Customer churn prediction using a novel meta-classifier: an investigation on transaction, Telecommunication and customer churn datasets

Abstract

With the advancement of electronic service platforms, customers exhibit various purchasing behaviors. Given the extensive array of options and minimal exit barriers, customer migration from one digital service to another has become a common challenge for businesses. Customer churn prediction (CCP) emerges as a crucial marketing strategy aimed at estimating the likelihood of customer abandonment. In this paper, we aim to predict customer churn intentions using a novel robust meta-classifier. We utilized three distinct datasets: transaction, telecommunication, and customer churn datasets. Employing Decision Tree, Random Forest, XGBoost, AdaBoost, and Extra Trees as the five base supervised classifiers on these three datasets, we conducted cross-validation and evaluation setups separately. Additionally, we employed permutation and SelectKBest feature selection to rank the most practical features for achieving the highest accuracy. Furthermore, we utilized BayesSearchCV and GridSearchCV to discover, optimize, and tune the hyperparameters. Subsequently, we applied the refined classifiers in a funnel of a new meta-classifier for each dataset individually. The experimental results indicate that our proposed meta-classifier demonstrates superior accuracy compared to conventional classifiers and even stacking ensemble methods. The predictive outcomes serve as a valuable tool for businesses in identifying potential churners and taking proactive measures to retain customers, thereby enhancing customer retention rates and ensuring business sustainability.