Submitted:
17 November 2026
Posted:
19 November 2026
You are already at the latest version
Abstract
The increasing adoption of artificial intelligence in construction management has created opportunities for data-driven schedule-risk prediction. However, the reliability and generalizability of machine-learning models across heterogeneous infrastructure projects remain insufficiently understood, particularly when project-level variability and temporal changes are not adequately considered. This study investigates the predictive performance, probability calibration, and transferability of machine-learning models for identifying activity-duration growth in infrastructure construction projects. Using an extensive dataset comprising 148,087 construction activities across 271 infrastructure projects, the study evaluates logistic regression and gradient-boosting models under conventional and project-balanced learning strategies. A rigorous evaluation framework incorporates five-fold project-held-out validation, project-cluster bootstrap uncertainty estimation, probability calibration, temporal validation, explainable artificial intelligence, and risk-based activity prioritization. Results indicate that standard gradient boosting achieves an area under the receiver operating characteristic curve (AUROC) of 0.633, while project-balanced learning achieves comparable discrimination (AUROC = 0.632) and improved project-level probability accuracy. However, temporal validation reveals substantial performance deterioration for certain model configurations, highlighting the influence of distributional shifts and project heterogeneity. Furthermore, model effectiveness in prioritizing high-risk construction activities varies across project sizes. These findings demonstrate the importance of project-aware evaluation, temporal robustness assessment, and uncertainty quantification when developing intelligent schedule-risk prediction frameworks. The study contributes methodological and practical insights for evaluating data-driven construction risk management systems and supporting more reliable infrastructure project decision-making.
Keywords:
construction scheduling
; infrastructure projects
; activity-duration growth
; temporal drift
; project heterogeneity
; probability calibration
; explainable machine learning
; schedule risk
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.