Machine learning-aided engineering services' cost overruns prediction in high-rise residential building projects: Application of random forest regression

Abstract

Current approaches to automating cost estimation mainly focus on construction costs. Yet, the two main services provided by design firms, namely ‘designing the project’, and ‘supervision of construction operations’ labelled as engineering services, despite their comparatively low cost, can significantly affect the total cost of construction projects as they can engender reworks, changes and disputes on project participants during the subsequent stages of the project. Continuous evaluation of engineering services' cost overruns (ESCO) is quintessentially important in order to prevent consequential problems later on in the project's development and use. Consequently, this research proposes a robust random forest (RF) regression model to predict ESCOs considering both project-related and organizational-related variables. A database consisting of 95 high-rise residential building projects designed during the past eight years in Iran, along with 12 related variables, were collected to develop and validate the model. The results were also compared with those of support vector regression (SVR) and multiple linear regression (MLR), which indicated that with an R2 value of 0.8680 and mean-absolute-error (MAE) of 3.88, the RF regression model performs better than those baseline models, namely SVR and MLR. This research presents two main contributions to the existing body of knowledge. From the practical point of view, it provides an efficient tool for design firms enabling them to screen and prioritize their projects from the cost overrun standpoint and to devise a contingency plan for them. From the theoretical point of view, it revealed that to mitigate ESCOs, three key factors should be given thorough consideration, namely: ‘the level of computer-aided design technologies adoption’; ‘level of communication among the project team’; and scope definition adequacy’ – cumulatively, these three factors contribute to 52.35% of ESCO variations.