Graph Learning-Based Network Dismantling via hierarchical feature fusion and margin-ranked optimization

Abstract Network dismantling—the strategic removal of nodes to disrupt connectivity in complex systems—is a foundational task in domains such as infrastructure protection, epidemic containment, and systemic risk mitigation. While existing approaches range from heuristic centrality metrics to optimization-based algorithms, they often lack generalization across topologies and fail to learn task-specific node influence. Recent graph learning-based methods, including CLMA, NRAND, and MaxShot-RL, introduce structural adaptability but typically rely on auxiliary modules such as contrastive heads, epidemic simulators, or reinforcement policies, increasing complexity and limiting scalability. Compared to recent reinforcement learning-based approaches such as MaxShot-RL, our method achieves higher scalability and dismantling accuracy while avoiding sample inefficiency and unstable convergence. In this work, we present GLND, a lightweight, self-contained dismantling framework that formulates node importance prediction as a supervised ranking problem. GLND integrates hierarchical attention to capture multi-scale topological patterns, a margin-based ranking loss to enforce score separation, and role-aware attention to amplify structurally pivotal nodes. Extensive experiments on both synthetic and real-world networks demonstrate that GLND consistently improves dismantling effectiveness, suppresses dynamic propagation more efficiently, and scales favorably with graph size, all while maintaining minimal architectural overhead. These results position GLND as a practical and generalizable tool for structural intervention across diverse complex networks.