Research map / Graph learning / GNN architectures
Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering
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 |
|---|---|---|---|
| Chebyshev spectral filters (localized, K=25) instead of non-parametric filters (ablation) | graph CNN on MNIST · MNIST GC32-P4-GC64-P4-FC512 | accuracy (%): 96.28 -> 99.14 | Helped |
| Chebyshev parametrization instead of spline parametrization of spectral filters | spectral graph CNN · MNIST GC32-P4-GC64-P4-FC512 | accuracy (%): 97.15 -> 99.14 | Helped |
| Chebyshev parametrization instead of spline parametrization of spectral filters | single graph conv layer GC10 · MNIST GC10 | accuracy (%): 97.26 -> 97.48 | Helped |
| Chebyshev filters instead of non-parametric filters on single layer (ablation) | graph CNN GC10 · MNIST GC10 | accuracy (%): 95.75 -> 97.48 | Helped |
| Use 8-NN 2D grid graph instead of random graph (ablation) | graph CNN on MNIST · MNIST GC32-P4-GC64-P4-FC512 | accuracy (%): 95.39 -> 99.14 | Helped |
| Use 8-NN 2D grid graph instead of random graph (ablation) | GC32 single layer · MNIST GC32 | accuracy (%): 96.88 -> 97.40 | Helped |
| Learned word2vec feature graph instead of bag-of-words-based graph (ablation) | GC32 on 20NEWS · 20NEWS | accuracy (%): 67.50 (bag-of-words) vs 68.26 (learned word2vec) | Mixed |
| Random feature graph instead of word2vec-based graph (ablation) | GC32 on 20NEWS · 20NEWS | accuracy (%): 68.26 -> 67.75 | Hurt |
| Proposed graph CNN vs classical CNN (time per minibatch) | graph CNN on MNIST · MNIST time per mini-batch, CPU | time (ms): 210 -> 1600 ms | Hurt |
| Graph CNN with Chebyshev filters vs. classical CNN accuracy on MNIST | graph CNN on MNIST · MNIST | accuracy (%): 99.33 vs 99.14 | No clear effect |
| Graph CNN (GC32) vs fully connected networks on 20NEWS | graph CNN on 20NEWS · 20NEWS | accuracy (%): 68.26 vs FC2500 64.64 and FC2500-FC500 65.76 | Helped |
| Graph CNN (GC32) vs multinomial naive Bayes on 20NEWS | graph CNN on 20NEWS · 20NEWS | accuracy (%): 68.26 vs 68.51 | Hurt |
Closest papers by meaning
- Simple Spectral Graph Convolution (ICLR 2021 · similarity 0.93)
- HoloNets: Spectral Convolutions do extend to Directed Graphs (ICLR 2024 · similarity 0.90)
- Generalizing CNNs to graphs with learnable neighborhood quantization (NeurIPS 2024 · similarity 0.89)
- Spatio-Spectral Graph Neural Networks (NeurIPS 2024 · similarity 0.89)
- A Manifold Perspective on the Statistical Generalization of Graph Neural Networks (ICML 2025 · similarity 0.89)
- SLOG: An Inductive Spectral Graph Neural Network Beyond Polynomial Filter (ICML 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.