MSTGN: Multimodal Spatio-Temporal Graph Neural Network for Telecommunication Customer Churn Prediction

This paper presents a model called MSTGN a new framework for customer churn prediction in telecommunications. Keeping existing customers is much cheaper than finding ones making churn prediction very important. Old methods struggle with the types of data and complex relationships between users. MSTGN solves these problems by combining data types into one model. We have tested MSTGN on a telecom dataset and a public dataset. It performed better than existing methods. Experiments on a provincial telecom dataset and the publicly available Orange Telecom dataset show that MSTGN outperforms existing methods in discrimination, calibration, and business-oriented metrics. Planned extensions include incorporating a wider range of data sources, experimenting with newer neural architectures, and conducting deeper economic analyses to improve the model’s practical value.