Ozkan, Mustafa Can; (2025) Towards traffic foundation models: graph-centric learning under network structural change and data sparsity. Doctoral thesis (Ph.D), UCL (University College London).

Text Ozkan_10215617_Thesis.pdf Access restricted to UCL open access staff until 1 November 2027. Download (13MB)

Abstract

This thesis presents a graph-centric framework for traffic prediction in dynamic road networks, focusing on scenarios where the network structure changes—such as road closures or newly developed networks—and historical traffic data is limited or no longer applicable. While traditional data-driven models perform well in static network structures with abundant data, they often fail when the topology evolves, spatiotemporal data coverage is sparse, and past patterns cannot be transferred to the new structure. This work addresses these challenges by enabling reliable forecasting under both structural changes and data scarcity. The proposed framework integrates three complementary strategies: each tailored to different levels of data availability. First, when a road network data with moderate observation exists, a subgraph augmentation method is used to train spatiotemporal graph neural networks (STGNNs), leveraging historical zero-flow events and anomalous patterns in neighbouring segments to adapt to structural changes. Second, in data-scarce environments (e.g., networks with only few days of historical data before the structure change), the framework deploys reinforcement learning (RL) agents trained on expert routing behaviours from data-rich cities. This knowledge is transferred to data-scarce road networks, where the agents adaptively respond to network changes and inject feedback into graph neural networks models via message-passing algorithms, enhancing predictions despite limited observations. Finally, for entirely new networks with no prior data, which implies a complete structural change (e.g. newly developed transport networks for a new city), a transformer-based traffic foundation model is introduced. Pretrained on data from over 25 global cities, the model employs multi-channel transfer learning to align spatiotemporal patterns across diverse urban topologies, enabling generalisation to new environments. Empirical evaluations across diverse urban settings demonstrate the framework’s versatility: STGNNs with subgraph augmentation excel with moderate data, RL-based models maintain accuracy under severe data constraints, and the foundation model effectively predicts traffic trends in new networks. Together, these contributions represent a major step toward adaptive, generalisable, and data-efficient traffic prediction. By uniting graph neural networks, reinforcement learning, and foundation model architectures, this work offers practical and scalable solutions for next-generation intelligent transportation systems across diverse urban environments.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as CSVJSONXML

Archive Staff Only

View Item