| Title |
A Machine Learning Framework for Early-Stage Underground Construction Cost Estimation in Apartment Projects Using mRMR Feature Selection |
| Authors |
허리정(Xu, Lijing) ; 홍영록(Hong, Rong-Lu) ; 김건우(Kim, Geon-U) ; 김주형(Kim, Ju-Hyung) |
| DOI |
https://doi.org/10.5659/JAIK.2026.42.9.405 |
| Keywords |
Apartment Complex; Underground Construction; Feature Selection; Cost Prediction; mRMR |
| Abstract |
Underground construction projects often exhibit substantial cost variability during the early stages due to limited data availability and
insufficient design detail, making accurate cost estimation challenging. Conventional statistical models, which rely on assumptions of linearity
and normality, are often limited in their ability to capture the nonlinear characteristics of underground construction costs. To address these
limitations, this study analyzes data from 62 real-world underground construction projects categorized into excavation and earth-retaining
works. The minimum Redundancy Maximum Relevance (mRMR) method is applied to identify key features while minimizing redundancy.
Based on the selected features, linear regression and five machine learning models, including Random Forest Regression (RFR) and XGBoost,
are developed and compared. In addition, SHapley Additive exPlanations (SHAP) is employed to evaluate variable contributions and improve
model interpretability. The results indicate that the optimal prediction model varies by construction type. RFR demonstrates stable performance
for excavation works, whereas XGBoost achieves higher predictive accuracy for earth-retaining works. Furthermore, mRMR-based feature
selection improves the performance of certain models by reducing redundancy among input variables. This study provides a comparative
evaluation of cost prediction models and assesses their applicability to different construction types under conditions of limited early-stage
project information. |