EDiS: cached edge-disjoint subgraphs cut GNN sparse-training cost and top benchmarks
Sai Karthik Navuluru · hf · 2026-10-09
EDiS is an edge-disjoint subgraph sparsification framework for GNNs that decouples one-time structural extraction from per-epoch graph composition. The graph is decomposed once into cacheable edge-disjoint subgraphs, then recombined under edge-budget constraints each epoch without re-extraction.
- Default construction uses feature-based scores and successive maximum-score covering forests; the composition mechanism supports alternative edge-selection rules
- Includes combinatorial analysis of the per-epoch sampler, deterministic preservation of high-score cut edges, and a selector-agnostic conditional bound
- Across 19 homophilic, heterophilic, and large-scale node-classification benchmarks against 17 baselines at equal edge budgets, EDiS achieves the highest mean score, lowest average rank, and smallest gap-to-best
- Ablations show structural decomposition and epoch variation help most under tight edge budgets
More from Research
- Why mathematicians call the OpenAI wall the 'death' of a field — the damage is real — stevenstrogatz · 2026-10-09
- First peer-reviewed paper on automating science with AI published in PNAS — Dr_Atoosa · 2026-10-09
- Opus 5.5 one-shots an animated explainer for new LLM hidden valence steering paper — repligate · 2026-10-09
- UCL's Memento 3 lets a frozen LLM self-improve via rulebooks, clearing all 25 ARC-AGI-3 games — UniversityCollegeLondon · 2026-10-09
- Intent-Eval benchmark shows rejected user changes still derail LLM multi-turn task execution — Junle Chen · 2026-10-09
- Conversational AI heads into real-patient clinical trial planned since 2023 — alan_karthi · 2026-10-09