Machine Learning-Based Prediction of Construction Labor Productivity Using Realistic Synthetic Datasets

This study responds to the vital problem of limited data in the field of construction by applying a synthetically sampled dataset to forecast labor productivity. Seven machine learning models, which are Linear Regression, Support Vector Machine, Artificial Neural Network, Decision Tree, Random Forest, Gradient Boosting, and TreeBagger, were trained and benchmarked. The results show that ensemble models and the Artificial Neural Network performed the best with the highest R2 score, effectively recovering the complex non-linear relationships and interaction effects embedded in the data-generation logic. These models showed high robustness, stability in cross-validation, and data noise resistance as evidenced by the sensitivity analysis. The analysis of feature importance ranked the worker experience and site conditions as the most powerful predictors for most models. A major conclusion is that the best performing models were extremely cross-situational, performing well with new distribution-shift samples. The study offers a useful benchmarking platform to the academic community and a workable data-based baseline, serving as methodological groundwork for any future automated productivity management system. The direction of future work is aimed at empirical testing and considerations of more complex deep learning networks related to dynamic forecasting.