Personalized Federated Learning with Neighbor Aggregation for Traffic Flow Prediction

As an important part of Intelligent Transportation System (ITS), traffic flow prediction has attracted more and more attention from academia and industry. However, since traffic flow data are distributed among different organizations, traditional centralized learning methods are no longer applicable, thus federated learning has begun to play an important role in traffic flow prediction tasks. In this paper, we design an improved traffic flow prediction model based on federated learning. At the same time, we consider the interaction of topology and feature space between traffic data owned by different organizations, and propose a method named Federated Neighbor Aggregation (FedNe). The server update the model parameters with both the physical connectivity and similarity relationships between the organizations, and finally generate a personalized model for each organization. Experimental results on real-world dataset show that our proposed method has the lowest average MAE and average RMSE. Finally, based on our method, when new clients participate, they can also quickly obtain their personalized models.