Implementation of Machine Learning Models for Predicting Internet Service Provider Customer Churn

Abstract

The telecommunications industry faces an extremely high level of competition, where the phenomenon of customer churn presents a significant challenge due to its impact on revenue decline and increased costs associated with acquiring new customers. This study aims to develop a churn prediction model using the Decision Tree algorithm and implement it in a web-based application to support customer retention strategies. The CRISP-DM methodology is employed, covering Business Understanding, Data Understanding, Data Preparation, Modeling, Evaluation, and Deployment. Experimental results show that the Decision Tree algorithm demonstrates strong performance in identifying non-churn customers, with a precision of 0.82, a recall of 0.91, and an F1-score of 0.86. However, its performance on the churn class remains limited, with a precision of 0.63, a recall of 0.44, and an F1-score of 0.52, highlighting the importance of addressing imbalanced data distribution to preserve existing data. The model underwent Learning Curve and Validation Curve analysis. The Learning Curve indicates a relatively stable model with a small gap, suggesting good generalization. The Validation Curve reveals that optimal performance is achieved at a moderate tree depth, avoiding the risk of overfitting at greater depths. Nevertheless, the main advantage of the Decision Tree is its interpretability, which highlights significant factors such as contract type, subscription duration, and additional services. The integration of the model into a web-based application also provides practical benefits through rapid churn risk monitoring, supporting the company’s strategic decision-making.