Research map / Graph learning / Graph self-supervised learning
Cross-Domain Graph Data Scaling: A Showcase with Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Pre-trains a self-conditioned discrete diffusion model on thousands of cross-domain graph structures and uses guided generation to augment downstream graph structures (features kept) for graph, link, and node-level tasks.
Problem
Graph pre-training methods struggle to scale across domains due to feature and structural heterogeneity, and existing graph data augmentation methods are task- or dataset-specific and do not transfer.
Approach
Pre-train a structure-only discrete (absorbing) diffusion model with self-supervised cluster labels on thousands of graphs from many domains, then at downstream time steer generation with an MLP guidance head (NOS-style gradient guidance) toward task objectives, and train standard GNNs on the generated structures paired with original node features.
Key results
- Graph classification: UniAug reaches 78.13% on DD and 92.28% on Reddit-B, with average rank 1.11 among graph-level methods, while pre-training baselines show negative transfer vs GIN on many datasets.
- Link prediction: UniAug-NCN improves over NCN on all seven datasets (e.g., Citeseer MRR 61.69 vs 58.03; Yeast Hits@10 42.11 vs 39.85).
- Node classification on heterophilic graphs: UniAug + Half-Hop gives the best accuracy on Cornell (72.43), Wisconsin (79.61), Texas (77.03) and Actor (34.97).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add UniAug structure augmentation | GIN · Graph classification datasets | accuracy | Helped |
| add UniAug structure augmentation | GCN · Link prediction datasets | MRR / Hits@10 | Helped |
| add UniAug structure augmentation | NCN · Link prediction datasets | MRR / Hits@10 | Helped |
| add UniAug structure augmentation | GCN · Heterophilic node classification | accuracy | Mixed |
| add UniAug structure augmentation | GCN · Homophilic node classification | accuracy: 81.75 ± 0.73 -> 81.78 ± 0.60; 70.71 ± 0.76 -> 71.17 ± 0.58; 79.53 ± 0.25 -> 79.54 ± 0.35 | No clear effect |
| add UniAug structure augmentation | GCN · Homophilic node classification | SD across homophily groups: 24.51 ± 1.06 -> 23.45 ± 0.90; 22.57 ± 0.80 -> 19.90 ± 0.81; 27.02 ± 0.56 -> 26.50 ± 0.55 | Helped |
| add Half-Hop to UniAug | UniAug · Cornell, Wisconsin, Texas, Actor, Chameleon, Squirrel | accuracy | Mixed |
| add graph-label guidance | UniAug · Graph property prediction | accuracy / MAE | Helped |
| vary link-prediction guidance objective among link, degree, and CN | UniAug-GCN · Link prediction datasets | MRR / Hits@10 | Mixed |
| add self-conditioning (ablation) | UniAug · Graph classification | accuracy: 71.11 ± 7.50 -> 71.50 ± 5.85 (Enzymes) | Helped |
| add diffusion guidance (ablation) | UniAug · Graph classification | accuracy: 62.17 ± 3.93 -> 71.50 ± 5.85 (Enzymes) | Helped |
| use in-domain guidance instead of cross-domain guidance (ablation) | UniAug · Graph classification | accuracy: 51.50 ± 7.64 -> 71.50 ± 5.85 (Enzymes) | Helped |
Surprising findings
- Existing graph pre-training methods (AttrMask, GraphMAE, etc.) frequently underperform a plain GIN/GCN baseline, i.e. negative transfer, when pre-trained cross-domain.
- Removing diffusion guidance causes large drops on graph classification (e.g., IMDB-B 71.10 to 53.80 when guidance removed), and cross-domain guidance hurts Enzymes badly (51.50).
Limitations
- No investigation of model size effects due to limited resources; the diffusion model is relatively small.
- Guidance objectives are heuristic choices; the paper does not find a single winning objective for link prediction.
- Results depend on sampling cost; fast sampling for discrete graph diffusion is left as future work.
Experimental setup
Datasets: Network Repository GitHub Star (TUDataset subset) Cora Citeseer Pubmed Power Yeast Erdos Flickr DD Enzymes Proteins
Benchmarks: DD Enzymes Proteins NCI1 IMDB-B IMDB-M Reddit-B Reddit-12K Collab ogbg-Lipo ogbg-ESOL ogbg-FreeSolv
Metrics: accuracy MRR Hits@10 MAE average rank standard deviation across homophily groups
Baselines: GIN GCN NCN AttrMask CtxtPred EdgePred InfoMax JOAO D-SLA GraphMAE S-Mixup GraphAug
Builds on: D3PM NOS DiGress Graph Transformer DCT
Closest papers by meaning
- Generative Modelling of Structurally Constrained Graphs (NeurIPS 2024 · similarity 0.89)
- ARDiff: Anisotropic Residual Diffusion for Heterogeneous Graph Learning (AAAI 2026 · similarity 0.89)
- SBGD: Improving Graph Diffusion Generative Model via Stochastic Block Diffusion (ICML 2025 · similarity 0.89)
- Directional diffusion models for graph representation learning (NeurIPS 2023 · similarity 0.89)
- Diffusion-Guided Graph Data Augmentation (NeurIPS 2025 · similarity 0.88)
- Unifying Generation and Prediction on Graphs with Latent Graph Diffusion (NeurIPS 2024 · similarity 0.88)
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.