Lightweight framework for IFC element classification using multi-view and heterogeneous graph with decision-level fusion

Ensuring semantic consistency in BIM interoperability is a critical challenge, as the geometric information of IFC elements is often preserved during IFC-based data exchange, whereas entity labels may become inconsistent or lost. However, existing methods are computationally expensive and fail to fully exploit the implicit representations embedded in IFC data. To address these challenges, this paper proposes IFCGeoNet, a lightweight framework for IFC element classification that integrates explicit shape and implicit representation features. The framework constructs IFC schema-level and file-level heterogeneous graphs, applies graph contrastive and transfer learning for implicit feature extraction, performs explicit feature extraction through multi-view shape encoding, and integrates both via a semantic group-aware decision-level fusion mechanism. Experiments on the enriched IFCNetCore dataset show that IFCGeoNet achieves an F1-score of 0.9128 with low computational cost. This study provides a robust and efficient approach for IFC element classification, laying the groundwork for consistent and accurate BIM data utilization.