Prediction of project activity delays caused by variation orders: a machine-learning approach

Project activity delays caused by variation orders (VOs) can compromise the achievement of timely project completion. Previous research on machine learning (ML) applications for delay predictions has mainly been concerned with delays on a whole project level, whereas predictions of delays in individual project activities have received less attention. This study is a pilot study to investigate how data from large project databases can be used for an ML analysis. The application is aimed at providing early warnings of delays related to VOs in construction projects. The study was performed following typical ML model development steps including data collection, data preprocessing, model training, and testing. A compound dataset was retrieved from project-planning software utilised in a large project. Four pilot tree-based ML models, namely, Decision Tree, Random Forest, AdaBoost, and Gradient Boosting, were trained and tested on a pre-processed dataset comprising 11194 activities. The overall best-performing model was Random Forest with 92.7% and 91.8% recall on DELAYED START and DELAYED FINISH, respectively. By emphasizing that project participants' competency and personal accountability might influence the timely implementation of scope adjustments, these findings advance the field of project management research. An approach like the use of tree-based ML algorithms is applicable for analyses of individual activities in other construction projects. Considering the capability of ML algorithms to capture complex interconnections in raw data extracted from project-planning software, further development of such ML models will enable the establishment of an AI-based Early Warning System (EWS) that can flag potential delays caused by VO requests.