Advancing Federated Learning for Unified Urban Spatio-Temporal Predictions (2026-2029)

This project aims to address pressing challenges in urban spatio-temporal predictions, such as data sparsity and noise, privacy concerns, data heterogeneity, and limited generalisability. It expects to generate transformative innovations in federated learning for spatio-temporal foundation models. Key contributions include a model transmission-free federated learning architecture featuring data condensation to generate synthetic yet informative knowledge carriers, a physics-guided spatio-temporal data enhancement framework, and robust defenses against potential attacks. These outcomes will broadly benefit the transportation, environment, and public safety sectors, enabling smarter, safer, more efficient, and sustainable urban communities.