An Ensemble of Advanced Machine Learning Models for Telecom Customer Churn Forecasting

Abstract

In the rapidly evolving telecommunications sector, maintaining profitability and growth depends on customer retention. This study provides an in-depth analysis of customer churn prediction with the aim of determining the important factors impacting customer attrition and developing a practical predictive model. In order to make churn prediction more accurate and consistent, the authors of this paper suggest a hybrid ensemble model (LGBM + RF) that mixes LightGBM with random forests. The model makes use of both approaches to improve the accuracy and generalizability of its predictions. Testing on the IBM Telco Customer Churn dataset shows that the proposed model performs better than both traditional ML and DL models. Compared to its forerunners, LGBM + RF achieves better results with a lower AUC of 0.998%, a higher PRE of 98.85%, a recall (REC) of 99.12%, and an F1-score (F1) of 98.56%. These models were previously known as CatBoost, XGBoost, Random Forest, KNN, and Deep-BP-ANN. Ensemble learning is successful in capturing complicated patterns of consumer behavior, according to the findings. This method offers practical information that telecom companies can apply to develop targeted customer retention mechanisms and enhance customer loyalty, making it a useful and valid tool for churn forecasting programs.