arXiv:2609.20026v1 Announce Type: cross Abstract: Urban traffic forecasting often relies on information distributed across stakeholders who may be unable to share raw data due to privacy or commercial constraints, motivating federated spatial-temporal approaches. In such federated settings, each client observes traffic over a distinct sensor subgraph with its own spatial topology and temporal dyn
FedeRICo (arXiv:2609.20026) is a federated spatial-temporal forecasting framework that addresses traffic flow prediction across heterogeneous client subgraphs without sharing raw data. It utilizes a dual-branch architecture that separates shared forecasting structures from client-specific residual corrections, coordinated via gradient-level alignment rather than direct parameter averaging.
To recover cross-client spatial dependencies severed by network partitioning, FedeRICo employs a boundary-aware residual communication protocol that exchanges transient, event-informative signals between physically adjacent clients. This approach preserves privacy by avoiding the transmission of raw traffic observations or persistent demand patterns.
The framework demonstrates superior performance on four real-world benchmarks (METR-LA, PEMS-BAY, PEMS03, PEMS07), reducing Mean Absolute Error by 18.8% compared to the FedAvg baseline and narrowing the performance gap to centralized models to within 3.2%–7.3%. It effectively handles regional heterogeneity and dynamic congestion propagation while maintaining competitive computational costs.
Problem and setting. FedeRICo addresses traffic flow prediction in a federated, multi-stakeholder environment where raw sensor data cannot be centralized because of privacy, ownership, or commercial constraints. Each participating client observes traffic over its own sensor subgraph, meaning the clients are heterogeneous not only in traffic volume and temporal patterns, but also in spatial topology: road networks, sensor placement, neighborhood structure, and regional dynamics differ across regions. This makes standard federated learning assumptions problematic, since clients are not merely non-IID in data distribution; their underlying graph structures and spatiotemporal regimes are also distinct.
Key idea and contribution. The paper introduces a federated framework that uses region-influenced coupling to exploit spatial relationships across clients without sharing raw observations. Rather than treating each client as an isolated local model, FedeRICo couples regional patterns so that a client can benefit from related or neighboring spatial-temporal information during training. The intended effect is to align local spatiotemporal representations, improve the transferability of learned features across regions, and make federated updates more robust to topological heterogeneity and non-identical traffic dynamics. This is a design choice aimed at the specific difficulty of graph-structured traffic data, where spatial context is as important as local time-series behavior.
Why it matters. Urban traffic forecasting is increasingly a distributed problem: municipalities, transit authorities, navigation providers, and infrastructure operators all hold complementary but sensitive data. A method that can learn from these data silos while preserving privacy is practically important, especially when central data sharing is infeasible. FedeRICo is relevant because it moves beyond generic federated averaging and instead targets the structural non-IID nature of regional traffic sensing. For federated learning over graph-based urban data, it offers a useful direction: incorporate inter-client spatial influence directly into the learning process, enabling better generalization across regions without requiring a centralized traffic data repository.