Abstract
Acquiring consumers is a challenge in the telecom industry since it reduces profits and slows growth. Customer churn prediction remains a pivotal challenge for telecommunications companies seeking to retain subscribers and sustain profitability in highly competitive markets. This study introduces XCL-Churn, an explainable ensemble learning framework that integrates XGBoost, CatBoost, and LightGBM within a soft-voting meta-architecture to achieve robust and interpretable churn prediction. The proposed framework leverages a comprehensive data preprocessing pipeline involving iterative Bayesian Ridge imputation for missing data, multi-stage scaling (Robust, Standard, and Min-Max), and feature refinement through a hybrid Boruta-Random Forest (RF) strategy. Class imbalance is addressed using the Synthetic Minority Oversampling Technique (SMOTE), ensuring equitable learning across churn and non-churn classes. The model is trained and validated on the large-scale Expresso Churn Prediction Challenge dataset from Zindi, encompassing approximately two million customer records with heterogeneous demographic, behavioral, and financial attributes. The novelty of this work lies in its holistic integration of advanced data imputation, multi-stage scaling, and hybrid feature selection within an explainable soft-voting ensemble that merges three state-of-the-art gradient boosting models. Experimental results demonstrate that the XCL-Churn ensemble achieves superior predictive performance-97.44% accuracy, 93.82% precision, 87.82% recall, and 91.25% F1-score-while maintaining computational efficiency with a reduced training time compared to traditional stacking and blending strategies. Furthermore, the incorporation of Explainable AI (XAI) techniques-LIME for local interpretability and SHAP for global feature importance-enhances transparency by elucidating the key behavioral and financial indicators driving customer churn decisions.