Research map / Graph learning / Graph self-supervised learning
One For All: Towards Training One Graph Model For All Classification Tasks
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 OFA, a single graph foundation model that unifies cross-domain graph classification via text-attributed graphs, NOI subgraphs with prompt nodes, and a graph prompting paradigm enabling supervised, few-shot, and zero-shot learning.
Problem
Building one graph model that handles node, link, and graph classification tasks across domains with heterogeneous features, and that supports in-context (few-shot and zero-shot) learning on graphs.
Approach
Describe all nodes and edges as natural-language text-attributed graphs embedded by one LLM into a shared space, unify tasks via nodes-of-interest subgraphs with an NOI prompt node, and perform in-context learning by appending prompt graphs with class-description nodes so every task becomes binary classification on class nodes.
Key results
- In supervised joint training, OFA-llama2-13b reaches 78.25% accuracy on PubMed node classification and 77.51% on ogbn-arxiv, matching or exceeding GCN/GAT baselines.
- OFA-joint-lr achieves 70.20% zero-shot accuracy on FB15K237 (20-way) and 56.92% zero-shot on Cora 2-way transfer.
- On the transductive ogbn-arxiv 5-way 5-shot setting, OFA-joint-lr gets 61.45% versus Prodigy's 61.09%, but TLP-SURGL is stronger at 77.89%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train one joint OFA model across datasets instead of training independently | OFA with sentence transformer features · Supervised learning across datasets | performance across benchmarks | Helped |
| Use larger Llama2-13b text embeddings instead of smaller LLM embeddings | Joint OFA · Supervised learning across datasets | performance | Mixed |
| Use Llama2 embeddings instead of sentence transformer embeddings | OFA · Citation-network benchmarks | accuracy | Helped |
| Use e5-large-v2 embeddings instead of other tested LLMs | OFA · Molecular datasets | performance | Helped |
| Apply OFA-joint-lr with graph prompting for zero-shot classification | OFA-joint-lr · Node, link, and graph transfer tasks | accuracy or AUC | Mixed |
Surprising findings
- Zero-shot OFA on HIV (35.67 AUC) is below random, and on WN18RR 5-way zero-shot is 30.96%, indicating weak zero-shot transfer in some settings.
- Larger LLM encoders do not consistently help: OFA-llama2-7b/13b underperform the sentence transformer on PCBA (21.35/21.54 vs 24.83 APR).
Limitations
- Does not handle regression tasks.
- Cross-domain dataset coverage is limited to nine datasets.
- Llama2 joint models were trained for fewer epochs due to limited compute, and zero-shot results are weak on some datasets.
Experimental setup
Datasets: Cora PubMed ogbn-arxiv Wiki-CS FB15K237 WN18RR PCBA HIV ChEMBL
Benchmarks: Cora PubMed ogbn-arxiv Wiki-CS FB15K237 WN18RR PCBA HIV ChEMBL
Metrics: accuracy AUC AP
Baselines: GCN GAT GIN Prodigy GPN TENT GLITTER TLP-SURGL Galactica-1.3B GIMLET
Builds on: R-GCN Prodigy GraphPrompt SEAL
Closest papers by meaning
- One Prompt Fits All: Universal Graph Adaptation for Pretrained Models (NeurIPS 2025 · similarity 0.88)
- UniGTE: Unified Graph–Text Encoding for Zero-Shot Generalization across Graph Tasks and Domains (NeurIPS 2025 · similarity 0.88)
- Towards Graph Foundation Models: Learning Generalities Across Graphs via Task-Trees (ICML 2025 · similarity 0.88)
- A Unified Graph Language Model for Multi-Domain Multi-Task Graph Alignment Instruction Tuning (NeurIPS 2026 · similarity 0.87)
- SSTAG: Structure-Aware Self-Supervised Learning Method for Text-Attributed Graphs (NeurIPS 2025 · similarity 0.87)
- GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks (NeurIPS 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.