A hybrid stacked regression model for actual construction cost: Incorporating MPNet sentence embeddings, meta-learning, and Bayesian hyperparameters optimization

초록

Accurate prediction of actual construction costs remains a critical challenge in the construction industry, as most existing machine learning approaches rely primarily on structured numerical and categorical inputs while neglecting the rich contextual information embedded in unstructured project descriptions. This study proposes a hybrid stacked regression framework that integrates MPNet-based sentence embeddings with structured features for construction cost prediction. Specifically, the all-mpnet-base-v2 model is employed to extract contextual representations from free-text project descriptions, which are then concatenated with numerical and categorical inputs and used within a stacking ensemble of regression models. The model was developed on a dataset of 472 Korean government housing projects and benchmarked against different models using TF-IDF and Word2Vec representations, as well as a structured-only baseline. The results demonstrate that the proposed model achieves an R2 of 0.9251 and an RMSE of approximately $5.24 million, corresponding to a 38% reduction in RMSE compared to the baseline. Post-hoc SHAP analysis further reveals that MPNet embedding dimensions contribute substantially to predictions, capturing latent cost drivers such as subsurface uncertainties and administrative delays that are not explicitly represented in structured data. These findings demonstrate that contextual sentence embeddings can serve as effective inputs for regression models, enabling more accurate and interpretable construction cost estimation. © 2026 Elsevier B.V.

키워드