Authors: Komal, Balvir Singh Thakur
Abstract: A major challenge for the telecommunication industry is the frequent customer switching, resulting in substantial revenue loss and increased customer acquisition costs. Most of the existing studies on churn prediction focus on the predictive accuracy and pay less attention to the interpretability of the model and how to convert the model output to business-facing visual tools. In this study, an integrated framework is proposed with XGBoost for churn classification, SHAP (Shapley Additive explanations) for global and local interpretability and Tableau for the interactive dashboard visualization. The Telco Customer Churn dataset with 7,043 customer records and 21 features is publicly available and is used for experimentation. The accuracy of the XGBoost model on a held-out test set of 1,409 customers is 80.41%, with better results on the majority (retained) class than on the minority (churned) class, which is consistent with the moderate imbalance between classes in the dataset. This outcome will be contextualized through baseline comparison against Logistic Regression and Random Forest. SHAP analysis reveals that contract type, tenure, monthly charges, total charges and internet service type are the dominant churn drivers globally, and local SHAP explanations illustrate how these factors combine for individual customers. In addition to the SHAP-ranked churn drivers, an interactive Tableau dashboard was created to visualize churn distribution by contract type, payment method, and internet service type for business-facing exploration of these insights. The study concludes that the combination of predictive accuracy with explainability and visualization leads to a more actionable churn-management framework than accuracy-oriented approaches alone.
