Research map / Graph learning / Graph self-supervised learning
GraphBridge: Towards Arbitrary Transfer Learning in GNNs
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 GraphBridge, a side-tuning framework (GSST and GMST) that adapts pre-trained GNNs to arbitrary graph tasks and domains (graph, node, point cloud) with input/output bridges and a small trainable side network.
Problem
Pre-trained GNNs are typically tied to a single domain and task format, so transferring them to tasks with different input/output dimensions and across large domain gaps suffers from heterogeneity and negative transfer.
Approach
Wrap any frozen pre-trained GNN with input adapters, task-specific prediction heads, and a trainable side network; GSST fuses layer-wise activations with a random MLP side network, while GMST additionally merges the pre-trained backbone with a randomly initialized backup model to dilute source bias.
Key results
- On Graph2Graph transfer (GraphCL pre-training), GSST reaches 70.9 average ROC-AUC vs 70.3 for full fine-tuning, a 0.6% improvement.
- On Node2Node transfer with GCN and GraphCL, GMST reaches 73.1% on Cora vs 74.4% for fine-tuning, while on PubMed with GIN it outperforms scratch training by 6.8%.
- GraphBridge tunes only about 5%-20% of parameters, and on Graph2Node Cora GMST outperforms fine-tuning by 5-10% across pre-training methods and backbones.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GSST instead of fine-tuning | GraphCL-pretrained GIN · Graph2Graph molecular benchmarks (average) | ROC-AUC (%): 70.3 -> 70.9 | Helped |
| Use GSST instead of fine-tuning | SimGRACE-pretrained GIN · Graph2Graph molecular benchmarks (average) | ROC-AUC (%): 66.3 -> 66.4 | No clear effect |
| Use GMST instead of training from scratch | GraphCL-pretrained GIN · CiteSeer | accuracy (%): 55.10 -> 58.80 | Helped |
| Use GMST instead of fine-tuning | GraphCL-pretrained GIN · PubMed | accuracy (%): 67.30 -> 72.60 | Helped |
| Use GMST instead of fine-tuning | GraphCL-pretrained GCN · Cora, Node2Node | accuracy (%): 74.40 -> 73.10 | Hurt |
| Use GMST instead of fine-tuning | GraphCL-pretrained GCN · CiteSeer, Graph2Node | accuracy (%): 52.60 -> 61.90 | Helped |
| Use GSST instead of fine-tuning | GraphCL-pretrained GCN · ModelNet10 transfer from ogbg-molhiv | accuracy (%): 78.1 er{\color{gray}±1.7} -> 81.7 er{\color{gray}±2.9} | Helped |
| Add GSST tuning rather than direct inference (ablation) | GraphCL-pretrained model · Cora, node-level pre-training, GCN | accuracy (%): 13.80 -> 63.30 | Helped |
| Add GMST tuning rather than direct inference (ablation) | GraphCL-pretrained model · Cora, node-level pre-training, GCN | accuracy (%): 13.80 -> 73.10 | Helped |
| Add GMST tuning rather than direct inference (ablation) | GraphCL-pretrained model · Cora, graph-level pre-training, GCN | accuracy (%): 13.50 -> 74.80 | Helped |
| Add GSST tuning rather than direct inference (ablation) | GraphCL-pretrained model · Cora, graph-level pre-training, GCN | accuracy (%): 13.50 -> 52.20 | Helped |
Surprising findings
- GSST outperformed fine-tuning and GMST on Graph2PtCld when pre-trained on molhiv, which the authors attribute to the similarity between point-cloud classification and graph classification.
- Simplified side-tuning variants GBST and GAST underperform the final GSST/GMST and in hard tasks even fall below adapter and fine-tuning baselines.
Limitations
- Does not surpass full fine-tuning in all settings, especially on Amazon and Flickr.
- Evaluated on a limited set of 16 datasets and only a preliminary exploration of arbitrary transfer.
- Authors note the framework's scope is not yet extended to more benchmarks.
Experimental setup
Datasets: ZINC-full BACE BBBP ClinTox HIV SIDER Tox21 MUV ToxCast ogbn-arxiv Cora CiteSeer
Benchmarks: BACE BBBP ClinTox HIV SIDER Tox21 MUV ToxCast Cora CiteSeer PubMed Amazon-Computers
Metrics: ROC-AUC test accuracy
Baselines: Full fine-tuning MetaGP MetaFP Adapter-GNN GBST GAST
Builds on: side-tuning GraphCL SimGRACE MetaFP Adapter-GNN
Closest papers by meaning
- How Much Can Transfer? BRIDGE: Bounded Multi-Domain Graph Foundation Model with Generalization Guarantees (ICML 2025 · similarity 0.90)
- Contextual Structure Knowledge Transfer for Graph Neural Networks (AAAI 2025 · similarity 0.89)
- G-Adapter: Towards Structure-Aware Parameter-Efficient Transfer Learning for Graph Transformer Networks (AAAI 2024 · similarity 0.87)
- Unified Graph Neural Networks Pre-training for Multi-domain Graphs (AAAI 2025 · similarity 0.87)
- TINED: GNNs-to-MLPs by Teacher Injection and Dirichlet Energy Distillation (ICML 2025 · similarity 0.86)
- Fantastic Gains and Where to Find Them: On the Existence and Prospect of General Knowledge Transfer between Any Pretrained Model (ICLR 2024 · similarity 0.86)
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.