Generating sparse origin-destination flows on shared mobility networks using probabilistic graph neural networks

Shared mobility services, such as bike sharing, have gained immense popularity and emerged as an integral part of sustainable urban mobility solutions. The planning of such systems requires forecasting the potential origin- destination (OD) flows between mobility sites (e.g., bike sharing stations) within the proposed network. Existing methods primarily focus on mobility flows between large regions, and do not generalize well to detailed planning applications due to the high spatial resolution required, with increased uncertainty and data sparsity. This study proposes a zero-inflated negative binomial graph neural network (ZINB-GNN) to generate sparse OD flows while capturing complex spatial dependencies. To reflect sparsity, OD flows are modeled as following ZINB distributions parameterized via feed-forward networks. To capture spatial dependencies, localized graphs are constructed to represent proximity between OD pairs, with spatial features encoded using GNNs. ZINB-GNN is validated through a case study of the bike sharing system in New York City. The results verify its prowess in both prediction accuracy and uncertainty quantification under real-world network expansion scenarios. We also demonstrate its interpretability by revealing important factors affecting OD flows. These findings can directly inform the planning of bike sharing systems, and the methodology may be adapted for other shared mobility systems.