Exploring architectural floor plan appropriateness in context of Bangladesh leveraging graph neural networks in spatial context

Abstract

This paper investigates the use of Graph Neural Networks (GNNs) for classifying architectural floor plans and establishing the applicability of international floor plans with respect to Bangladeshi architectural standards. Flooring plan data is mainly derived from Chinese residential designs, which are converted into graph-based representations where rooms represent the nodes, and the connections through doors form the edges. Node features are prepared that include room area, centroid coordinates of the room, and room type, while door connections form unweighted edges. Three GNN models—GCN, GraphSAGE, and GAT are tested to evaluate their effectiveness in this binary classification task. GraphSAGE yielded the best performance among all the three GNN models tested, showing 87.09% test accuracy and an AUC-ROC score of 0.9512, with good generalization on unseen data. This work illustrates how GNNs can capture spatial relations from architectural data to enable scalable solutions for cross-cultural design evaluation and urban planning. It contributes to the increasingly important intersection of AI and Architecture by going beyond image-based traditional approaches and introducing a framework that automatically assesses the appropriateness of architectural designs concerning different cultural contexts.