A Feature Interaction Network for Customer Churn Prediction

Customer churn prediction is an active research topic for the data mining community and business managers in this rapidly growing society. The ability to detect churn customers precisely is something that every company would wish to achieve. With the great success of DNNs, several churn prediction models based on DNNs are proposed in recent years. However, traditional DNNs cannot learn high-order feature interactions and deal with one-hot vectors well. In this paper, we proposed a feature interaction network (FIN), which aims to enhance the inherent relations of discrete features and learn high-order feature interactions. This network contains two modules: an entity embedding network and a factorization machine network with several sliding windows. From the experiments, it is observed that our proposed model has a better predictive performance than several state-of-the-art models on 4 public datasets.