The integration of Graph Neural Networks (GNNs) with Building Information Modeling (BIM) unveils transformative potential for the semantic enrichment of BIM data, offering significant advancements to the Architecture, Engineering, and Construction (AEC) industry. This paper addresses the critical issue of spatial recognition and classification in BIM through the proposed nodeenhanced, self-supervised graph neural network model, Node-Enhanced GraphBidirectional Encoder Representation from Transformer (NE-Graph-BERT), which incorporates edge features. This model represents each space as a node, utilizing spatial relational edges as inputs and is trained on an independently developed knowledge base consisting of 14 space types and 4 relational features across three major categories of architectural space layouts, enabling automatic identification of spatial features.
| Authors | Chen, Yian and Jiang, Huixian |
|---|---|
| Year | 2025 |
| Venue | Engineering Applications of Artificial Intelligence |
| DOI | 10.1016/j.engappai.2025.110365 |
| Source Database | EBSCO (Academic Search Ultimate) |
| Bridge-to-GNN | Category A |
| GNN Architecture | GNN (heuristic — verify) |
| Graph Encoding | Not yet extracted from PDF |
| AEC Task | Semantic Enrichment (inferred from title) |
| Cohort | Mature Applications (2024–2026) |
| Primary Metric | Accuracy (heuristic — verify) |
|---|---|
| Primary Value | 97.05% (heuristic — verify) |
| GNN architectures detected | GNN (general) |
|---|---|
| Reported metrics + values | F1 Score: 96.75% · Precision: 97.08% |
| Quality (heuristic, 6-flag) | 0/6 · rigor: Low Code/Data: —Benchmark: —Baseline: —CV/Split: —Ablation: —Reproducible: — |
Part of the GML/GNN in AEC Systematic Review (PRISMA 2020) — 112 papers, 2020–2026