Spatio-temporal traffic series prediction is essential in intelligent transportation systems, benefiting various applications such as route planning, vehicle dispatching, and congestion prediction. To tackle privacy leakage aroused by centralized forecasting methods, Federated Learning (FL), a privacy-preserving approach for decentralized model training into disjointed federated clients, has garnered widespread adoption in numerous traffic prediction endeavors. However, existing FL-based approaches ignore spatio-temporal heterogeneity among federated clients, including spatial feature skew, temporal coverage skew, and data quality skew. This makes them inapplicable and unsuitable to real-world scenarios and exhibits subpar prediction performance. To this end, we perform the first study of heterogeneous-aware traffic prediction in the federated environment, proposing a unified and effective framework named Fed4TP. It offers general federated capability for various centralized forecast models, supporting flow, speed, and occupancy prediction tasks. To address spatial feature heterogeneity, Fed4TP develops multi-dimensional personalized federated learning with positive samples contrastive learning for clustering to achieve personalized aggregation and global sharing across diverse clients. To overcome temporal coverage heterogeneity, Fed4TP designs a time window-based federated training mechanism, sequentially training client models and learning missed traffic information with varying time coverage. To tackle data quality heterogeneity, Fed4TP introduces a dual-driven method, i.e., global detection and local denoising, to improve client data quality. Extensive experiments on 4 real-life datasets verify the effectiveness and scalability superiority of Fed4TP in various federated-based traffic prediction tasks, compared with 24 well-known and state-of-the-art baselines. The source code and data of this work are available at https://github.com/ZJU-DAILY/Fed4TP.

Spatio-temporal prediction plays a crucial role in intelligent transportation, weather forecasting, and urban planning. While integrating multi-modal data has shown potential for enhancing prediction accuracy, key challenges persist: (i) inadequate fusion of multi-modal information, (ii) confounding factors that obscure causal relations, and (iii) high computational complexity of prediction models. To address these challenges, we propose E$^2$-CSTP, an Effective and Efficient Causal multi-modal Spatio-Temporal Prediction framework. E$^2$-CSTP leverages cross-modal attention and gating mechanisms to effectively integrate multi-modal data. Building on this, we design a dual-branch causal inference approach: the primary branch focuses on spatio-temporal prediction, while the auxiliary branch mitigates bias by modeling additional modalities and applying causal interventions to uncover true causal dependencies. To improve model efficiency, we integrate GCN with the Mamba architecture for accelerated spatio-temporal encoding. Extensive experiments on 4 real-world datasets show that E$^2$-CSTP significantly outperforms 9 state-of-the-art methods, achieving up to 9.66% improvements in accuracy as well as 17.37%-56.11% reductions in computational overhead.