Federated Spatio-Temporal Learning for Cross-Modal Transport Predictions

Within the DigiK project "Monitoring nachhaltiger Tourismusdestinationen", we present a prediction framework where operators of different transport modes collaborate to forecast traffic conditions while preserving privacy. Privacy is a challenge particularly pronounced in tourism destinations, where transport systems typically involve multiple independent operators managing diverse and sensitive mobility data. Modern transportation systems are shaped by multiple interdependent modes of transport, and exploiting their interdependencies can improve forecasting accuracy. In tourism contexts specifically, this multi-operator landscape creates two compounding challenges: data-privacy constraints prevent the sharing of raw data among providers, and variation in data distributions and volumes across operators introduces substantial heterogeneity. To address these challenges, we propose FedSTD, a federated learning framework for Spatial and Temporal Decoupled learning. FedSTD enables transport providers to collaboratively train predictive models without sharing raw data, preserving privacy for sensitive spatial information. Our approach handles cross-modal heterogeneous distributions by decoupling shared temporal patterns from mode-specific spatial features, supporting effective knowledge transfer while maintaining model personalization. We further introduce a new real-world cross-modal dataset with extended temporal coverage, enabling longer prediction horizons. Experiments on two real-world datasets demonstrate that FedSTD improves long-term flow forecasting and outperforms existing baselines. The method offers tourism destinations actionable insights into the benefits of privacy-preserving collaborative predictions, directly supporting the sustainability monitoring goals of our DigiK project