Towards a large scale practical churn model for prepaid mobile markets

Abstract

Communication Service Providers (CSPs) are increasingly focusing on effective churn management strategies due to fierce competition, price wars and high subscriber acquisition cost. Though there are many existing studies utilizing data mining techniques for building churn prediction models, addressing aspects such as large data volumes and high dimensionality of subscriber data is challenging in practice. In this paper, we present a large scale churn management system, utilizing dimensionality reduction and a decision-tree ensemble learning. We consider subscriber behavior trends as well as aggregated key performance indicators (KPIs) as features and use a combination of feature selection strategies for dimensionality reduction. We report favorable results from evaluating our model on a real-world dataset consisting of approximately 5 million active subscribers from a popular Asian CSP. The model output is presented as an actionable lift chart which will help marketers in deciding the target subscriber base for retention campaigns. We also present the practical aspects involved in productionizing our model using the Apache Oozie workflow engine.