In the contemporary subscription-based economy, customer retention is structurally paramount to the long-term viability and profitability of telecommunications enterprises. The financial asymmetry between Customer Acquisition Cost (CAC) and Customer Retention Cost (CRC) mandates the development of highly accurate, proactive customer churn prediction systems. This comprehensive research paper presents an in-depth empirical analysis of machine learning algorithms deployed to forecast customer defection. We rigorously evaluate a spectrum of predictive models, progressing from traditional linear classifiers (Logistic Regression) to complex, non-linear distance-based models (Support Vector Machines), and state-of-the-art ensemble architectures (Random Forest and eXtreme Gradient Boosting). A critical challenge in churn analytics—the severe class imbalance inherent in real-world datasets—is addressed through the application of the Synthetic Minority Over-sampling Technique (SMOTE) combined with rigorous cross-validation strategies. Utilizing a robust telecommunications dataset, our extensive feature engineering and hyperparameter optimization reveal that tree-based ensemble methods, particularly XGBoost, significantly outperform baseline models. XGBoost achieved a superior Area Under the Receiver Operating Characteristic Curve (ROC-AUC) of 0.88, optimizing the critical trade-off between Precision and Recall. Furthermore, the integration of SHapley Additive exPlanations (SHAP) provides a granular, interpretable analysis of feature importance, identifying contract duration, customer tenure, and specific service combinations as the primary drivers of attrition. The findings equip business stakeholders with interpretable, high-fidelity predictive intelligence required to operationalize targeted, cost-effective customer retention interventions.
Pandey et al. (Fri,) studied this question.