Churn prediction methods based on mutual customer interdependence

Most widespread churn prediction models assume customer independence, ignoring the well-documented propagation of churn influence in a customer network. Although this customer interdependence can be modelled by social network analysis and shallow node representation learning algorithms, these methods are too inefficient and impractical for use in large corporate systems. An efficient solution that incorporates both customer features and interconnections is a graph neural network; however, its potential for churn prediction is still understudied. This paper provides an overview of the existing approaches and outlines the properties of graph neural networks that make them a promising end-to-end solution.