Graph Representation Learning: Embedding Multimodality BIM Models into Graphs

Abstract: Existing AI research in BIM often overlooks the multimodal nature of BIM data. This study proposes a graph representation learning approach to embed multimodal design data into high-dimensional vectors for supporting learning. Specifically, large text embedding models are utilized to encode object attributes, while geometries are embedded by a point cloud-based approach. The individual embeddings are concatenated and processed by GraphSAGE to leverage graph topology. An object classification experiment on a self-constructed graph dataset achieved 97.7% accuracy and an F1 score of 0.97 using attributes and graph topology. Future work will refine geometry embeddings and explore broader BIM applications.