Customer Churn Prediction Using Decision Tree and Random Forest with a SMOTE Approach to Support Customer Intelligence in the Telecommunications Industry

Abstract

Customer churn represents a critical issue in the telecommunications sector due to its impact on customer retention levels and long-term business continuity. Consequently, companies require reliable methods to recognize customers who are likely to terminate their subscriptions. This study focuses on predicting customer churn by applying Decision Tree and Random Forest algorithms to the Telco Customer Churn dataset. The research methodology adopts the CRISP-DM framework, which encompasses data preparation, feature engineering, class balancing through the Synthetic Minority Over-sampling Technique (SMOTE), model construction, and performance evaluation. Four classification approaches were examined, including Decision Tree Gini, Decision Tree Entropy, Decision Tree Pre-Pruning, and Random Forest. Hyperparameter tuning was performed using GridSearchCV, whereas model effectiveness was assessed through Accuracy, Precision, Recall, F1-Score, and ROC-AUC metrics. The experimental results reveal that the Random Forest model produced the highest performance, achieving an accuracy of 84.88% and a ROC-AUC value of 92.94%. Furthermore, the feature importance analysis identified Contract, MonthlyCharges, and tenure as the variables with the strongest contribution to churn prediction. These findings indicate that the proposed approach can enhance Customer Intelligence by generating valuable insights into customer behavior and assisting organizations in developing more effective customer retention initiatives