Research map / Graph learning / Graph self-supervised learning
Multi-Domain Graph Foundation Models: Robust Knowledge Transfer via Topology Alignment
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 MDGFM, a multi-domain graph foundation model that uses domain/shared/balance tokens, graph structure learning with kNN refinement for topology alignment, and dual-prompt tuning to transfer to unseen domains.
Problem
Multi-domain graph pre-training suffers from large topology differences across domains (homophilic vs heterophilic), sparse and noisy edges, and adversarial attacks, which hinder robust transfer to unseen graph domains.
Approach
Decouples feature and topology via domain, shared, and balance tokens; refines each source and target graph with a structure-learning module (kNN over features plus aggregated structure) to align topology; pre-trains a GCN contrastively with the refined graph as extra positives; and transfers via a dual prompt (meta prompt from source-domain token combinations plus a specific prompt) with frozen e
Key results
- In one-shot node classification, MDGFM reaches 44.83% on Cora and 40.77% on Cornell, the best among all baselines on all seven datasets.
- On Penn94 with K=500, MDGFM achieves 63.25% versus 52.88% for GCOPE and 51.52% for MDGPT.
- MDGFM outperforms the second-best model by up to 18.04% in one-shot and 8.11% in few-shot settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MDGFM instead of the strongest reported baseline | few-shot node classification · Cora, Citeseer, Pubmed, Cornell, Squirrel, Chameleon, and Penn94 (1-shot) | accuracy | Helped |
| Use MDGFM instead of the strongest reported baseline | few-shot node classification · Cora, Citeseer, Pubmed, Cornell, Squirrel, Chameleon, and Penn94 (3/5-shot) | accuracy | Helped |
| Increase the number of labeled examples from 10 to 500 | MDGFM on Penn94 · Penn94 | accuracy: 54.43±0.93 -> 63.25±0.82 | Helped |
| Remove the shared token (ablation) | MDGFM · Most datasets | accuracy | Hurt |
| Remove the graph structure learning module and refined adjacency (ablation) | MDGFM · Few-shot node classification datasets | accuracy | Hurt |
| Use features only in graph structure learning, without topology information (ablation) | MDGFM · Few-shot node classification datasets | accuracy | Hurt |
| Remove the balance token (ablation) | MDGFM · Few-shot node classification datasets | accuracy | Hurt |
| Remove Pubmed from the source domains (ablation) | MDGFM · Cornell 1-shot | accuracy: 40.77±5.96 -> 39.66±5.51 | No clear effect |
| Remove Squirrel from the source domains (ablation) | MDGFM · Cornell 1-shot | accuracy: 40.77±5.96 -> 40.78±5.44 | No clear effect |
| Remove Cora and Citeseer from the source domains (ablation) | MDGFM · Cornell 1-shot | accuracy: 40.77±5.96 -> 39.96±5.99 | No clear effect |
| Remove Pubmed and Squirrel from the source domains (ablation) | MDGFM · Cornell 1-shot | accuracy: 40.77±5.96 -> 42.68±5.64 | Helped |
| Remove Citeseer, Cora, and Pubmed from the source domains (ablation) | MDGFM · Cornell 1-shot | accuracy: 40.77±5.96 -> 40.08±5.74 | No clear effect |
Surprising findings
- Removing some source domains sometimes improves performance, which the authors attribute to conflicts or interference among multi-domain graphs.
- Supervised GCN/GAT occasionally outperform graph pre-training methods, suggesting negative transfer.
Limitations
- Evaluated only on node classification with a GCN backbone and small-to-medium graphs.
- Theoretical guarantees rely on covariate-shift and invariance assumptions that are loosely connected to the implemented method.
- Relies on fixed-size PCA projection and kNN hyperparameters tuned per dataset.
Experimental setup
Datasets: Cora Citeseer Pubmed Cornell Chameleon Squirrel Penn94
Benchmarks: Cora Citeseer Pubmed Cornell Squirrel Chameleon Penn94
Metrics: accuracy
Baselines: GCN GAT DGI GraphCL GPPT GPF GCOPE MDGPT SAMGPT
Builds on: GCOPE MDGPT GPF GraphCL DGI
Closest papers by meaning
- How Much Can Transfer? BRIDGE: Bounded Multi-Domain Graph Foundation Model with Generalization Guarantees (ICML 2025 · similarity 0.90)
- GRAVER: Generative Graph Vocabularies for Robust Graph Foundation Models Fine-tuning (NeurIPS 2025 · similarity 0.88)
- Unified Graph Neural Networks Pre-training for Multi-domain Graphs (AAAI 2025 · similarity 0.88)
- TFGDA: Exploring Topology and Feature Alignment in Semi-supervised Graph Domain Adaptation through Robust Clustering (NeurIPS 2024 · similarity 0.88)
- Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement (NeurIPS 2025 · similarity 0.88)
- Contextual Structure Knowledge Transfer for Graph Neural Networks (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.