Scaffold: Support Graph Theory Based Sparsification for Graph Neural Networks
Scaffold keeps GNN sparse graphs useful by controlling both reroute length and bottleneck pressure.
The paper presents an unsupervised sparsification method based on support graph theory preconditioners. Its criterion pairs dilation, for how far removed-edge communication must detour, with congestion, for how much those detours pile onto retained edges. Across 19 benchmarks, the authors report the best aggregate rank among evaluated methods. With 10%-50% of edges retained, Scaffold often recovers or nears full-graph GNN performance while cutting memory below half and reducing total training time, including sparsification overhead. ArXiv · AI/CL/LG's note
The paper presents an unsupervised sparsification method based on support graph theory preconditioners. Its criterion pairs dilation, for how far removed-edge communication must detour, with congestion, for how much those detours pile onto retained edges. Across 19 benchmarks, the authors report the best aggregate rank among evaluated methods. With 10%-50% of edges retained, Scaffold often recovers or nears full-graph GNN performance while cutting memory below half and reducing total training time, including sparsification overhead. ArXiv · AI/CL/LG's note
score 4