble Learning with SMOTE and ADASYN

Abstract

Customer churn prediction is an important task in the telecommunications sector, particularly when class imbalance affects the ability of machine learning models to identify churned customers. This study evaluates the performance of several machine learning algorithms using the Telco Customer Churn dataset, which contains 7,043 customer records and 21 features. Five classification algorithms, namely Support Vector Machine (SVM), Decision Tree, Random Forest, XGBoost, and CatBoost, were evaluated. A Voting Ensemble model combining Random Forest, XGBoost, and CatBoost was also developed. To address class imbalance, SMOTE and ADASYN oversampling techniques were applied to the training data. The results showed that ensemble-based models generally outperformed individual classifiers. The Voting Ensemble model combined with SMOTE achieved the best performance, obtaining 97% accuracy, 0.95 precision, 0.93 recall, 0.95 F1-score, and 0.96 ROC-AUC. The findings indicate that integrating ensemble learning with oversampl