Construction labor productivity estimation through machine learning: Performance comparison of regression algorithms

Labor productivity is a critical factor that directly affects project performance in terms of time, cost, and quality. However, accurately predicting this productivity is challenging due to variability in construction site conditions, individual characteristics of the workforce, and environmental factors. In this study, a total of 18 different regression algorithms, including ensemble methods, tree-based structures, linear regression models, support vector regressions, and artificial neural networks, were systematically evaluated to predict productivity. In the modeling process, a multi-source dataset from large-scale energy infrastructure projects was utilized. The data includes attributes such as environmental conditions, team experience, task complexity, and daily production output for six different activity types. All models are trained within a robust framework, featuring a 70% training and 30% testing separation, 10-fold cross-validation, and hyperparameter optimization using GridSearchCV. Performance evaluation is based on metrics such as R2, RMSE, and MAE. The results of the comparative analysis revealed that the ensemble models, particularly CatBoost, XGBoost, and Bagging, outperformed the others in almost all activity types, achieving higher generalization accuracy and lower prediction error. Beyond model accuracy, attribute importance analyses also provided estimates of the determinants of productivity. Variables such as task complexity, temperature, and team experience were among the prominent factors. These findings demonstrate that data-driven models can also be applied to identify variables that influence productivity, thereby supporting planning processes. The study contributes to the literature by providing a comprehensive model comparison framework and offering practical implications for productivity management in construction projects.