FedTP: A Federated Learning Framework for Traffic Prediction

Accurate traffic forecasting is essential for intelligent transportation systems, yet most existing approaches rely on centralized training paradigms that raise privacy concerns and often neglect fairness across heterogeneous nodes. Federated learning (FL) provides a privacy-preserving alternative by enabling collaborative training without raw data sharing. However, standard FL algorithms, such as FedAvg and FedProx, primarily optimize global accuracy while exacerbating inter-client disparities, leading to biased service quality in critical regions. To address these challenges, we propose FedTP, a federated learning framework that integrates gradient conflict elimination into the aggregation process. By projecting conflicting client gradients onto a conflict-free subspace, FedTP harmonizes local updates, thereby improving fairness across clients without compromising overall predictive accuracy. Extensive experiments on real-world traffic datasets demonstrate that FedTP achieves competitive accuracy compared to centralized baselines while significantly reducing performance disparities among clients.