A Comparative Study of Feature Selection and Resampling Techniques for Customer Churn Prediction with Explainable AI in the Telecommunications Sector

This study presents an integrated modelling pipeline for churn prediction that includes preprocessing, sampling, feature selection, and explainable AI (XAI) techniques. The dataset, constructed by merging six separate files, was preprocessed through contextual imputation, removal of redundant features, categorical encoding, and Min-Max normalization. Three sampling strategies—random oversampling, undersampling, and SMOTE—were applied and visually analyzed for distribution improvements to address the class imbalance.A key contribution of the study is the comparative evaluation of three feature selection methods (Boruta, a genetic algorithm-based machine learning approach (GA-ML), and a hybrid method that combines genetic algorithms with deep learning (GA-ML+DL)) under different sampling conditions. Genetic algorithm-based approaches were enhanced using varied fitness functions and tested in hybrid setups involving both traditional machine learning models (Random Forest, XGBoost, CatBoost) and a one-dimensional convolutional neural network. In addition, explainable AI methods such as LIME and SHAP were employed to increase model transparency and visualize the impact of selected features. These steps improved prediction accuracy and dataset balance and enhanced interpretability, making the models more trustworthy and insightful for decision-makers.