Research map / Graph learning / Graph self-supervised learning
Rethinking Kernel Methods for Node Representation Learning on Graphs
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 K3, combining kernel and softmax losses, instead of FastGCN | Node classification · Cora, Citeseer, and Pubmed (FastGCN split) | accuracy (%): 85.0 -> 88.40 ± 0.24; 77.6 -> 80.28 ± 0.03; 88.0 -> 89.42 ± 0.01 | Helped |
| Use K3 rather than JK-Concat | Node classification · Cora and Citeseer (JK split) | accuracy (%): 89.10 ± 1.10 -> 89.24 ± 0.31; 78.30 ± 0.80 -> 80.78 ± 0.28 | Helped |
| Use the RBF rather than dot-product base kernel (ablation) | Kernel variants K1 and K2 · Citeseer and Pubmed (FastGCN split) | accuracy (%): 77.92 ± 0.25 -> 78.68 ± 0.38; 89.22 ± 0.17 -> 89.36 ± 0.21 | Helped |
| Replace the proposed single-step feature aggregation with GCN or GAT feature mapping (ablation) | K1 · Cora and Citeseer (FastGCN split) | accuracy (%): 86.68 ± 0.17 -> 87.04 ± 0.09 (Cora); 77.92 ± 0.25 -> 77.12 ± 0.23 (Citeseer, GCN); 77.92 ± 0.25 -> 77.92 ± 0.19 (Citesee | Hurt |
| Add the kernel similarity loss to softmax training (ablation) | Feature mapping gθ · Cora, Citeseer, and Pubmed (FastGCN split) | accuracy (%): 87.56 ± 0.14 -> 88.40 ± 0.24; 79.80 ± 0.03 -> 80.28 ± 0.03; 89.24 ± 0.14 -> 89.42 ± 0.01 | Helped |
| Increase neighborhood aggregation from 1-hop to the default 2-hop (ablation) | K3 · Cora, Citeseer, and Pubmed (FastGCN split) | accuracy (%): 85.56 ± 0.02 -> 88.40 ± 0.24; 77.73 ± 0.02 -> 80.28 ± 0.03; 88.98 ± 0.01 -> 89.42 ± 0.01 | Helped |
| Increase aggregation from 2-hop to 3-hop (ablation) | K3 · Cora, Citeseer, and Pubmed (FastGCN split) | accuracy (%): 88.40 ± 0.24 -> 88.25 ± 0.01; 80.28 ± 0.03 -> 80.13 ± 0.01; 89.42 ± 0.01 -> 89.53 ± 0.01 | Mixed |
| Reduce the MLP from 2 layers to 1 layer (ablation) | K3 · Cora, Citeseer, and Pubmed (FastGCN split) | accuracy (%): 88.40 ± 0.24 -> 82.60 ± 0.01; 80.28 ± 0.03 -> 77.63 ± 0.01; 89.42 ± 0.01 -> 85.80 ± 0.01 | Hurt |
| Increase the MLP from 2 layers to 3 layers (ablation) | K3 · Cora, Citeseer, and Pubmed (FastGCN split) | accuracy (%): 88.40 ± 0.24 -> 86.33 ± 0.04; 80.28 ± 0.03 -> 78.53 ± 0.20; 89.42 ± 0.01 -> 89.46 ± 0.05 | Mixed |
| Replace learned hop weights with fixed c^h weights (ablation) | K3 · Cora, Citeseer, and Pubmed (FastGCN split) | accuracy (%): Default: 88.40 ± 0.24, 80.28 ± 0.03, 89.42 ± 0.01; fixed-weight results: c = 0.25: 69.33 ± 0.09, 74.48 ± 0.03, 84.68 ± 0 | Hurt |
| Train node embeddings with kernel similarity rather than classification loss alone | K3 versus GCN and GAT · Citeseer | embedding cluster compactness | Helped |
| Apply the proposed kernel framework to link prediction | Kernel node representation method · Cora, Citeseer, and Pubmed link prediction | AUC: 93.1 ± 0.06 vs 91.4 ± 0.01 (Cora); 90.9 ± 0.08 vs 90.8 ± 0.02 (Citeseer); 94.5 ± 0.03 vs 96.4 ± 0.00 (Pubmed) | Mixed |
Closest papers by meaning
- Convolutional Kernel Networks for Graph-Structured Data (ICML 2020 · similarity 0.92)
- Exploring Consistency in Graph Representations: from Graph Kernels to Graph Neural Networks (NeurIPS 2024 · similarity 0.91)
- DHAKR: Learning Deep Hierarchical Attention-Based Kernelized Representations for Graph Classification (AAAI 2025 · similarity 0.91)
- Unsupervised Neighborhood Propagation Kernel Layers for Semi-supervised Node Classification (AAAI 2024 · similarity 0.90)
- Graph Classification via Reference Distribution Learning: Theory and Practice (NeurIPS 2024 · similarity 0.89)
- Temporal Graph Neural Tangent Kernel with Graphon-Guaranteed (NeurIPS 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.