Scalable machine learning framework for predicting critical links in urban networks

Efficient identification of critical links in urban road networks is essential for optimizing traffic management, infrastructure planning, and resource allocation. Existing methods, such as simulation-based approaches, are computationally expensive and often impractical for large-scale networks. This study proposes a scalable machine learning framework capable of training on a subset of network links (20%) and predicting the criticality of remaining links with approximately 7% percentage mean error. The framework integrates structural, functional, and newly proposed features, offering a comprehensive representation of road network dynamics. Validated on two diverse datasets, namely, Luxembourg (LuST) and Monaco (MoST), the framework achieves high precision (~72% and ~73% in single-city scenarios) and robust cross-city performance (~70% for LuST -> MoST and ~66% for MoST -> LuST). Random Forest and Gradient Boosting emerged as the top-performing models, consistently delivering the best precisions and lowest number of errors. The inclusion of dynamic traffic metrics and advanced preprocessing techniques further enhanced predictive accuracy and generalization capabilities. This study highlights the potential of machine learning for scalable critical link evaluation, demonstrating its applicability to large-scale networks with limited data. The findings provide actionable insights for urban traffic management and open pathways for future research, including domain adaptation, temporal modeling, and integration with real-time systems.