Machine Learning Approaches to Advance Cycling

Accurate bicycle volume estimation is vital for urban planning but is hindered by extreme data sparsity, with over 99% of road segments lacking counts. This thesis develops Graph Neural Network (GNN)-based methods to address this challenge across four studies. It begins with a survey of GNNs in transportation, then evaluates their performance under sparsity using Melbourne’s bicycling network. To improve results, it introduces BikeVAE-GNN, combining GNNs with Variational Autoencoders, and INSPIRE-GNN, an RL-boosted sensor placement strategy. Together, these approaches overcome sparsity limitations, enabling more accurate bicycle volume prediction and infrastructure planning for healthier, data-driven urban mobility systems.