EDiS: Edge Disjoint Subgraph Sparsification Framework for Graph Neural Networks
EDiS creates a framework to make sparse graph training more efficient by separating structural extraction from per-epoch graph composition. It decomposes a large graph once into reusable, edge-disjoint subgraphs and then combines them across different epochs while respecting strict edge budgets. This allows for fast, high-quality training on complex graphs without repeatedly recomputing expensive structural information.