Explaining Telecom Customer Churn through Communication-Oriented Machine Learning and SHAP Analysis

Abstract

In subscription-based communication markets, customer churn threatens both revenue and long-term audience relationships. Although machine learning models are widely used to target at-risk subscribers, many implementations remain opaque and difficult to translate into communication strategies. This article analyzes an anonymized telecom customer churn dataset to examine how communication-related attributes of the customer relationship shape churn risk and how explainable machine learning can support retention communication. Logistic regression, random forest, and gradient boosting models are estimated on secondary data for 7,032 subscribers, with class imbalance handled through weighted training and performance evaluated on a held-out test set. Shapley Additive Explanations (SHAP) are then applied to the best-performing tree-based model, with predictors grouped into four constructs: relationship and contract attributes, digital communication and billing channels, support and value-added services, and demographic and household characteristics. The models achieve good discrimination, and SHAP analyses show that contract type, tenure, and billing-related monetary variables dominate churn prediction, followed by support and value-added services and digital billing channels, while demographics play a limited role. Aggregating explanations at the construct level yields stage-specific insights into early versus late relationship drivers of churn and illustrates how churn analytics can be aligned with media and communication theory to design more targeted retention communication.