Objective Multi-Metric Fusion for Critical Node Identification via CRITIC and Global-Local Context Modeling

Accurately identifying critical nodes in complex networks and applying targeted protection strategies significantly enhances network security. Traditional importance metrics rely on a single topological feature and cannot fully capture node influence. Existing multi-attribute fusion methods integrate multiple structural sources but typically use fixed weights or predefined rules, failing to adaptively adjust attribute contributions based on local and global network characteristics, which limits their generalization across diverse networks. To address this, we propose the CRITIC-based Objective Weighting and Multi-Metric Fusion Method (COWMF). COWMF first builds a Graph Attention Network with Virtual Global-Local Integration (GAT-VGL), taking four low-complexity topological metrics, degree centrality (DC), H-index, degree and neighborhood information centrality (DNC), and k-shell, as input. Through a learnable attention mechanism, GAT-VGL adaptively aggregates multi-hop neighborhood information and explicitly incorporates global structural information via a virtual node to achieve whole-graph topological awareness, generating a global influence score with good discriminative power and high computational efficiency. This score is then integrated with DC and DNC into an improved CRITIC-based objective weighting fusion scheme, enabling adaptive synergy among local connectivity, semi-local radiation, and global structure. Experiments on six real-world networks of varying types and scales show that COWMF demonstrates relatively stable and competitive performance in both simulated attack and susceptible-infected-recovered (SIR) spreading simulations, two complementary experiments, demonstrating satisfactory disruptive capability and propagation influence. Its importance scores exhibit high monotonicity across all networks, with good discriminative power.