Key points are not available for this paper at this time.
Apartment buildings are significantly popular among South Korean construction companies. However, design changes present a common yet challenging aspect, often leading to cost overruns. Traditional cost prediction methods, which primarily rely on numerical data, have a gap in fully capitalizing on the rich insights that textual descriptions of design changes offer. Addressing this gap, this research employs machine learning (ML) and natural language processing (NLP) techniques, analyzing a dataset of 35,194 instances of design changes from 517 projects by a major public real estate developer. The proposed models demonstrate acceptable performance, with R-square values ranging from 0.930 to 0.985, underscoring the potential of integrating structured and unstructured data for enhanced predictive analytics in construction project management. The predictor using Extreme Gradient Boosting (XGB) shows better predictive ability (R2 = 0.930; MAE = 16.05; RMSE = 75.09) compared to the traditional Multilinear Regression (MLR) model (R2 = 0.585; MAE = 43.85; RMSE = 101.41). For whole project cost changes predictions, the proposed models exhibit good predictive ability, both including price fluctuations (R2 = 0.985; MAE = 605.1; RMSE = 1009.5) and excluding price fluctuations (R2 = 0.982; MAE = 302.1; RMSE = 548.5). Additionally, a stacked model combining CatBoost and Support Vector Machine (SVM) algorithms was developed, showcasing the effective prediction of cost changes, with or without price fluctuations.
Ahn et al. (Tue,) studied this question.