An adaptive analytics framework for customer retention through integrative feature optimization and ensemble learning

Abstract

An adaptive analytics workflow is presented for customer churn prediction, combining Principal Component Analysis for dimensionality reduction, a hybrid Modified Particle Swarm Gravitational Search Optimization (MPSO-GSO) for feature selection and hyperparameter tuning, and an ensemble learning stage combining XGBoost and LightGBM through weighted voting. Applied to an e-commerce dataset, the complete framework achieves AUC = 0.99 and accuracy = 0.98, outperforming standalone XGBoost (AUC = 0.98) and LightGBM (AUC = 0.97). Stratified 5-fold cross-validation and paired t-tests confirm the statistical significance of this improvement (p < 0.01). Subsequent SHAP analysis interprets the feature contributions, demonstrating that this integrative, optimization-based approach substantially improves the quality of churn prediction.