Research map / Graph learning / Knowledge and temporal graphs
Sparse Structure Learning via Graph Neural Networks for Inductive Document Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| learn sparse inter-sentence edges from document-specific contextual dependencies rather than use static word co-occurrence edges (ablation) | Our proposal · R8, R52, Ohsumed, MR, 20NG | accuracy: WordCooc: 97.20$\pm$0.29, 93.82$\pm$0.15, 68.08$\pm$0.32; our graph w/ reg: 97.81$\pm$0.14, 95.48$\pm$0.26, 70.59$\pm$0. | Helped |
| remove inter-sentence edges (ablation) | Our proposal · R8, R52, Ohsumed, MR, 20NG | accuracy: 97.81$\pm$0.14 -> 97.29$\pm$0.21; 95.48$\pm$0.26 -> 94.80$\pm$0.20; 70.59$\pm$0.38 -> 69.72$\pm$0.27; 79.74$\pm$0.19 -> | Hurt |
| connect all words between sentences instead of adaptively selecting sparse edges (ablation) | Our proposal · R8, R52, Ohsumed, MR, 20NG | accuracy: 97.81$\pm$0.14 -> 97.40$\pm$0.25; 95.48$\pm$0.26 -> 94.35$\pm$0.10; 70.59$\pm$0.38 -> 67.57$\pm$0.30; 79.74$\pm$0.19 -> | Mixed |
| add regularization to preserve original syntactic structure (ablation) | Our proposal · R8, R52, Ohsumed, MR, 20NG | accuracy: 97.76$\pm$0.16 -> 97.81$\pm$0.14; 95.32$\pm$0.21 -> 95.48$\pm$0.26; 70.53$\pm$0.30 -> 70.59$\pm$0.38; 79.74$\pm$0.11 -> | No clear effect |
| vary the adaptive-sampling temperature (ablation) | Our proposal · R8, R52, Ohsumed, MR, 20NG | accuracy: R8: 97.50$\pm$0.29, 97.34$\pm$0.13, 97.44$\pm$0.39, 97.81$\pm$0.14, 97.35$\pm$0.24; R52: 95.16$\pm$0.18, 95.48$\pm$0.26, | Mixed |
| use learned sparse document graphs rather than sentence-disjoint graphs (ablation) | Our proposal · R8 and R52 | accuracy: 97.29$\pm$0.21 -> 97.76$\pm$0.16; 94.80$\pm$0.20 -> 95.32$\pm$0.21 | Helped |
| use learned sparse document graphs rather than sentence-disjoint graphs (ablation) | Our proposal · Ohsumed, MR, 20NG | accuracy: 69.72$\pm$0.27 -> 70.53$\pm$0.30; 78.61$\pm$0.12 -> 79.74$\pm$0.11; 84.92$\pm$0.28 -> 85.15$\pm$0.33 | Helped |
| use sentence-level structure and sparse dynamic context in an inductive graph model | Our proposal · MR | accuracy: 78.93$\pm$0.65 -> 79.74$\pm$0.19 | Helped |
| use sentence-level structure and sparse dynamic context in an inductive graph model | Our proposal · R8 | accuracy: 97.34$\pm$0.25 -> 97.81$\pm$0.14 | Helped |
| use sentence-level structure and sparse dynamic context in an inductive graph model | Our proposal · R52 | accuracy: 94.15$\pm$0.18 -> 95.48$\pm$0.26 | Helped |
| use sentence-level structure and sparse dynamic context in an inductive graph model | Our proposal · Ohsumed | accuracy: 66.39$\pm$0.65 -> 70.59$\pm$0.38 | Helped |
| use sentence-level structure and sparse dynamic context in an inductive graph model | Our proposal · 20NG | accuracy: 84.65$\pm$0.31 -> 85.26$\pm$0.28 | Helped |
Closest papers by meaning
- Two Heads Are Better Than One: Boosting Graph Sparse Training via Semantic and Topological Awareness (ICML 2024 · similarity 0.85)
- Scale-Free Graph-Language Models (ICLR 2025 · similarity 0.85)
- Label-free Node Classification on Graphs with Large Language Models (LLMs) (ICLR 2024 · similarity 0.85)
- Improved Graph Contrastive Learning for Short Text Classification (AAAI 2024 · similarity 0.85)
- Fully-inductive Node Classification on Arbitrary Graphs (ICLR 2025 · similarity 0.84)
- Graph-KV: Breaking Sequence via Injecting Structural Biases into Large Language Models (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.