FedSTD: Federated Learning with Spatio-Temporal Decoupling for Cross-Modal Transport Prediction

Modern transportation systems are shaped by multiple interdependent modes of transport. Understanding their interactions is essential for effective management and planning. Cross-modal flow and demand predictions can improve forecasting accuracy for individual modes by exploiting these interdependencies. However, data-privacy constraints limit data sharing among transport providers, and the variation in data distributions and volumes within and across silos introduces substantial heterogeneity. To address these challenges, we propose FedSTD, a federated learning framework decoupling spatial and temporal learning. FedSTD enables transport providers to collaboratively train predictive models without sharing raw data, preserving privacy, especially for sensitive spatial information. Our approach addresses cross-modal heterogeneous data distributions by decoupling shared temporal patterns from mode-specific spatial features. This design supports effective knowledge transfer while maintaining model personalization. In addition, we present a new real-world cross-modal dataset with extended temporal coverage, enabling longer prediction horizons than existing benchmarks. Extensive experiments on two real-world cross-modal datasets demonstrate that FedSTD improves long-term flow forecasting accuracy and outperforms existing baselines. Beyond prediction performance, we analyze the trade-offs between model design and federated collaboration strategies, offering practical insights for transport providers regarding the benefits of collaborative predictions under privacy constraints.