GNN-based prediction of closeness centrality in urban road networks

Identifying key nodes in urban road networks is essential for effective planning, management, and resilience assessment. we propose an end-to-end graph neural network (GNN) framework for efficient key node identification. Our approach utilizes road network data from Guangzhou’s Tianhe District, where we identify the top 5% high-centrality nodes through perturbation analysis to create a large-scale training dataset. The proposed model integrates multi-path embedding with Feature Tokenizer Transformer (FT-Transformer) branches to capture attribute-level attention features, while a hierarchical GATv2 architecture enhances structural representation learning. Experimental results indicate that the proposed method attains competitive detection accuracy, recall, and F1-score performance.