| id |
caadria2025_479 |
| authors |
Cui, Ziqi and Lou, Shangyu |
| year |
2025 |
| title |
How do the Spatial Layout of Street Elements and Geometric Features Affect Human Perception: A prediction model and explainable method based on Graph Neural Network |
| source |
Dagmar Reinhardt, Christiane M. Herr, Anastasia Globa, Jielin Chen, Taro ?Narahara, Nicolas Rogeau (eds.), ARCHITECTURAL INFORMATICS - Proceedings of the 30th CAADRIA Conference, Tokyo, 22-29 March 2025, Volume 4, pp. 459–468 |
| summary |
Understanding citizen feedback on the built environment is essential to promote inclusive urban development. However, existing urban perception models built based on machine learning techniques have limitations such as poor interpretability and insufficient feature embedding. This study explores whether incorporating the spatial relationships and geometric features of individual street elements can enhance the accuracy and interpretability of perception prediction models. We used quantified spatial and geometric features of urban street elements as input for a graph neural network (GNN) model, a deep learning approach designed to process unstructured data. Compared to traditional methods, the GNN model achieved approximately a 6.3% on average improvement in accuracy across six perception dimensions, highlighting the importance of integrating spatial and geometric features into predictive modelling. Additionally, the study trained an explainer to analyze prediction results, enabling precise identification of critical subgraphs within images, which can determine the key edges, node combinations, and node features that significantly influence perception scores. Compared to previous approaches, this method provides more granular explanation results, offering valuable insights for urban planners to design environments that better align with public perceptions. |
| keywords |
Street spatial relationship, Urban perception, Graph neural network, Explainability in prediction model, Computer vision |
| series |
CAADRIA |
| full text |
 file.pdf (1,890,638 bytes) |
| last changed |
2025/03/21 12:10 |
|
