Probabilistic risk identification and assessment model for construction projects using elicitation based bayesian network

While risks in construction projects have severe consequences on the project schedule, budget, quality, and safety, the realm of Risk Management (RM) falls short in terms of efficiency, productivity, and automation. Artificial Intelligence technologies, especially Machine Learning, can address these issues and utilize risk data effectively for informed decision-making. However, due to the infrequent and unstructured data registration in projects, deterministic RM approaches with a frequentist inference are inapplicable to such small databases and cannot represent the actual risk exposure accurately. This research proposes two solutions to compensate for the data scarcity issue: a) Elicitation, which allows for the integration of subjective and experience-based expert opinions with the existing objective project database, and b) Synthetic data generation using Generative Adversarial Networks (GANs) for data augmentation. A probabilistic model based on a Bayes inference is developed, where experts' opinions are quantified and used for learning the structure and primary parameters in a Bayesian Networks (BN) representing the overall risk network of the case study. A case study of 44 construction projects in Italy is utilized for belief updates in the network, and cross-validation and elicitation methods are employed to validate the results. The results confirm the effectiveness of both solutions, as the overall model accuracy increased by 18% using GANs for synthetic generation and the collective experts' opinions served as a basis to prevent the overfitting of the model to the limited project data. These findings underscore the superiority of probabilistic ML approaches in limited databases, contributing to the body of knowledge in the construction RM field and to the enhancement of precision and productivity of RM practices in the industry.