Hybrid AI framework for customer churn classification and business risk assessment

Customer turnover is a major problem in the telecom sector, Consequently, identifying clients who are most likely to depart might aid in retention decisions. Single-model methods might lack the flexibility to represent complex customer behavior, and class imbalance can impact the identification of churn cases. This study introduces a framework for churn prediction based on machine-learning (ML) techniques that employs CatBoost, LightGBM and XGBoost as base learners and Logistic Regression as a meta-learner in the Stacking Ensemble. The framework utilizes data pre-processing, domain knowledge-based feature engineering, outlier detection, and SMOTE class imbalance techniques for enhancing predictive learning. Feature engineering and SMOTE are incorporated to improve the learning process. Experimental results demonstrate that Stacking Ensemble obtains an accuracy of 81.10%, a recall of 60.86%, an F1-score of 63.08%, and an ROC-AUC of 84.60%, whereas XGBoost scores the best ROC-AUC of 84.63%. Furthermore, churn probabilities are predicted and utilized for customer risk segmentation and financial impact assessment, helping to identify customer groups that have a higher potential revenue exposure. In summary, proposed framework shows the potential of ensemble ML for churn prediction and offers data-driven insights to help companies prioritize customer retention and protect revenue in the telecom industry.