Churn Prediction in Telecom Industry using Machine Learning Ensembles with Class Balancing

Abstract

Telecommunication service providers are going through a very competitive and challenging time to retain existing customers by offering new and attractive services (e.g., unlimited local and international calls, high-speed internet, new phones). It is therefore imperative to analyse and predict customer churn behaviour more accurately. One of the major challenges to analyse churn data and build better prediction model is the imbalance nature of the data. Customer behaviour for churn and non-churn scenarios may contain resembling features. Using a single classifier or simple oversampling method to handle data imbalance often struggles to identify the minority (churn) class data. To overcome the issue, we introduce a model that uses sophisticated oversampling technique in conjunction with ensemble methods, namely Random Forest, Gradient Boost, Extreme Gradient Boost, and AdaBoost. The hyperparameters of the baseline ensemble methods and the oversampling methods were tuned in several ways to investigate their impact on prediction performances. Using a widely used publicly available customer churn dataset, prediction performance of the proposed model was evaluated in term of various metrics, namely, accuracy, precision, recall, F-1 score, AUC under ROC curve. Our model outperformed the existing models and significantly reduced both false positive and false negative prediction.