Research map / Graph learning / GNN architectures
Towards Better Graph Representation Learning with Parameterized Decomposition & Filtering
Mingqi Yang, Wenjie Feng, Yanming Shen, Bryan Hooi
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 |
|---|---|---|---|
| learnable decomposition (eps,k) instead of fixed Laplacian, with linear filter (channel-shared) (ablation) | PDF shd · ZINC | test MAE: 0.219 -> 0.167 | Improved |
| learnable decomposition (eps,k) instead of fixed Laplacian, with linear filter (channel-independent) (ablation) | PDF idp · ZINC | test MAE: 0.172 -> 0.150 | Improved |
| more expressive f (Lin -> 1-layer -> 2-layer MLP) (ablation) | PDF idp · ZINC | test MAE: 0.150 -> 0.088 -> 0.066 | Improved |
| more expressive f (Lin -> 1-layer -> 2-layer MLP) (ablation) | PDF shd · ZINC | test MAE: 0.167 -> 0.160 -> 0.112 | Improved |
| restrict representations to 2-hop sparse subspace (ablation) | PDF idp 2L · ZINC | test MAE: 0.066 -> 0.111 | Hurt |
| channel-independent (idp) instead of channel-shared (shd) decomposition/filtering (ablation) | PDF 2L · ZINC | test MAE: 0.112 (shd) -> 0.066 (idp) | Improved |
| PDF (idp, 2-layer MLP, learnable decomposition) vs. GCN/GIN baselines | PDF · ZINC | test MAE: 0.066 vs GPS 0.070, GIN 0.526 | Improved |
| PDF vs. GPS on ogbg-ppa (sparse 1-hop subspace) | PDF · ogbg-ppa | ACC (%): 80.10 vs 80.15 | No clear effect |
| PDF on TUDataset, 6 of 8 datasets best accuracy | PDF · TUDataset | accuracy (%): best on NCI1, NCI109, ENZYMES, PTC_MR, IMDB-B, RDT-B; not MUTAG (FGSD 92.12) or PROTEINS (DiffPool 78.1) | Mixed |
Closest papers by meaning
- Position: Spectral GNNs Rely Less on Graph Fourier Basis than Conceived (ICML 2025 · similarity 0.93)
- Unifying Homophily and Heterophily for Spectral Graph Neural Networks via Triple Filter Ensembles (NeurIPS 2024 · similarity 0.92)
- GrokFormer: Graph Fourier Kolmogorov-Arnold Transformers (ICML 2025 · similarity 0.90)
- HoloNets: Spectral Convolutions do extend to Directed Graphs (ICLR 2024 · similarity 0.90)
- How Universal Polynomial Bases Enhance Spectral Graph Neural Networks: Heterophily, Over-smoothing, and Over-squashing (ICML 2024 · similarity 0.90)
- 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.
