Research map / Graph learning / GNN architectures
Learning Discrete Structures for Graph Neural Networks
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 |
|---|---|---|---|
| Jointly learn graph edge probabilities and GCN weights rather than use the fixed input graph | GCN · Citeseer and Cora with 25%, 50%, 75%, or 100% edges retained | test accuracy (%): Citeseer: 64.42 (1.4) -> 71.92 (1.0), 66.82 (0.7) -> 73.26 (0.6), 69.26 (0.6) -> 74.58 (0.9), 70.72 (0.5) -> 75.04 (0.4) | Helped |
| Add randomly sampled edges during standard GCN training | GCN · Citeseer and Cora, varying retained-edge fractions | test accuracy (%): Citeseer: 64.42 (1.4) -> 64.22 (1.3), 66.82 (0.7) -> 67.28 (0.9), 69.26 (0.6) -> 69.66 (0.5), 70.72 (0.5) -> 70.86 (0.7) | No clear effect |
| Initialize edge probabilities from a kNN graph, then optimize them | GCN · Wine, Cancer, Digits, Citeseer, Cora, 20news, FMA | test accuracy (%): LDS vs kNN-GCN: Wine 93.2 (3.1) -> 97.3 (0.4); Cancer 93.8 (1.4) -> 94.4 (1.9); Digits 91.3 (0.5) -> 92.5 (0.7); Citesee | Mixed |
| Use a sparse learned graph instead of the dense LDS adjacency (ablation) | kNN-LDS · Seven classification datasets | test accuracy (%) | Helped |
| Use multiple inner optimization steps when computing STE hypergradients rather than alternating minimization (τ=0) (ablation) | LDS · Citeseer, all settings shown in the τ experiment | test accuracy (%) | Helped |
| Increase τ further to 20 (ablation) | LDS · Citeseer τ experiment | test accuracy (%) | No clear effect |
Closest papers by meaning
- Bi-GCN: Binary Graph Convolutional Network (CVPR 2021 · similarity 0.91)
- Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Learning (AAAI 2018 · similarity 0.91)
- Simplifying Graph Convolutional Networks (ICML 2019 · similarity 0.91)
- A Manifold Perspective on the Statistical Generalization of Graph Neural Networks (ICML 2025 · similarity 0.89)
- On the Interplay between Graph Structure and Learning Algorithms in Graph Neural Networks (ICML 2025 · similarity 0.89)
- Generalizing CNNs to graphs with learnable neighborhood quantization (NeurIPS 2024 · similarity 0.89)
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.