Research map / Graph learning / GNN architectures
Specformer: Spectral Graph Neural Networks Meet Transformers
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 a set-to-set spectral filter with eigenvalue encoding, self-attention, and learned bases | Specformer · Synthetic spectral-filter recovery (Low-pass, High-pass, Band-pass, Band-rejection, Comb) | sum of squared error: JacobiConv: 0.0003, 0.0064, 0.0213, 0.0156, 0.2933; Specformer: 0.0002, 0.0026, 0.0017, 0.0014, 0.0057 | Helped |
| Use Specformer instead of the strongest listed baseline | Node classification models · Squirrel | accuracy: 57.72±0.59 -> 64.64±0.81 | Helped |
| Use Specformer instead of the strongest listed baseline | Node classification models · Cora | accuracy: 88.98±0.46 -> 88.57±1.01 | Hurt |
| Use Specformer instead of the strongest listed baseline | Node classification models · Citeseer | accuracy: 80.78±0.79 -> 81.49±0.94 | Helped |
| Use Specformer instead of the strongest listed baseline | Graph-level models · MolHIV | AUROC: 0.8094 ± 0.0057 -> 0.7889 ± 0.0124 | Hurt |
| Add eigenvalue encoding (ablation) | Specformer with Medium decoder and no attention · Squirrel | accuracy: 33.05 -> 63.78 | Helped |
| Add eigenvalue encoding (ablation) | Specformer with Medium decoder and no attention · Citeseer | accuracy: 80.57 -> 81.17 | Helped |
| Add self-attention to the eigenvalue encoder (ablation) | Specformer with Medium decoder and eigenvalue encoding · Squirrel | accuracy: 63.78 -> 64.64 | Helped |
| Add self-attention to the eigenvalue encoder (ablation) | Specformer with Medium decoder and eigenvalue encoding · Citeseer | accuracy: 81.17 -> 81.49 | Helped |
| Add self-attention to the eigenvalue encoder (ablation) | Specformer with Medium decoder and eigenvalue encoding · MolPCBA | AP: 0.2933 -> 0.2970 | Helped |
| Use the Small decoder instead of the Medium decoder (ablation) | Specformer with eigenvalue encoding and self-attention · Squirrel | accuracy: 64.64 -> 64.51 | No clear effect |
| Use the Small decoder instead of the Medium decoder (ablation) | Specformer with eigenvalue encoding and self-attention · Citeseer | accuracy: 81.49 -> 81.47 | No clear effect |
Closest papers by meaning
- Spectral Graph Neural Networks are Incomplete on Graphs with a Simple Spectrum (NeurIPS 2025 · similarity 0.90)
- Position: Spectral GNNs Rely Less on Graph Fourier Basis than Conceived (ICML 2025 · similarity 0.90)
- HoloNets: Spectral Convolutions do extend to Directed Graphs (ICLR 2024 · similarity 0.90)
- On the Expressive Power of Spectral Invariant Graph Neural Networks (ICML 2024 · similarity 0.90)
- Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025 · similarity 0.90)
- Unifying Homophily and Heterophily for Spectral Graph Neural Networks via Triple Filter Ensembles (NeurIPS 2024 · similarity 0.90)
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.