Comparative Analysis of Machine Learning Models for Telecom Customer Churn Prediction

Abstract

Customer churn has consistently been a significant issue within the telecommunications sector. The timely and accurate identification of high-risk churn customers is essential for improving customer satisfaction and operational profitability. With the rise of data-driven decision-making, predictive modelling has become a critical tool for telecom operators to mitigate churn risks. This study leverages the publicly available Telco Customer Churn dataset from Kaggle and employs R to construct and evaluate three classical machine learning models: Logistic Regression, Decision Tree, and K-Nearest Neighbours (KNN). Then, after preprocessing, feature selection, and hyper-parameter tuning, the models are evaluated with score metric and multi-metric evaluation to predict customer churn. The results of the research showed that it is 81.1% accurate, 72.5% precise in accuracy and interpretability compared to other models since it is a very interpretable model, as you can see from the Receiver Operating Characteristic (ROC) curve also. All models have specific weaknesses in predicting churned users; however, as a whole, they manage to achieve good accurate results. This study can offer an empirical and technical foundation for southern industrial companies to analyze churn and do data management, which is vital in the telecom industry.