Research map / Graph learning / Graph self-supervised learning
GRAVER: Generative Graph Vocabularies for Robust Graph Foundation Models Fine-tuning
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.
Proposes GRAVER, a graph foundation model that pre-trains with disentangled ego-graph vocabularies and fine-tunes few-shot prompts after augmenting support samples with graphon-generated vocabularies routed by an MoE-CoE network.
Problem
Few-shot prompt fine-tuning of graph foundation models is unstable and inefficient because of random support set selection and structural mismatch between pre-training and target graphs.
Approach
Extracts factor-aware ego-graph subgraph vocabularies, models each class's vocabularies with structure and feature graphons to generate new vocabularies, and uses a hierarchical MoE-CoE router to compose them into in-context augmentations of the few-shot support samples.
Key results
- One-shot node classification: GRAVER reaches 48.00% on Cora (cross-dataset) with std 2.52, above SAMGPT's 46.79%.
- One-shot graph classification on ogbn-arXiv (cross-domain): 67.10% for GRAVER versus 62.90% for the runner-up GCOPE.
- Paper reports average std reduction of 54.0% (node) and 54.6% (graph) relative to baselines and average gains of 2.8% (node) and 3.2% (graph) over the runner-up.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GRAVER instead of the runner-up baseline | Few-shot node classification · Cross-dataset and cross-domain settings | accuracy: 2.8% | Helped |
| Use GRAVER instead of the runner-up baseline | Few-shot graph classification · Cross-dataset and cross-domain settings | accuracy: 3.2% | Helped |
| Use generative vocabulary augmentation | GRAVER · One-shot node classification | standard deviation: 54.0% | Helped |
| Use generative vocabulary augmentation | GRAVER · One-shot graph classification | standard deviation: 54.6% | Helped |
| Add semantic-independence promotion (SIP) (ablation) | GRAVER · Few-shot classification | accuracy | Helped |
| Add vocabulary augmentation (VA) (ablation) | GRAVER · Few-shot classification | accuracy | Helped |
| Use MoE-CoE selective routing instead of uniform fusion (ablation) | GRAVER · Few-shot classification | accuracy | Helped |
| Add raw text to node features | GRAVER · Zero-shot transfer | accuracy | Helped |
| Add LLM-generated class-aware descriptions to features | GRAVER · Low-shot transfer | accuracy | Helped |
Limitations
- Fine-tuning incurs higher GPU memory than baselines due to richer vocabulary tokens.
- Space complexity is not analyzed theoretically.
- Theoretical bounds rely on Lipschitz and Gaussian-type assumptions; the slack function is only loosely justified.
Experimental setup
Datasets: Cora CiteSeer PubMed ogbn-arXiv ogbn-Tech ogbn-Home Wiki-CS
Benchmarks: Cora CiteSeer PubMed ogbn-arXiv ogbn-Tech ogbn-Home Wiki-CS
Metrics: accuracy standard deviation over 20 runs
Baselines: GCN GAT GCC DGI InfoGraph GraphCL DSSL GraphACL GPPT GraphPrompt GPF ProNoG
Builds on: GFT G-Mixup GraphPrompt GPF InfoNCE
Closest papers by meaning
- Fine-Tuning Graph Neural Networks by Preserving Graph Generative Patterns (AAAI 2024 · similarity 0.89)
- GOFA: A Generative One-For-All Model for Joint Graph Language Modeling (ICLR 2025 · similarity 0.89)
- Towards Graph Foundation Models: Learning Generalities Across Graphs via Task-Trees (ICML 2025 · similarity 0.89)
- Multi-Domain Graph Foundation Models: Robust Knowledge Transfer via Topology Alignment (ICML 2025 · similarity 0.88)
- Position: Graph Foundation Models Are Already Here (ICML 2024 · similarity 0.88)
- Generative Graph Pattern Machine (NeurIPS 2025 · 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.