Explainable churn prediction in telecom with tabular ML five model benchmark and SHAP analysis

Abstract

Telecom operators need churn models that are both accurate and explainable for operational decision making. We study the Kaggle Telco Customer Churn dataset with a class ratio of approximately two point seven seven to one in favor of non churn. We benchmark five tabular learners namely Logistic Regression with Elastic Net, Support Vector Classifier with radial basis kernel, Random Forest, XGBoost, and a Neural Network based on a multilayer perceptron. Model selection uses nested cross validation with an inner search for hyperparameters and an outer five fold stratified evaluation that reports accuracy, precision, recall, F1 score, ROC AUC, and PR AUC. The multilayer perceptron attains 92.28% accuracy with precision 0.88, recall 0.82, and an F1 score of approximately 0.85. The remaining models deliver accuracy between 80 and 89%. Wilcoxon signed rank tests across folds show consistent gains for the multilayer perceptron. Two sided results are at the level of a trend while one sided tests that encode the directional hypothesis that the multilayer perceptron is better are significant. SHAP, partial dependence, and individual conditional expectation reveal a monotonic risk decrease with tenure with a tipping point around three or more years. Risk rises from zero to five technical tickets and then saturates and is amplified by higher monthly charges and month to month contracts with fiber users most sensitive. Recommended actions are contract migration, rapid issue resolution, and targeted credits for early life high charge customers with multiple tickets. Our contribution is a compact and reproducible pipeline that pairs strong tabular models with model agnostic explanations suitable for deployment and analyst use.