Research map / Graph learning / GNN architectures
How Powerful are Spectral Graph Neural Networks
Xiyuan Wang, Muhan Zhang
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 |
|---|---|---|---|
| remove nonlinearity: use linear spectral GNN with Jacobi basis (JacobiConv) instead of nonlinear spectral GNNs | spectral GNNs on real-world node classification · 10 real-world datasets (Table 2) | accuracy (%): up to 12% gain on Squirrel; best on 9 of 10 datasets | Improved |
| add nonlinearity via 2-layer ReLU MLP in place of linear transform (ablation) | JacobiConv · Cora, Computers, Photo, Squirrel etc. (Table 3) | accuracy (%): about 6% average loss vs JacobiConv | Hurt |
| use Jacobi basis instead of monomial/Chebyshev/Bernstein bases in linear GNN (ablation) | linear spectral GNN · real-world datasets (Table 3) | accuracy (%): more than 0.8% average over other bases | Improved |
| use multiple per-output-dimension filter functions instead of one shared filter (ablation) | JacobiConv · real-world datasets (Table 3) | accuracy (%): about 1.3% average gain | Improved |
| add polynomial coefficient decomposition (PCD) (ablation) | JacobiConv · real-world datasets (Table 3) | accuracy (%): about 0.8% average gain | Improved |
| replace node features with random features (no learnable random features) (ablation) | JacobiConv · real-world datasets (Table 4) | accuracy (%): Cora 88.98 to 23.46 | Hurt |
| Jacobi basis vs Bernstein basis in linear filters on synthetic image filter learning (ablation) | linear GNN · synthetic image filter learning (Table 1) | average squared loss: Jacobi 0.0003 vs Bernstein 0.0110 on Low filter | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BernNet — BernNet: Learning Arbitrary Graph Spectral Filters via Bernstein Approximation (NeurIPS 2021) · better on 4 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- On the Expressive Power of Spectral Invariant Graph Neural Networks (ICML 2024 · similarity 0.94)
- Position: Spectral GNNs Rely Less on Graph Fourier Basis than Conceived (ICML 2025 · similarity 0.93)
- Homomorphism Expressivity of Spectral Invariant Graph Neural Networks (ICLR 2025 · similarity 0.92)
- Spectral Graph Neural Networks are Incomplete on Graphs with a Simple Spectrum (NeurIPS 2025 · similarity 0.92)
- Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025 · similarity 0.92)
- SLOG: An Inductive Spectral Graph Neural Network Beyond Polynomial Filter (ICML 2024 · similarity 0.92)
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.
