FedeRICo: Federated Region-Influenced Coupling for Traffic Flow Prediction

Summary

FedeRICo is a federated learning framework for traffic flow prediction in which each client holds a disjoint sensor subgraph of a road network. The authors argue that parameter-averaging FL dilutes client-specific representations and that graph partitioning severs cross-boundary spatial dependencies. The framework combines two mechanisms: (1) a dual-branch GWNet-based architecture, where a shared branch is coordinated across clients via gradient-direction alignment (adapted from FedIGL's bi-gradient principle) while a private residual branch, supervised on the shared branch's residuals, is deliberately diversified across clients; and (2) client-to-client boundary messaging, where sender-side encoders transmit learned transient residual signals—periodic structure suppressed via temporal-embedding gating—to the top-k nearest nodes of physically adjacent clients, without sharing raw observations. On METR-LA, PEMS-BAY, PEMS03, and PEMS07(M), FedeRICo reports the best federated MAE/RMSE/MAPE on every dataset, cutting METR-LA MAE by 18.8% versus FedAvg (3.93 to 3.19) and 9.1% versus the strongest baseline FedDis (3.51 to 3.19), within 3.2-7.3% of centralized GWNet. Ablations attribute the gains to both mechanisms; results are stable across 4-20 clients, and per-round cost (21.3s plus 40.7s) undercuts FedDis (148.4s). A reconstruction audit finds boundary messages add no inversion signal over metadata. Main limitations: single-run results without seeds, error bars, or significance tests despite small margins (e.g., 0.12 MAE on PEMS-BAY); no comparison with gradient-correction FL methods (SCAFFOLD, FedDyn, FedIGL); and privacy supported only by one limited attacker audit with no formal guarantees.