Development of a data-driven framework to predict waste generation and evaluate influential factors: Machine learning innovations in construction waste management

Abstract

The construction industry is responsible for a significant section of waste production among all industries, emphasizing the importance of waste management and revealing the inefficiency of current methods. Accurate estimation of waste and associated factors is the first step to developing a waste management plan, which has remained a challenge. This research develops a machine learning-based framework based on above 500 buildings to predict waste production in different building sections for five major construction materials: concrete, steel, bricks and blocks, tiles and stones, and wood. Using Extreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), Decision Tree Regression (DTR), and Random Forest Regression (RFR), this study identifies the optimal model for each material and construction phase. It also applies SHapley Additive exPlanations (SHAP) analysis to determine key influencing factors. The findings indicate that XGBoost achieved the highest accuracy in 8 out of 13 predictions, with waste estimation reaching over 98 % precision in several cases. The façade stage exhibited the highest variance, posing a greater risk of waste unpredictability. Project duration was the most critical factor, while Building Information Modeling (BIM) had minimal impact. These insights support data-driven waste management practices, helping reduce environmental impact and improve construction efficiency.