Graph-structured gravity model enhances transferable pedestrian flow prediction

Abstract Understanding and modeling pedestrian flow patterns within cities is essential for promoting public health, social equity, and environmental sustainability. However, pedestrian mobility data are often sparse, imbalanced, or unreliable, particularly in low-resource contexts, forcing planners to transfer models developed in one city to others, often with poor outcomes. This challenge calls for models that are generalizable and robust to pedestrian mobility data limitations. In this study, we introduce a hybrid graph-based method for constructing behaviorally meaningful destination choice sets, coupled with a Transferable Graph Laplacian Gravity Network (TL-GNet), a learning-enhanced gravity model that integrates graph topology and spatial regularization to estimate pedestrian flows. Evaluated using household travel survey data from four cities of Melbourne and Brisbane (Australia), and Seattle and Chicago (USA), our model consistently outperforms traditional and deep gravity models, particularly in predicting low-volume pedestrian flows. TL-GNet demonstrates strong spatial transferability, achieving significantly higher predictive accuracy across cities without retraining. The results highlight the potential of graph-informed models to support pedestrian planning in data-scarce urban environments and advance transferable approaches to modeling active and sustainable mobility.