Feature-centric machine learning framework for demolition supervision fee prediction: evidence from Seoul, Korea

초록

Demolition supervision fees directly affect onsite safety management; however, existing data-driven approaches may not adequately capture complex relationships among demolition project characteristics. This study proposes a feature-centric machine learning framework for demolition supervision fee prediction using 915 demolition supervision contract records from Seoul, Korea. Four feature sets-basic, unit-based, interaction, and log-transformed interaction variables-were evaluated across multiple machine learning models. Interaction-aware feature representations generally improved predictive performance over basic and unit-based configurations. CatBoost with the log-transformed interaction feature set achieved the best single hold-out performance, with a mean absolute error (MAE) of 7,811,301 KRW, root mean squared error (RMSE) of 13,602,176 KRW, coefficient of determination (R2) of 0.880, and weighted absolute percentage error (WAPE) of 23.08%. Repeated train-test split evaluation confirmed that Random Forest provided the most favorable average MAE, R2, and WAPE across repeated data partitions. SHapley Additive exPlanations analysis identified log(Duration), log(Duration & times; Cost), and log(GFA) as key predictors, indicating that supervision fee formation is strongly shaped by duration-related and project-scale effects. Thus, interaction-aware feature engineering improves prediction accuracy and interpretability, supporting preliminary budgeting, fee reasonableness checks, and transparent contract preparation in demolition supervision practice.

키워드