A comparative analysis of graph neural networks for estimating free-flow travel times in urban road networks

Abstract

Urban transit planning in densely populated areas is an urban design challenge that constantly requires effective solutions. Urban road networks can be represented by graphs, helping decision-makers visualize and further incorporate information learned from the cities’ structure. Graph Neural Networks (GNNs) are artificial neural networks designed to learn from graphs’ information and structure. However, important implicit assumptions of general-purpose GNNs do not hold for urban road network graphs and learning with this type of graph may require specialized GNNs. This study investigates the performance of both general-purpose and specialized GNNs to estimate free-flow travel times in urban road networks of twelve cities worldwide. Emphasis is given to Relational Fusion Networks (RFNs), specialized GNNs with competitive results in the literature. Edge regression results of the selected GNNs are obtained for comparison. They show that RFNs perform better and highlight the role of different fusion functions on the regression performance. Additionally, this study explores RFNs’ capacity to extrapolate learned knowledge to different cities. The investigation also searches through cities’ structural characteristics, aiming to comprehend their influence on RFN’s learning mechanisms and its capacity to extrapolate knowledge.