ABSTRACT
This paper analyses how well decision trees (DTs) predict the construction duration in the predesign phase. To answer this question, the authors first evaluate the expected prediction accuracy with a survey in the German AEC industry. Second, they compare the prediction accuracy of five DTs (Random Forest, GBR, XGBoost, LightGBM and CatBoost) with artificial neural networks (ANN) and linear regression models (LRMs) in two exemplary data sets from residential projects. The study uses performance indicators mean absolute error (MAE) and mean absolute percentage error (MAPE) as metrics. The results reveal that DTs perform better, with the underlying data sets, than ANNs and LRMs. The expected prediction accuracy of 26% is fulfilled in data set 1 with a MAPE of 13.48% and is nearly reached in data set 2 with a 26.45% MAPE. This shows the potential of using DTs in practice as more and more data in construction is generated. From a practical perspective, the explainabilty of DTs should be further tested in predicting the construction duration.
