Churn Prediction Using Dynamic RFM-Augmented Node2vec

摘要

Current studies on churn prediction in telco apply network analytics to analyze and featurize call graphs. While the suggested approaches demonstrate a lot of creativity when it comes to deriving new features from the underlying networks, they also exhibit at least one of the following problems: they either do not account properly for dynamic aspects of call networks or they do not exploit the full potential of joint interaction and structural features and additionally, they usually address these in a non-systematic manner which involves hand-engineering of features. In this study, we propose a novel approach in which we address each of the identified issues. In a nutshell, first, we propose slicing a monthly call graph to capture dynamic changes in calling patterns. Second, we devise network designs which conjoin interaction and structural information. Third, we adapt and apply the node2vec method to learn node representations in a more automated way and to avoid the need for feature handcrafting.