Machine learning approaches for customer churn prediction in a telecom operator using real-world data

Customer churn prediction is a critical issue in the telecommunications industry, as customer attrition can directly impact revenue and business sustainability. This study aims to develop and evaluate machine learning models for customer churn prediction using real-world data from a telecom operator. The dataset, collected over a twelve-month period, integrates Call Detail Records (CDR) and Customer Relationship Management (CRM) data. It consists of ۳,۱۵۰ customer records with behavioral and usage-based features. The class distribution is imbalanced, with churners representing approximately ۱۶% of the dataset. Several machine learning algorithms, including Logistic Regression, k-Nearest Neighbors, Random Forest, and Support Vector Machine (SVM), were developed and compared. Class imbalance was addressed using oversampling techniques, and hyperparameter tuning was applied to improve model performance. The results indicate that the tuned SVM model combined with oversampling achieved the best performance, with an F۱-score of ۰.۸۹ and an accuracy of ۰.۹۶. The findings demonstrate that even classical machine learning models can serve as effective tools for identifying customers at risk of churn and supporting data-driven retention strategies. The proposed framework provides practical value for telecom operators in designing targeted customer retention campaigns. This study contributes empirical evidence from an underrepresented telecom market.