An Explainable Machine Learning Framework for Telecom Customer Churn Prediction Using RFM Clustering and LightGBM USING R

Abstract

Customer churn prediction is a very sensitive activity in the telecommunication sector since it is cheaper to retain customers than to acquire them. This paper will suggest a hybrid analytic and machine learning platform that combines Recency-Frequency-Monetary (RFM)-based customer segmentation, K-means clustering, and Light Gradient Boosting Machine (LightGBM) model with SHAP-based explainable AI (XAI) to improve prediction accuracy and model interpretability. LightGBM algorithm performs both for the original data and the clustering data set. The experiment result shows that the cluster data set improves the prediction accuracy by 0.8068 and F1 by 0.8760, and the original data accuracy is 0.7765 and F1 is 0.8395. And the AUC value cluster data and original data are 0.8484 and 8582, which means the strong predictive capability of the prediction. SHAP analysis will be employed in order to show the most significant features influencing customer churn. According to the findings, the most important predictors of different categories of customers are the type of contract, tenure, monthly charges, and total charges. Specifically, the month-to-month contracts and increased monthly fees are closely linked with a greater degree of churn risk. The importance of the features also differ within clusters, which highlight the customer behavior distinctions The findings provide that the customer segmentation RFM model with ML algorithm and explainable AI techniques can provide deep insight into customer behaviour and an important churn prediction model strategy