Improved Decision Tree, Random Forest, and XGBoost Algorithms for Predicting Client Churn in the Telecommunications Industry

Abstract

Traditional machine learning models, especially decision trees, face great challenges when applied to high-dimensional and imbalanced telecommunication datasets. The research presented in this paper aims to enhance the performance of traditional Decision Tree (DT), Decision Tree with grid search (DT+), random forest (RF), and XGBoost (XGB) models. This is accomplished by augmenting them with robust preprocessing techniques, as well as optimizing them through grid search. We then evaluated how well the enhanced models can accurately predict customer churn and compared their performance metrics in detail. We utilized a dataset derived from the benchmark Cell2Cell dataset by applying combined preprocessing methods including KNN imputation, normalization, and resampling with SMOTE Tomek to address class imbalance. The findings reveal that XGBoost outperformed all other models with an accuracy of 0.82, demonstrating strong precision, recall, and F1 scores. RF also delivered robust results, achieving an accuracy of 0.82, benefiting from its ensemble nature to improve generalization and reduce overfitting.