Machine Learning Based Customer Churn Prediction

Abstract

Although substantial work has been devoted to improving the performance of customer churn prediction models, research on their interpretability and understanding of feature importance remains limited. However, for businesses to develop cost-effective customer retention strategies, it is vital to recognize which customer attributes are most influential in churn prediction. This paper explores the role of feature importance in forecasting churn within the telecommunications industry, utilizing a dataset containing nearly 40 customer attributes and employing three transparent, tree-based machine learning models. Specifically, we applied Random Forest, Decision Tree, and Extra Trees Classifier algorithms, leveraging their internal feature ranking mechanisms to evaluate the impact of individual features. We then compared the importance scores generated by these models to identify the most significant predictors overall. All three models achieved high classification performance, though the Decision Tree model demonstrated slightly lower accuracy than the others. Interestingly, the top-ranked features showed substantial overlap across the models, confirming their suitability for this analysis. Among the algorithms, Random Forest and Extra Trees Classifier, with their more intricate structures, provided superior results. Our findings highlight that variables such as customer tenure, service usage, and referral counts are the key determinants of churn, which aligns logically with customer engagement and loyalty patterns.