Abstract
The growing demand for resilient and safe bridge infrastructure necessitates advanced approaches to construction risk assessment capable of addressing high dimensionality, uncertainty, and interpretability challenges. This study proposes a multi-level explainable artificial intelligence (XAI) framework that integrates domain-specific feature learning, ensemble prediction, and probabilistic causal reasoning for comprehensive bridge construction risk assessment. At Level 1, supervised autoencoders extract latent structural representation features from high-dimensional construction data, preserving the distinct characteristics of structural, environmental, managerial, resource, and safety domains. Level 2 employs a stacked ensemble model combining an Explainable Boosting Machine (EBM) with a calibrated HistGradientBoostingClassifier to achieve accurate and interpretable risk predictions, supported by SHAP-based explanations. At Level 3, a factor graph model captures causal dependencies and quantifies uncertainty, enabling probabilistic inference and scenario-based risk evaluation under varying operational and environmental conditions. Results demonstrate robust predictive performance, transparent domain-level risk attribution, and well-calibrated uncertainty estimates. The proposed framework advances data-driven and explainable decision support for bridge construction risk mitigation, effectively addressing the challenges posed by complex, heterogeneous datasets arising from BIM and sensor-enabled construction environments.