Model Agnostic Meta-Learning for traffic assignment on disrupted networks

Abstract

Accurate traffic flow estimation from origin-destination matrices is challenging when historical and future conditions diverge significantly — a common occurrence during network disruptions such as road closures, flooding, or public demonstrations. These are precisely the scenarios where reliable predictions matter most, yet standard machine-learning models struggle to quickly adapt to unseen graph structures with different demand patterns. This paper presents a meta-learning framework combined with a graph convolutional neural network (GNN) to address this limitation. Rather than requiring exhaustive training datasets covering all possible disruption scenarios, the architecture is trained to rapidly adapt to new network topologies and OD matrices with minimal data. Applied to simultaneous changes in road network structure and travel demand, the proposed model achieves an R² of approximately 0.85 on unseen network closures, significantly easing the burden on transport practitioners designing training datasets.