Machine Learning Algorithms for Construction Projects Delay Risk Prediction

TLDR

Construction projects frequently suffer delays due to complex, interdependent risk sources, and while machine learning offers a promising approach to address such complexity, its adoption in the sector remains early. This study aims to develop machine learning models for accurate project delay risk analysis and prediction using objective data. The authors compiled a multivariate dataset of past project performance and risk factors, performed exploratory analysis to uncover system complexity, and trained decision‑tree and naïve‑Bayes classifiers, evaluating their predictive performance with cross‑validation and performance metrics. The naïve‑Bayes model outperformed the decision‑tree model, demonstrating that machine learning can support evidence‑based, proactive risk management in construction projects.

Abstract

Projects delays are among the most pressing challenges faced by the construction sector attributed to the sector's complexity and its inherent delay risk sources' interdependence. Machine learning offers an ideal set of techniques capable of tackling such complex systems; however, adopting such techniques within the construction sector remains at an early stage. The goal of this study was to identify and develop machine learning models in order to facilitate accurate project delay risk analysis and prediction using objective data sources. As such, relevant delay risk sources and factors were first identified, and a multivariate data set of previous projects' time performance and delay-inducing risk sources was then compiled. Subsequently, the complexity and interdependence of the system was uncovered through an exploratory data analysis. Accordingly, two suitable machine learning models, utilizing decision tree and naïve Bayesian classification algorithms, were identified and trained using the data set for predicting project delay extents. Finally, the predictive performances of both models were evaluated through cross validation tests, and the models were further compared using machine-learning-relevant performance indices. The evaluation results indicated that the naïve Bayesian model provides a better predictive performance for the data set examined. Ultimately, the work presented herein harnesses the power of machine learning to facilitate evidence-based decision making, while inherent risk factors are active, interdependent, and dynamic, thus empowering proactive project risk management strategies.