Comparative analysis demonstrates tabular models outperform graph architectures for accident risk screening in public construction projects, highlighting gradient boosting efficacy.
This study compares tabular machine learning models and attribute-based heterogeneous graph models for construction site accident risk screening using 214,293 public construction project records with a 0.91% accident rate. Six models were compared under five repeated stratified splits and a common hyperparameter search, with PR-AUC as the primary metric. XGBoost achieved the highest PR-AUC of 0.268 ± 0.022, ranked first in all five splits, and captured 35.3% of accidents within the top 1% of sites. The graph-based models exceeded the accident-rate baseline but did not outperform the tabular models, indicating limited benefits without direct relational identifiers. Under missing-signal control, the estimated dependence on missingness varied from −2.1% to −23.2% across splits, indicating that single-split estimates are unreliable.
No takes yet. Share an insight, caveat, or question.
Moon et al. (2026) studied this question.
Synapse has enriched 5 related papers. Consider them for comparative context: