Bike-sharing ridership prediction for network expansion using graph neural networks

Ridership prediction in station-based bike-sharing services improves station planning, fleet management, and network design. Ridership inflow and outflow prediction at the station level has received significant attention through trip production and attraction models. However, station-to-station ridership has been studied less, despite its widespread applications in use cases such as bike-lane planning or fleet electrification. This study introduces a Graph Neural Network (GNN) to model station-to-station ridership using a customized Graph Sample and Aggregate framework to generate node embeddings and minimize the weighted Mean Squared Error for peak periods. The model incorporates the characteristics of the network, sociodemographic features, and station properties. We present the case study of Bikeshare Toronto to train and test the GNN model and benchmark it against other standard prediction methods. We show that the GNN outperforms linear regression, spatial regression, XGBoost, and artificial neural networks due to its ability to capture the impact of the network structure on ridership patterns. We incorporate the GNN model in five design scenarios focusing on urban core connectivity, suburban access, transit integration, equitable accessibility, and tourist hubs. Each scenario is strategically developed to prioritize and address unique urban challenges. To enhance the model’s application in real-world planning, we embedded the model in a web-based tool for the Cities of Vancouver and Toronto, allowing for further “what-if” scenario analysis in bike-sharing network planning.