Client churn prediction with call log analysis

Client churn prediction is a classic business problem of retaining customers. Recently, machine learning algorithms have been applied to predict client churn and have shown promising performance comparing to traditional methods. Despite of its success, existing machine learning approach mainly focus on structured data such as demographic and transactional data, while unstructured data, such as emails and phone calls, have been largely overlooked. In this work, we propose to improve existing churn prediction models by analysing customer characteristics and behaviours from unstructured data, particularly, audio calls. To be specific, we developed a text mining model combined with gradient boosting tree to predict client churn. We collected and conducted extensive experiments on 900 thousand audio calls from 200 thousand customers, and experimental results show that our approach can significantly improve the previous model by exploiting the additional unstructured data. Copyright © 2018 Springer International Publishing AG, part of Springer Nature.