Collaborative User Mobility Prediction in Distributed Edge Computing Framework

User attachment forecasting with Deep Learning (DL) is an effective tool for proactive mobility management in dense Beyond 5G (B5G)/6G deployments with reduced cell sizes. However, collecting user data in a central cloud to facilitate DL model learning causes extensive overhead and privacy concerns. Distributed edge cloud-based federated learning solves these issues, but it faces challenges in handling out-of-distribution data from decentralized edges at the network periphery, and the model biases due to data heterogeneity. This paper addresses these limitations by proposing a fully distributed Collaborative User Mobility Prediction (CUMP) framework that mitigates the out-of-distribution data issue through collaboration among initial layers of DL models in edges that are selected using inter-edge mobility rates. The remaining part of each model only trains on local data, preserving biases towards their respective edges. This enhances the generalization, robustness, and predictive performance of the DL models. Results show that CUMP outperforms conventional global learning and state-of-the-art distributed personalized federated learning and cyclic incremental institutional learning by 63%, 12%, and 10% in predicting the next Point of Attachment (PoA) of a user and by 70%, 22%, and 28% in predicting user dwell time in current PoA, respectively. Thus, CUMP improves prediction performance while reducing network and storage overheads while preserving privacy.