Optimization of Predictive Ensemble Model for Customer Churn Through the Integration of SMOTE-ENN and GridSearchCV

Customer churn is a phenomenon where customers terminate their subscription with a company's offerings and switch to another provider. In the telecommunications industry, churn presents a significant challenge as customer acquisition costs are higher than retention costs. In developing churn prediction models, a major challenge is the class imbalance problem, where a significantly higher number of non-churn instances compared to churn instances leads to biased model predictions that favor the majority class, thereby reducing the overall performance and reliability of the predictive model. In response to this challenge, this study proposes an ensemble machine learning approach using Extreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LightGBM), enhanced through the application of Synthetic Minority Over-sampling Technique combined with Edited Nearest Neighbors (SMOTE-ENN) and hyperparameter optimization via Grid Search Cross-Validation (GridSearchCV). The integration of SMOTE-ENN effectively addresses class imbalance by generating synthetic minority samples and eliminating noisy or borderline instances, while GridSearchCV systematically identifies the optimal parameter configurations. Experimental results demonstrate that the optimized XGBoost and LightGBM models achieved high prediction accuracies of 96.49% and 96.32%, respectively. These findings confirm that the proposed approach significantly enhances model performance and offers a robust framework for developing accurate and reliable churn prediction systems. This approach is distinctive in its integration of SMOTE-ENN and GridSearchCV within a single pipeline, enabling the model to address class imbalance and perform hyperparameter optimization. Such an integrated design contributes to improved classification performance, particularly in predictive modeling tasks for churn affected by class imbalance.