Research map / Graph learning / GNN architectures
Revisiting 2D Convolutional Neural Networks for Graph-Based Applications
Yecheng Lyu, Xinming Huang, Ziming Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add vertex separation penalty (lambda>0) (ablation) | GPGL · MUTAG (21x aug) | accuracy (%): 80.51 -> 86.31 | Improved |
| use alpha=1.25 instead of 1.50 (ablation) | GPGL · MUTAG (21x aug) | accuracy (%): 83.04 -> 86.31 | Improved |
| use alpha=1.25 instead of 1.00 (ablation) | GPGL · MUTAG (21x aug) | accuracy (%): 85.14 -> 86.31 | Improved |
| use lambda=1000 instead of 200 or 5000 (ablation) | GPGL · MUTAG (21x aug) | accuracy (%): 85.26 -> 86.31 | Improved |
| add 2D max-pooling to PointNet on grid layout | PointNet · MUTAG (101x aug) | accuracy (%): 66.55 -> 86.23 | Improved |
| use MSM-CNN instead of ResNet50 on GPGL | GPGL · MUTAG (101x aug) | accuracy (%): 91.02 -> 94.18 | Improved |
| add 2D max-pooling and upsampling to PointNet on H-GPGL | PointNet · ShapeNet | mcIoU (%): 80.1 -> 82.0 | Improved |
| use Delaunay triangulation instead of 5-NN graph | H-GPGL MSM-CNN · ShapeNet | mcIoU (%): 82.5 -> 83.8 | Improved |
| use circular initialization instead of spectral | GPGL · MUTAG | vertex loss ratio (%): 10.09 -> 1.06 | Improved |
| use hierarchical GPGL with 16 or 32 cuts instead of GPGL alone | H-GPGL · 2048-node graph | running time: 90s -> 5.5s | Improved |
| increase data augmentation of grid layouts from 1x to 101x | GPGL MSM-CNN · MUTAG | accuracy | Improved |
| overlapped points in grid cells | H-GPGL MSM-CNN · ShapeNet | mcIoU (%): -0.22 mcIoU, -0.21 miIoU | Hurt |
Closest papers by meaning
- Generalizing CNNs to graphs with learnable neighborhood quantization (NeurIPS 2024 · similarity 0.90)
- pathGCN: Learning General Graph Spatial Operators from Paths (ICML 2022 · similarity 0.90)
- Hyperbolic Graph Convolutional Neural Networks (NeurIPS 2019 · similarity 0.89)
- Simplifying Graph Convolutional Networks (ICML 2019 · similarity 0.89)
- Learning Convolutional Neural Networks for Graphs (ICML 2016 · similarity 0.89)
- DHGCN: Dynamic Hop Graph Convolution Network for Self-Supervised Point Cloud Learning (AAAI 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.
