Abstract: With the continuous development of building intelligence, building information modeling (BIM) integrates key information such as building geometry, component attributes and spatial relationships. However, when conducting structural analysis and intelligent design, the lack of effective representation of architectural space characteristics makes it difficult to intuitively present the adjacent relationships between components and the overall layout of the building. The adjacency relationship of building components (building graph) is an important manifestation of spatial characteristics, but existing research has not been able to effectively extract this relationship. In view of the above, this paper proposes a building graph construction method based on feature-augmented graph neural networks (GNNs). The research process consists of building graph construction, GNN design, and training. This method effectively overcomes the problem of difficulty in obtaining adjacent relationships when acquiring building graphs. The experimental results show that this method achieves an accuracy rate of 97.23% in obtaining the adjacent relationships of building components, providing strong support for the efficient generation of data from BIM to building graphs and the implementation of downstream tasks.