Explaining and Predicting Mobile Network–Driven Churn Using Machine Learning and SHapley Additive exPlanations–Based Attribution

Customer churn remains a critical concern in the mobile telecommunications industry, where retaining subscribers is essential for long-term profitability. This study presents an explainable machine learning framework to detect and interpret network-driven churn by integrating demographic data with key network performance indicators, including signal quality metrics e.g., reference signla received power, signal to interference plus noise ratio, download/upload throughput, and latency-related measurements derived from the radio access network. Using the CatBoost classifier on a proprietary dataset of 115 264 subscribers constructed with a balanced 50/50 class distribution to prevent majority-class bias, the model achieves an F1 score of 81.0% and an area under the receiver operating characteristic curve of 0.884, demonstrating strong discriminative performance. A central contribution is the introduction of a systematic SHapley Additive exPlanations (SHAP)-based attribution method that distinguishes network-driven churn from other churn types. For each churn prediction, positive SHAP values are extracted and categorized as network-related or non-network-related; instances where network features contribute the dominant share of the total positive SHAP attribution are labeled as network-driven churn. The framework is validated on both the proprietary Turkcell dataset and three publicly available telecom churn benchmarks, demonstrating generalizability. The proposed approach bridges predictive modeling and realworld network optimization, offering telecom operators a scalable and interpretable tool for targeted customer retention.