Research on the causal mechanism of prefabricated building accidents: a comprehensive framework integrating association rule mining and graph neural network

Purpose This study investigates the causal mechanisms underlying prefabricated building construction safety accidents (PBCSA) and aims to provide a systematic framework for identifying key risk factors and propagation pathways in complex construction environments. Design/methodology/approach Based on 267 accident cases, 5 accident types and 101 causal factors were extracted. A weighted causal network was constructed using association rule mining (ARM), followed by structural analysis of network properties. A Multivariate Feature Graph Convolutional Network (MF-GCN) was developed to identify critical causal nodes, and a depth-first search algorithm was applied to extract representative causal pathways. Findings The constructed causal network exhibits clear small-world characteristics and a heavy-tailed degree distribution, indicating a structured rather than random risk formation mechanism. MF-GCN significantly outperforms traditional centrality measures in identifying key causal factors, with root and direct causes accounting for approximately 75% of the most influential nodes. Risk propagation is primarily driven by a small number of recurrent causal chains, where factors such as improper lifting, inadequate risk assessment and insufficient safety distance act as key bridging nodes across accident types. Originality/value This study integrates the “2–4” accident classification model with ARM and graph learning, providing a unified framework for causal structure discovery in construction safety systems. The proposed MF-GCN enhances the identification of influential risk nodes in complex accident networks, while the extracted causal chains reveal cross-type risk propagation mechanisms, offering actionable insights for proactive safety management.