Synthetic data as a contribution and a reward in multi-agent urban traffic forecasting

Abstract

Collaborative urban traffic forecasting is an emerging area of research. Given the growing number of stakeholders and owners of traffic data, the collaboration between them is extremely important for successful citywide traffic forecasting and management. A stable collaboration between data owners requires a fair and efficient scheme for valuing contributions and distributing rewards. This study proposes the use of synthetic data sets both as a contribution and a reward in collaborative forecasting. This novel approach has several significant advantages over other collaboration schemes like federated model training: it is model-agnostic, privacy-preserving, flexible and enables market mechanisms for contribution valuation. The empirical part of the study illustrates the applicability of the proposed approach, using the real-world split of a road network between organisations. The experimental setup utilises an attention-based spatial-temporal graph neural network as a forecasting model and dynamic factors as an approach to synthetic data generation. The obtained results are analysed and found promising, and the advantages of the proposed approach are articulated.