MoE Enhanced Federated Learning for Spatiotemporal Prediction

Researchers propose MoE-FedTP, a federated learning framework using Mixture-of-Experts networks to improve traffic prediction across cities while preserving privacy. The system enables data-rich cities to share knowledge with data-scarce regions by dynamically fusing expert networks tailored to different urban environments, achieving superior accuracy without centralized data collection.

MoE-FedTP addresses a critical bottleneck in intelligent transportation systems: traffic prediction accuracy suffers when cities lack sufficient sensor data. The framework represents a meaningful advance in federated learning architecture, moving beyond generic cross-city knowledge transfer to handle spatiotemporal heterogeneity—the reality that traffic patterns vary significantly across urban geographies. By decomposing expertise into specialized expert networks paired with a dynamic gating mechanism, the system avoids one-size-fits-all model limitations while maintaining the privacy protections inherent to federated approaches.

The technical innovation centers on partial parameter sharing among experts derived from different source cities, enabling fine-grained personalization without requiring massive model downloads or centralized aggregation. This design choice reduces computational overhead compared to full federated models while preserving the ability to capture diverse urban dynamics.

For smart city infrastructure planners and transportation authorities, this research offers practical implications. Data-scarce cities can leverage predictive capabilities from well-instrumented regions without exposing sensitive traffic or infrastructure data to third parties—a critical consideration for municipalities balancing innovation with security. The experimental validation across four real-world datasets suggests the approach generalizes beyond controlled environments.

Looking forward, the intersection of federated learning and urban computing remains underexplored. Success here could catalyze adoption in other spatiotemporal domains—energy grids, pollution monitoring, emergency response—where privacy and heterogeneity pose similar challenges. The lightweight MoE architecture's resource efficiency may prove particularly valuable for resource-constrained municipalities.

• →MoE-FedTP enables privacy-preserving knowledge transfer between data-rich and data-scarce cities for traffic prediction.
• →Dynamic gating mechanism fuses expert networks to capture diverse urban traffic patterns without centralized data collection.
• →Partial parameter sharing reduces computational overhead while maintaining fine-grained personalization across heterogeneous environments.
• →Framework consistently outperforms state-of-the-art federated and cross-city prediction baselines on real-world datasets.
• →Lightweight architecture design positions the approach for scalability to other spatiotemporal prediction domains.