HFedBike: A Hybrid Federated Learning System for Urban Bike-Sharing Demand Forecasting

Publication Type:

Conference Paper

Source:

27th IEEE International Conference on Mobile Data Management (MDM 2026), IEEE, Athens, Greece (2026)

Keywords:

Bike-Sharing Systems, Edge computing, Federated Learning, Smart Mobility

Abstract:

This study presents a hybrid federated learning system for urban bike-demand forecasting that balances distributed training with controlled coordination. The approach introduces a structured communication scheme in which local model updates are exchanged within small groups and selectively propagated across the network, thereby avoiding the need for a central aggregation point. The system is evaluated on large-scale, real-world mobility and weather data in a realistic multi-node deployment with heterogeneous edge devices. Results show consistent improvements in predictive accuracy over non-collaborative and temporally adaptive approaches, while maintaining stable operation across all evaluation cycles. In addition, the proposed design significantly reduces communication overhead, demonstrating its effectiveness as a practical and scalable solution for real-world smart mobility scenarios.