A wide range of networked systems is spatially embedded, with structural organization shaped by topological connectivity and spatial distribution. Existing studies on critical component identification mainly focus on nodes or edges, while giving relatively limited attention to regional characteristics and inter-regional topological coupling. Accordingly, key region identification in three-dimensional spatial networks via graph neural networks (3D-KRIG) is proposed. The method partitions space into cubic cells, maps the original network into a region graph, and extracts regional structural features. On this basis, graph sample and aggregation (GraphSAGE) is employed to aggregate neighborhood information and learn regional embeddings, followed by a multi-layer perceptron (MLP) that outputs key region scores. Experimental results show that 3D-KRIG generally outperforms seven baseline methods on synthetic networks and unmanned aerial vehicle (UAV) swarm simulation networks, effectively identifying key regions with significant influence on network connectivity. Runtime analysis further indicates that the proposed method achieves stable key region identification performance with limited additional computational cost.
 

three-dimensional spatial networks, key region identification, graph neural networks, network dismantling, unmanned aerial vehicle swarm