Telecom Customer Churn Prediction with Explainable Machine Learning

In the fast-paced era of digital transformation, the telecoms industry is grappling with an increasing issue of customer churn, which endangers stable revenue and operational efficiency. This paper therefore proposes an interpretable machine learning framework that aims to identify customers who are likely to churn more accurately, while also enhancing the interpretability of the model. The performance of the model is evaluated using the publicly available Iranian churn dataset by means of data preprocessing (duplicate removal, outlier detection, feature selection and class imbalance mitigation), training of seven machine learning models tuned with hyper-parameters and selection of five key performance indicators. Experimental results demonstrate that the CatBoost model achieves excellent performance, with an accuracy of 0.9554, precision of 0.9010, recall of 0.8072, F1 score of 0.8503 and AUC-ROC of 0.9823 — surpassing the performance of the other six models. Additionally, this study explored the stacking model, which ultimately proved to be inferior to the CatBoost model. Finally, the optimal model was interpreted using the SHAP technique. The analysis identified three key characteristics affecting customer churn: seconds of use (negative correlation), call failures (positive correlation) and frequency of SMS (negative correlation). This led to the derivation of an operational strategy for customer retention in the telecommunications industry.