Multi-class imbalanced data problem solving for predicting internet service cancellation

Abstract

Churn is common problem across various industry, especially in the telecommunication sector. In order to reduce customer churn rates and complete effectively, many telecom companies are recognizing the need to improve customer experience and service. Therefore, understanding the reasons behind customer churn is essential. In this research, we apply machine learning multi-class classification technique for predicting customer churn to the internet service provider industry. We propose a two-fold approach to address the issue: 1) Handling imbalanced data at the dataset level by utilizing the SMOTE technique to synthetically generate additional data, and 2) Addressing imbalanced data at the learning process level by adjusting the data weights using the XGBoost technique. The resulting XGBoost based classifier is able to predict accurately the majority classes and as well as the minority classes. The combination of SMOTE and XGBoost techniques demonstrates effectiveness in solving imbalanced problem in the case of multi-class datasets.