At the IEEE International Conference on Edge Computing and Communications - IEEE EDGE 2025 held in Helsinki, Finland, from July 7 to 12, 2025, the paper “Semi-decentralized Training of Spatio-Temporal Graph Neural Networks for Traffic Prediction” authored by Ivan Kralj, Lodovico Giaretta, Gordan Ježić, Ivana Podnar Žarko, and Šarūnas Girdzijauskas was presented. The paper was accepted after an international review and presented on-site by Ivan Kralj.
The work is the result of a collaboration between IoTLab at FER and RISE Research Institutes of Sweden. The paper explores semi-decentralized training approaches for Spatio-Temporal Graph Neural Networks (ST-GNNs) in the context of smart mobility systems, where large-scale, geographically distributed traffic sensors continuously generate high-frequency data. To address scalability and reliability limitations of centralized learning, the authors propose a cloudlet-based framework in which sensors are grouped by proximity and local ST-GNN models are trained cooperatively using different distributed learning paradigms.