Search-guided regression ensembles for accurate, interpretable, and uncertainty-aware construction cost estimation

Abstract

Accurate and interpretable construction cost estimation remains a major challenge due to complex nonlinear dependencies, heterogeneous data distributions, and inherent uncertainty in project parameters. To overcome these limitations, this study proposes the Search-Guided Regression Ensemble (SGRE), a novel hybrid framework that unifies dynamic learner selection, uncertainty quantification through prediction intervals, and explainable model interpretation using SHAP (SHapley Additive exPlanations). The framework integrates six complementary base learners, namely K-Nearest Neighbors (KNN), Decision Tree (DT), Natural Gradient Boosting (NGB), Support Vector Regression (SVR), Multilayer Perceptron (MLP), and Bayesian Ridge Regression (BR), and introduces two ensemble variants: Forward Search-Guided Regression Ensemble (F-SGRE) and Backward Elimination Search-Guided Regression Ensemble (BE-SGRE). These search-guided strategies adaptively construct parsimonious ensembles that enhance predictive accuracy, stability, and reliability while maintaining interpretability. Comprehensive evaluations demonstrate that SGRE not only achieves superior prediction performance compared to traditional single and fixed ensemble models but also produces well-calibrated prediction intervals that provide reliable uncertainty bounds around model predictions. Furthermore, SHAP analysis reveals consistent feature importance across models, identifying "Formwork" as the dominant cost driver, followed by Tributary Area and Concrete, thereby reinforcing the framework's transparency and practical trustworthiness. Overall, the proposed SGRE framework establishes a robust, explainable, and uncertainty-aware paradigm for construction cost estimation, supporting resilient infrastructure, sustainable transportation, and resource efficiency in modern construction management.