Unveiling the Power of Hybrid Balancing Techniques and Ensemble Stacked and Blended Classifiers for Enhanced Churn Prediction

For businesses, customer retention is crucial as it is more cost-effective than acquiring new customers. Identifying potential customer churn early allows for the development of effective retention strategies. With advancements in technology and data storage, machine learning has become a popular approach for predicting customer churn. To counteract data imbalance, researchers have utilized minority oversampling methods, particularly the Synthetic Minority Over-sampling Technique (SMOTE). Innovations in this area include hybrid techniques like SMOTE Tomek-Links and SMOTE ENN, which have shown effectiveness in data resampling. Traditional classifiers like Logistic Regression, Naïve Bayes, Support Vector Machine, and K-Nearest Neighbors have been surpassed in performance by ensemble classifiers such as XgBoost, LightGBM, and CatBoost. Yet, there is limited research on the combination of SMOTE hybrid techniques with these advanced ensemble classifiers for churn prediction. This study aims to contribute to the field by integrating hybrid balancing techniques with ensemble classifiers and introducing new stacked and blended models. The findings reveal that a stacked model incorporating SMOTE ENN achieved impressive results: 96.46% accuracy, 97% F1 score, and 97.40% PR-AUC. This was closely followed by the CatBoost-SMOTE ENN model, which scored 95.32% in accuracy, 96% F1 score, and 96.50% PR-AUC. In contrast, ADASYN and standard SMOTE techniques did not significantly affect model performance.