SGT: an energy-efficient spiking graph transformer for graph representation learning

Attention mechanisms have been widely explored in graph data learning, significantly enhancing the representational capabilities of Graph Neural Networks (GNNs). However, conventional attention models generally require exhaustive modeling of all node relationships, leading to high computational complexity and substantial resource consumption, especially on large-scale graphs. To overcome this limitation, this paper proposes SGT, an energy-efficient Spiking Graph Transformer built upon Spiking Neural Networks (SNNs). By exploiting the sparse computation and event-driven characteristics of SNNs, SGT selectively establishes feature interactions among critical nodes, thereby effectively reducing the computational burden associated with attention mechanisms. Experimental evaluations on multiple benchmark datasets demonstrate that SGT achieves superior computational efficiency while maintaining competitive predictive performance, indicating its strong potential for scalable graph learning.