Transfer Learning in Spatiotemporal Graph Neural Networks for Traffic Forecasting: Taxonomy, Analysis, and Future Directions
Published in SSRN preprint, 2025
Accurate traffic forecasting is vital for many intelligent transportation system applications. Graph neural network (GNN)-based spatiotemporal models achieve remarkable success by capturing the spatial dependencies among sensors and the temporal dynamics of traffic flow. However, these models rely heavily on large volumes of high-quality sensor data, limiting their applicability in data-scarce regions. Transfer learning offers a promising solution by enabling the reuse of knowledge learned from data-rich cities or road networks. In recent years, numerous transfer learning frameworks have been proposed to enhance the generalisation, transferability, and data efficiency of GNN-based traffic forecasting models.
This survey systematically reviews and categorises 42 methods into a unified taxonomy through meta-analysis. We compile a comprehensive list of datasets and open-source implementations, highlighting key challenges and future research directions. This is the first comprehensive review of transfer learning in GNN-based spatiotemporal traffic forecasting. We also provide an actively maintained GitHub repository that tracks the latest papers, codes, and datasets in this domain.
Recommended citation: Lone, Soban Nasir, Mohamed Abouelela, Taeyoung Yu, Jiwon Kim, and Constantinos Antoniou. "Transfer Learning in Spatiotemporal Graph Neural Networks for Traffic Forecasting: Taxonomy, Analysis, and Future Directions." SSRN preprint (2025). https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5933134
Download Paper