Research map / Graph learning / GNN architectures
LanczosNet: Multi-Scale Deep Graph Convolutional 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 |
|---|---|---|---|
| Use long diffusion scales instead of only short scales (ablation) | LanczosNet · QM8 validation | MAE (× 1.0e^{-3}): 10.60 -> 10.54 | Helped |
| Combine short and long diffusion scales (ablation) | LanczosNet · QM8 validation | MAE (× 1.0e^{-3}): 10.54 -> 10.41 | Helped |
| Set the Lanczos step to 10 rather than 20 (ablation) | LanczosNet with long scales {10, 20, 30} · QM8 validation | MAE (× 1.0e^{-3}): 10.54 -> 10.44 | Helped |
| Increase the Lanczos step from 10 to 20 (ablation) | LanczosNet with long scales {10, 20, 30} · QM8 validation | MAE (× 1.0e^{-3}): 10.44 -> 10.54 | Hurt |
| Add a 3-layer learnable spectral filter (ablation) | LanczosNet · QM8 validation | MAE (× 1.0e^{-3}): 10.54 -> 10.44 | Helped |
| Use a 5-layer rather than 3-layer learnable spectral filter (ablation) | LanczosNet · QM8 validation | MAE (× 1.0e^{-3}): 10.44 -> 10.54 | Hurt |
| Learn node embeddings rather than fixing them as one-hot (ablation) | LanczosNet with 3-layer spectral filter · QM8 validation | MAE (× 1.0e^{-3}): 10.44 -> 10.26 | Helped |
| Learn node embeddings rather than learning a graph kernel (ablation) | AdaLanczosNet · QM8 validation | MAE (× 1.0e^{-3}): 10.99 -> 10.20 | Helped |
| Tune the scale parameters and use the listed long-scale set (ablation) | LanczosNet with learned node embeddings · QM8 validation | MAE (× 1.0e^{-3}): 10.26 -> 9.56 | Helped |
| Use AdaLanczosNet rather than LanczosNet | Graph convolution models · QM8 test | MAE (× 1.0e^{-3}): 9.58 -> 9.97 | Hurt |
| Use LanczosNet rather than GAT | Citation node classification · Cora public | accuracy: 82.6 $\pm$ 0.7 -> 79.5 $\pm$ 1.8 | Hurt |
| Use AdaLanczosNet rather than the strongest baseline DCNN | Citation node classification · Cora 3% training | accuracy: 76.7 $\pm$ 2.5 -> 77.7 $\pm$ 2.4 | Helped |
Closest papers by meaning
- Break the Ceiling: Stronger Multi-scale Deep Graph Convolutional Networks (NeurIPS 2019 · similarity 0.89)
- Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering (NeurIPS 2016 · similarity 0.89)
- Simple Spectral Graph Convolution (ICLR 2021 · similarity 0.88)
- HoloNets: Spectral Convolutions do extend to Directed Graphs (ICLR 2024 · similarity 0.87)
- Simplifying Graph Convolutional Networks (ICML 2019 · similarity 0.87)
- Generalizing CNNs to graphs with learnable neighborhood quantization (NeurIPS 2024 · similarity 0.86)
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.