Graph neural network modeling of pedestrian space vitality in campus waterfront space: a Flow–Dwell perspective for design translation

In campus landscapes, waterfront pedestrian spaces support both movement and staying activities and may contribute to activity- and health-supportive environments. This study examined the associations between built-environment features and waterfront pedestrian-space vitality across 42 spatial units at two campuses of the South China University of Technology (20 at Wushan Campus and 22 at Guangzhou International Campus). Wi-Fi probe data were collected during weekday and weekend observation periods on December 11–14, 2025 (Wushan Campus) and December 18–21, 2025 (Guangzhou International Campus) to derive two device-based vitality proxies: Flow for transient activity and Dwell for staying-oriented activity, distinguished using unit-specific traversal-time thresholds. A Flow–Dwell quadrant framework was used to classify the spatial contexts, and graph neural network node regression was employed to incorporate spatial adjacency relationships. Feature perturbation was further used to quantify the model-prediction sensitivity at the node level and to support design interpretation. The results showed that (1) Flow and Dwell exhibited distinct spatial distributions and environmental sensitivity patterns; (2) Flow was better predicted by graph neural networks than Dwell (best overall R 2 = 0.48, versus 0.04 for Dwell), whereas Dwell showed a substantially weaker model fit; and (3) the selected GNNs outperformed the best-performing conventional baselines for both targets, reducing the RMSE by 36.2% for Flow and 14.0% for Dwell. This study provides a spatially explicit framework for interpreting campus waterfront vitality and translating model-derived sensitivity patterns into design priorities, offering evidence for planning activity-supportive campus environments, and contributing to public-health-oriented spatial design.