Inferring building functions from a weighted graph isomorphic network based on the building-POI graph

Building function is the practical use of a structure, playing a crucial role in urban planning and risk management. However, prevailing studies utilizing graph neural networks (GNNs) often regard building function inference as a node classification task, which fails to solve the problem of model performance bias toward residential buildings caused by sample imbalance. To address this limitation, we construct the building-POI graph by regarding POIs distributed on a building as graph nodes, incorporate the vertical spatial location of POIs and the contribution of different types of POIs to building function, and convert the building function inference into a graph classification task. Then, we propose a weighted graph isomorphic network (WGIN) by improving the aggregators and readout of GIN for distinguishing buildings. To validate the efficacy of the proposed method, we compare it with existing approaches. The results demonstrate that the graph construction method and the proposed classification model outperform alternative methods in building function inference, achieving the highest accuracy of 82.27%, an F1 value of 74.57%, and a kappa coefficient of 0.716. Notably, our approach alleviates the challenge of accuracy differences across various building types and holds immense significance for supplementing and validating existing data and optimizing urban structures.