Research map / Graph learning / Graph self-supervised learning
Graph Generative Model for Benchmarking Graph Neural Networks
Minji Yoon, Yue Wu, John Palowitch, Bryan Perozzi, Russ Salakhutdinov
Read the paperCheck your idea against 137,217 papers
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 label conditioning from CGT (ablation) | CGT · Cora, Citeseer, Pubmed (9 GNNs) | Pearson: 0.943 vs 0.592 without label | Improved |
| Add layer-based position embeddings (ablation) | CGT · Cora, Citeseer, Pubmed (9 GNNs) | Pearson: 0.943 vs 0.411 without | Improved |
| Add ancestor-only attention masks (ablation) | CGT · Cora, Citeseer, Pubmed (9 GNNs) | Pearson: 0.943 vs 0.329 without | Improved |
| Add all three modules (label, position, attention) vs plain Transformer (ablation) | CGT · Cora, Citeseer, Pubmed (9 GNNs) | MSE: 0.017 vs 0.034 | Improved |
| Apply k-anonymity (k=100 to 1000) during quantization (ablation) | CGT · Cora, GNN benchmark correlation | Pearson: 0.934 no privacy -> 0.916 (k=100), 0.862 (k=500), 0.030 (k=1000) | Hurt |
| Use DP-SGD to train the transformer for edge distributions (ablation) | CGT · Cora, GNN benchmark correlation | Pearson: 0.112 at eps=1e6 vs 0.934 no privacy; 0.890 only at eps=1e9 | Hurt |
| Apply DP k-means (eps=1, delta=0.01) for node attribute privacy (ablation) | CGT · Cora, GNN benchmark correlation | Pearson: 0.874 (eps=1) vs 0.934 no privacy | No clear effect |
| Replace privacy method with edge addition, deletion, or attribute noise baselines | GNN benchmark MSE on Cora · Cora, 4 GNNs | MSE: K-anonymity K=5 MSE 0.001 with 20% re-identification; noise-based variants range 0.021 to 0.548 | Mixed |
| CGT vs baselines (VGAE, GraphVAE, GraphAF, GraphDF, GraphEBM) | benchmark effectiveness · 7 public graph datasets, 9 GNNs | MSE and Pearson/Spearman: up to 33% lower MSE, 0.80 higher Pearson, 1.03 higher Spearman | Improved |
| Scale CGT to ogbn-arxiv and ogbn-products via computation graph sampling | CGT · ogbn-arxiv, ogbn-products | Pearson: Pearson 0.989; training 1.1 hr (arxiv), 14.7 hr (products) | Improved |
Closest papers by meaning
- Generative Graph Pattern Machine (NeurIPS 2025 · similarity 0.87)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.87)
- Geometric Generative Modeling with Noise-Conditioned Graph Networks (ICML 2025 · similarity 0.87)
- Can Graph Neural Networks Expose Training Data Properties? An Efficient Risk Assessment Approach (NeurIPS 2024 · similarity 0.87)
- Graph Generative Pre-trained Transformer (ICML 2025 · similarity 0.87)
- Towards Addressing Frontiers in Graph Generation (AAAI 2025 · similarity 0.87)
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.
