Construction safety predictions with multi-head attention graph and sparse accident networks

The reliability of risk assessment is crucial for designing effective construction safety management strategies. Construction safety prediction using machine learning models is still suboptimal, considering representativeness, interpretability, and efficiency challenges. This paper enhances the representativeness, interpretability, and accuracy of construction safety prediction models using a multi-head graph attention network (GAT) and a novel sparse construction safety network. Through its accommodation of connectivity information between accident records, the proposed approach enhances the interpretability and representativeness of non-graph-based machine learning models. In addition, it improves the message aggregation of the embedded accident information and adaptability to hyperparameter variations of the state-of-the-art graph convolutional network. The evaluation of multi-head GAT on three sparse construction safety networks achieved accuracies of 86.2%, 87.1%, and 86.9%, which implies the benefit of the self-attention mechanism in learning the importance of connected accident records. This substantiates the reliability of the proposed approach for integration within a risk assessment procedure, the outcome of which is instrumental in designing effective management strategies.