Comparative Analysis of Machine Learning Algorithms for Customer Churn Prediction in the Telecommunications Industry

Abstract

This study jointly benchmarks Logistic Regression, Random Forest, and Support Vector Machine (SVM) for customer churn prediction and links the resulting feature importance profile to retention strategy. The IBM Telco Customer Churn dataset (7,043 records, 20 predictor features covering demographic, account, and service attributes) was preprocessed through missing value imputation, label encoding, and feature standardization, then split into 80% training and 20% testing subsets using stratified sampling. Model performance was evaluated using accuracy, precision, recall, F1-score, and Area Under the ROC Curve (AUC). Recall and AUC were prioritized because a missed churner costs more than an unnecessary retention offer, while AUC summarizes ranking quality across all intervention thresholds. Results show that Random Forest achieved the highest accuracy (80.62%) and AUC (0.842), followed closely by Logistic Regression, which recorded a slightly lower accuracy (79.99%) but the highest recall (0.548), while SVM obtained the lowest AUC (0.792) despite comparable accuracy. Random Forest feature importance analysis indicated that contract type, tenure, total charges, and monthly charges were the strongest predictive attributes, together accounting for more than half of the model’s total importance; these are predictive associations rather than causal determinants of churn. These findings indicate that tree based ensembles offer a marginal advantage over linear and margin based classifiers on structured churn data, and that contract and tenure related variables deserve priority in retention strategy. The observed differences between models are descriptive, as performance was estimated from a single stratified train test split with lightly tuned hyperparameters and no significance testing.