Inductive Representation Learning on Feature Rich Complex Networks for Churn Prediction in Telco

Abstract

In the mobile telecommunication industry, call networks have been used with great success to predict customer churn. These social networks are complex and rich in features, because the telecommunications operators have a lot of information about their customers. In this paper we leverage a novel framework called GraphSAGE for inductive representation learning on networks with the goal of predicting customer churn. The technique has an advantage over previously proposed representation learning techniques because it leverages node features in the learning process. It also features a supervised learning process, which can be used to predict churn directly, as well as an unsupervised variant which produces an embedding. We study how the number of node features impacts the predictive performance of churn models as well as the benefit of a complete learning process, compared to an embedding with supervised machine learning techniques. Finally, we compare the performance of GraphSAGE to that of standard local models.