Construction cost prediction for small-to-medium datasets using Bayesian optimization and multi-task learning

Purpose Accurate early-stage construction cost prediction is critically important for decision-making. However, traditional single-task models are inherently limited in their ability to simultaneously capture the interdependencies between total cost and its subcosts (e.g. labor, material, and machinery). This study proposes a multitask learning framework, the Construction Cost Bi-Stacked Network with Bayesian Optimization (CCBSN-BO), to simultaneously predict total, labor, material, and machinery costs, capturing correlations among itemized subcosts. Design/methodology/approach The model is trained on multi-source preliminary design data from 82 building projects (22 residential, 60 public) in Hefei, China (2019–2024). It integrates XGBoost and LightGBM in a dual-stack ensemble architecture, optimized via Bayesian hyperparameter search. The proposed model achieved a mean absolute percentage error (MAPE) of 18.06% for total cost prediction, significantly outperforming benchmark models including Random Forest (32.99%), XGBoost (36.25%), and LASSO (37.84%), while attaining a high coefficient of determination (R2) of 0.92. Comparative verification revealed a clear performance hierarchy, with the framework excelling in accuracy and stability. Further, SHAP analysis provides interpretability of cost transmission mechanisms. Findings SHAP analysis reveals quantifiable cost transmission mechanisms. Floor area affects labor costs about 2.15 times more strongly than material costs, supporting standardized design to reduce labor-intensive spending. The foundation type shows contrasting cost profiles in this dataset; Pile foundations are associated with higher labor and machinery/equipment costs but lower material costs, whereas shallow foundations (e.g. strip or raft) exhibit the opposite pattern; Even under safety constraints, component-level cost optimization remains important. Irregular geometries are associated with concurrent increases in labor and machinery costs, which can be mitigated through prefabrication. Special facilities (e.g. intelligent systems) are linked to higher upfront costs but may reduce reliance on labor and heavy equipment in later stages, potentially lowering lifecycle expenditures. Practical implications The CCBSN-BO framework provides a decision-support tool for early-stage cost management. It enables value-driven design optimization by allowing designers to quantify the cost implications of design alternatives before detailed specifications are available and to balance functional requirements with cost efficiency. In doing so, the framework bridges predictive analytics and practical design optimization, providing a basis for refined cost management in the critical early phase of a project. Originality/value This research offers an accurate and interpretable tool for early-stage cost estimation. It quantifies interdependencies among detailed cost components, supporting refined cost management and value engineering. The CCBSN-BO framework improves multi-cost prediction and informs design strategies—such as standardization to reduce labor-cost sensitivity and prefabrication to manage cost interactions from complex geometries.