<p>Accurate early stage construction cost prediction is essential for project budgeting, financial planning, and strategic decision making. However, reliable estimation remains difficult because early project information is limited, heterogeneous, and strongly interdependent across design conditions, project characteristics, and market context. Conventional statistical and machine learning methods usually represent these factors as unordered tabular inputs and therefore have limited ability to capture their hierarchical organization and high order dependencies. To address this issue, this study proposes an adaptive hypergraph learning framework for actual construction cost prediction. First, a prior hierarchical hypergraph is constructed to encode cost factors and their domain-informed relationships across design information, project characteristics, and economic indicators. Second, an adaptive hyperedge refinement mechanism is introduced to modulate relation strength according to project context, enabling the learning framework to preserve expert knowledge while learning project-specific structural variation. Third, a confidence-calibrated aggregation mechanism is designed to estimate latent factor reliability and calibrate feature aggregation during hypergraph learning. To avoid interpreting this internal reliability mechanism as full probabilistic uncertainty quantification, an additional prediction interval analysis is introduced to evaluate interval coverage and width under heteroscedastic cost ranges. Experiments on a real-world dataset of 50 school projects using 5-fold cross-validation show that the proposed method achieves a MAPE of <InlineEquation ID="IEq1"><EquationSource Format="TEX">\(10.94 \pm 1.02\%\)</EquationSource></InlineEquation>, an MAE of <InlineEquation ID="IEq2"><EquationSource Format="TEX">\(15.12 \pm 2.18\)</EquationSource></InlineEquation> million HKD, an RMSE of <InlineEquation ID="IEq3"><EquationSource Format="TEX">\(19.46 \pm 2.41\)</EquationSource></InlineEquation> million HKD, and an <InlineEquation ID="IEq4"><EquationSource Format="TEX">\(R^2\)</EquationSource></InlineEquation> of <InlineEquation ID="IEq5"><EquationSource Format="TEX">\(0.57 \pm 0.06\)</EquationSource></InlineEquation>. Compared with strong tree-based ensemble baselines, including Random Forest, XGBoost, and LightGBM, the proposed learning framework achieves moderate but consistent gains. In addition, when the proposed training strategy is applied to existing neural and graph-based predictors, it improves both point prediction accuracy and prediction interval quality compared with alternative training strategies. The framework also provides interpretable insights into both factor importance and learned relation usage. These results indicate that integrating hierarchical factor organization, adaptive structural learning, and confidence-calibrated representation learning offers an effective direction for early stage construction cost prediction under small-sample conditions.</p>

Adaptive hypergraph learning improves early stage construction cost prediction

期刊论文
出版模式：

开放获取
发表日期：

2026年7月20日

查看全文

Zhao Lilong

摘要

Accurate early stage construction cost prediction is essential for project budgeting, financial planning, and strategic decision making. However, reliable estimation remains difficult because early project information is limited, heterogeneous, and strongly interdependent across design conditions, project characteristics, and market context. Conventional statistical and machine learning methods usually represent these factors as unordered tabular inputs and therefore have limited ability to capture their hierarchical organization and high order dependencies. To address this issue, this study proposes an adaptive hypergraph learning framework for actual construction cost prediction. First, a prior hierarchical hypergraph is constructed to encode cost factors and their domain-informed relationships across design information, project characteristics, and economic indicators. Second, an adaptive hyperedge refinement mechanism is introduced to modulate relation strength according to project context, enabling the learning framework to preserve expert knowledge while learning project-specific structural variation. Third, a confidence-calibrated aggregation mechanism is designed to estimate latent factor reliability and calibrate feature aggregation during hypergraph learning. To avoid interpreting this internal reliability mechanism as full probabilistic uncertainty quantification, an additional prediction interval analysis is introduced to evaluate interval coverage and width under heteroscedastic cost ranges. Experiments on a real-world dataset of 50 school projects using 5-fold cross-validation show that the proposed method achieves a MAPE of \(10.94 \pm 1.02\%\), an MAE of \(15.12 \pm 2.18\) million HKD, an RMSE of \(19.46 \pm 2.41\) million HKD, and an \(R^2\) of \(0.57 \pm 0.06\). Compared with strong tree-based ensemble baselines, including Random Forest, XGBoost, and LightGBM, the proposed learning framework achieves moderate but consistent gains. In addition, when the proposed training strategy is applied to existing neural and graph-based predictors, it improves both point prediction accuracy and prediction interval quality compared with alternative training strategies. The framework also provides interpretable insights into both factor importance and learned relation usage. These results indicate that integrating hierarchical factor organization, adaptive structural learning, and confidence-calibrated representation learning offers an effective direction for early stage construction cost prediction under small-sample conditions.