Network analysis and graph neural network (GNN)-based link prediction of construction hazards

Hazard recognition is critical for construction safety, especially for accident prevention. Traditional methods often fail to capture the dynamic and interdependent nature of construction hazards. To address this issue, this paper proposes a network-based framework that conceptualizes construction hazards as dynamic interactions between objects with hazardous attributes. A link prediction model using Graph Neural Networks (GNNs) is integrated in this framework to automatically explore latent interactions between hazard objects that are ignored by the existing dataset. By analyzing 4470 construction accident reports, this paper constructed a hazard network and revealed key structural properties, including hazard object centrality, cliques, and communities. The experimental results of link prediction showed that the GNN-based model demonstrated superior performance compared to traditional methods, with 81 % of GNN-predicted links validated by actual construction accident cases. This framework provides a practical solution for intelligent hazard recognition and proactive risk management in the construction industry. • An object-attribute model was proposed for intelligent hazard recognition. • Hazard clique was introduced to describe closely interacting hazard objects. • GNNs were employed to predict latent links between hazard objects. • 81 % GNN-predicted links were validated by actual construction accident cases.