EDiS: Edge Disjoint Subgraph Sparsification Framework for Graph Neural Networks
EDiS tries to make sparse GNN training variable across epochs without paying the extraction cost each time.
The framework decomposes a graph once into cacheable edge-disjoint subgraphs, then recombines them under edge-budget constraints for training. Its default method uses feature-based scores and successive maximum score covering forests, while allowing other edge selection rules. The paper reports the best mean benchmark score and lowest average rank across 19 node-classification benchmarks against 17 baselines at the same edge budget. Ablations say the gains are clearest when edge budgets are tight. HF Daily Papers' note
The framework decomposes a graph once into cacheable edge-disjoint subgraphs, then recombines them under edge-budget constraints for training. Its default method uses feature-based scores and successive maximum score covering forests, while allowing other edge selection rules. The paper reports the best mean benchmark score and lowest average rank across 19 node-classification benchmarks against 17 baselines at the same edge budget. Ablations say the gains are clearest when edge budgets are tight. HF Daily Papers' note
score 4