Graph Neural Networks Map How Campus Waterfront Spaces Move and Hold Their Visitors

A study of two university campuses in Guangzhou, China, has used Wi-Fi probe sensors and graph neural networks to model how pedestrian spaces along campus waterfronts are actually used, distinguishing for the first time in this setting between people passing through and people lingering. The research, published in Frontiers in Public Health, analyzed 42 spatial units along the water bodies of the South China University of Technology’s Wushan Campus and Guangzhou International Campus, and found that the two aspects of spatial vitality behave very differently: movement was far easier to predict from the surrounding built environment than staying behavior.

The researchers, led by Zhiye Yu and Ying Bao of the South China University of Technology with collaborators at KTH Royal Institute of Technology, Chongqing University, and the University of Queensland, framed the problem around a tension that most vitality studies blur. Waterfront pedestrian paths serve simultaneously as circulation routes, which the authors call “path” attributes, and as destinations for recreation and social interaction, their “place” attributes. Treating vitality as a single aggregate number, they argue, obscures the fact that these two behaviors may respond to different environmental features.

To separate them, the team deployed six TZ-4007 Wi-Fi probes across the campuses between December 11 and 14, 2025 at Wushan and December 18 to 21, 2025 at the Guangzhou International Campus, observing each spatial unit during two afternoon periods on both a weekday and a weekend day under comparable dry, mild conditions of roughly 18 degrees Celsius. The probes recorded anonymized device identifiers, timestamps, and signal strength values every five seconds. Records with weak signals, at or below −85 dBm, were filtered out to reduce detections likely originating outside the target unit.

From these raw detections the authors constructed two device-based indicators. Flow counts devices whose observed duration within a unit fell below a unit-specific traversal threshold, calculated from the unit’s length divided by a reference walking speed with a tolerance factor of 1.5. Dwell counts devices persisting beyond that threshold, indicating staying-oriented activity rather than transient passage. The two indicators are mutually exclusive by construction and jointly partition all detected devices. The authors are careful to note these are relative device-based patterns of spatial use, not exact pedestrian counts, since Wi-Fi probes detect phones rather than people and cannot capture devices with Wi-Fi disabled.

The built environment of each unit was described by nine features across four dimensions: circulation conditions such as average walkway width, environmental perception including motor interference and shading level, visual landscape captured through semantic segmentation of field photographs to derive green view ratio, visual openness, and water view ratio, and activity support measured as stay facility density and surrounding activity influence. Bridges were flagged as a binary variable. Photographs were taken at 1.5 meters height with a 28 mm focal length at roughly 15 evenly spaced points per unit to approximate pedestrian eye-level perception.

The analytical core of the study is a graph representation in which each spatial unit is a node and physical adjacency between neighboring units forms the edges, allowing a graph neural network to incorporate the conditions of adjacent spaces into each node’s prediction. Three architectures were tested: Graph Convolutional Network, GraphSAGE, and Graph Attention Network, each tuned through a reproducible random search over hidden dimensions, layer counts, dropout rates, and learning rates, evaluated across five predefined training-validation splits within each campus with three fixed random seeds per split.

The results revealed a striking asymmetry. For Flow, the GCN achieved the best overall validation RMSE of 0.5387 with an R-squared of 0.48. For Dwell, GraphSAGE performed best but managed only an R-squared of 0.04, meaning the model explained essentially none of the variance in staying behavior. Against conventional benchmarks, a mean baseline, ordinary least-squares regression, and a spatial lag model evaluated on identical splits, the selected graph models reduced RMSE by 36.2 percent for Flow and 14.0 percent for Dwell. A geographically weighted regression benchmark was excluded because each campus contained only 20 to 22 spatial units, too few for stable local coefficient estimates.

A permutation-based feature perturbation analysis then probed which environmental features the trained models were most sensitive to, aggregating results across seeds, splits, campuses, and a Flow-Dwell quadrant classification of the units. Water view ratio ranked highest for Flow at both campuses, a shared pattern despite the two sites differing markedly: Wushan has a mature, discontinuous water system with a direct pedestrian-water interface, established in 1934, while the 2022 international campus has a continuous water system with a buffered interface. Motor interference and stay facility density ranked highly at Wushan, where mixed pedestrian-vehicle use persists and waterside seating is plentiful, but dropped sharply at the newer campus, which maintains stricter separation and has no waterside seating at all, producing zero sensitivity for that feature.

The authors stress that these sensitivity values capture the magnitude of model response to perturbing a feature, not the direction of its effect or any causal influence on behavior. Quadrant-level rankings for contexts containing only two or three units, such as the low-flow, high-dwell quadrants, are treated as descriptive and exploratory. Validation performance may also carry some optimistic bias because the same validation subsets served both early stopping and model selection. The observation window was short and seasonal, and the study did not measure physical activity, mental health, or restorative experience directly.

What the framework offers, the authors argue, is a translation layer between model output and design judgment. By combining a unit’s behavioral quadrant, its feature-sensitivity ranking, and its current conditions standardized against the campus median, the method distinguishes features worth maintaining, candidates for improvement, and cases where highly sensitive features are already well provided. In illustrative units, reducing pedestrian-vehicle conflict emerged as a candidate priority at a movement-oriented Wushan space with high measured motor interference, while a low-activity unit at the international campus with good water views, greenery, and openness warranted maintenance and further site-level examination rather than physical upgrades. The authors position all such interventions as design hypotheses requiring onsite validation, with the public health contribution remaining at the environmental-support level rather than as evidence of direct health effects.

Subject of Research: Graph neural network modeling of pedestrian movement and staying activity in campus waterfront spaces using Wi-Fi probe data

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

Article References: Yu, Z., Bao, Y., Li, S., Liu, X., Pei, Y., & He, B.-J. (2026). Graph neural network modeling of pedestrian space vitality in campus waterfront space: a Flow–Dwell perspective for design translation. Frontiers in Public Health, 14, Article 1963816. https://doi.org/10.3389/fpubh.2026.1963816

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.1963816

Keywords: graph neural network, campus waterfront, pedestrian vitality, Wi-Fi probe, built environment, spatial analysis, urban design, public health, Flow-Dwell, feature sensitivity, spatial units, machine learning