Optimizing infrastructure procurement: A predictive model for procurement delivery method selection in New Zealand context

Procurement delivery method selection is critical for successful construction projects, enabling informed decisions, efficient capacity management, risk minimization, and optimal outcomes. Traditional selection methods, such as weightage systems, often require extensive features, data, and assessments, which can be subjective and biased, leading to suboptimal decision-making and hindering strategic planning. Hence, this study aims to develop a data-driven procurement delivery method selection tool, leveraging machine learning as a decision support system tailored for New Zealand's infrastructure. A dataset of 2,276 infrastructure pipeline projects with project-related features from New Zealand Infracom was used to train machine learning models through the Bagging Ensemble technique. Models, including SVM, AdaBoost, KNN, Random Forest, XGBoost, and LightGBM, were evaluated to identify the best performer in predicting procurement delivery methods. Generative Adversarial Networks (GAN) combined with SMOTE were applied for data augmentation, enhancing robustness and accuracy. Random Forest achieved the highest accuracy (92.58%), followed by XGBoost (92.25%) and LightGBM (91.92%), with validation on an external dataset confirming 95.82% accuracy and reliability. By incorporating evidence-based decision-support tools, this research equips stakeholders with actionable insights, supporting informed and strategic infrastructure procurement and capacity building.