A Comparative Analysis of Machine Learning Models for Customer Churn Prediction in Subscription-Based Businesses

Customer churn prediction remains a critical task for subscription-based businesses aiming to enhance customer retention and reduce revenue loss. This study presents a comparative analysis of four machine learning models: Logistic Regression, Random Forest, XGBoost, and Artificial Neural Network applied to the Telco Customer Churn dataset. Extensive data preprocessing and feature engineering were performed, followed by rigorous performance evaluation using accuracy, precision, recall, F1-score, and ROC-AUC metrics. Among the models, XGBoost achieved the highest performance, with an accuracy of 87.5%, an F1-score of 82.8%, and an ROC-AUC of 92.4%. Additionally, SHAP (Shapley Additive Explanations) was used to enhance interpretability, highlighting that contract type, monthly charges, and customer tenure were key predictors of churn. The study provides actionable insights and validates XGBoost as a robust and interpretable solution for real-world churn prediction challenges.