Node-Level Federated Learning with Adaptive Personalized Aggregation for Spatio-Temporal Traffic Prediction

Accurate and real-time traffic flow prediction is crucial for Intelligent Transportation Systems. Recent advances in federated learning and spatio-temporal modeling have improved accuracy and privacy protection. However, existing methods often rely on global topology for spatial features, neglecting topology protection, and typically train a generic global model without considering local personalized features, limiting prediction performance. This paper proposes ST-PFLA (Spatio-Temporal Traffic Flow Prediction via Personalized Federated Learning with Adaptive Aggregation), a framework designed for node-level scenarios where clients only have information about their respective connections, to improve prediction accuracy and training efficiency while safeguarding topology privacy. In ST-PFLA, clients conduct prediction by combining spatial and temporal features extracted by the attention mechanism and local datasets respectively. The method aggregates only encoders across clients, retaining decoders locally for personalization. Each client performs an additional local training round to generate a guide model, which is used to inform the calculation of aggregation weights. Experimental results on two public datasets show that ST-PFLA can significantly enhance prediction accuracy while safeguarding topology privacy at lower training costs.