Research map / Graph learning / GNN architectures
RAGraph: A General Retrieval-Augmented Graph Learning Framework
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 RAGraph, which builds a key-value vector library of toy graphs from resource graphs and retrieves similar ones at inference to inject features and labels into a frozen pre-trained GNN via message-passing prompts, with tuning-free and tuned variants.
Problem
GNNs generalize poorly to unseen graph data that differs from training instances, and existing in-context/prompt approaches (e.g., PRODIGY, GraphPrompt) do not integrate retrieved external features and labels for dynamic, cross-task settings.
Approach
Retrieval-augmented graph learning: a toy graph vector base keyed by timestamp, hidden embedding, neighbor environment, and structural position code; retrieved toy graphs' hidden embeddings and task outputs are propagated via parameter-free or learnable message passing into the query node, with noisy-retrieval prompt tuning.
Key results
- On ENZYMES node classification (5-shot), RAGraph/NFT reaches 76.23% vs. 57.98% for PRODIGY/FT (best baseline).
- On COX2 graph classification, RAGraph/FT reaches 76.60% vs. 65.31% for PRODIGY/FT.
- On link prediction, RAGraph/NFT achieves 24.89 Recall on TAOBAO vs. 23.74 for PRODIGY/FT (per-mille scale).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use retrieval augmentation rather than Vanilla/FT | RAGraph · PROTEINS node classification (5-shot) | accuracy: 48.99$\scriptstyle{\pm 01.88}$ -> 59.83$\scriptstyle{\pm 00.40}$ | Helped |
| Use retrieval augmentation rather than Vanilla/FT | RAGraph · ENZYMES node classification (5-shot) | accuracy: 51.99$\scriptstyle{\pm 1.36}$ -> 76.23$\scriptstyle{\pm 01.63}$ | Helped |
| Use retrieval augmentation rather than Vanilla/FT | RAGraph · PROTEINS graph classification (5-shot) | accuracy: 57.04$\scriptstyle{\pm 03.88}$ -> 62.33$\scriptstyle{\pm 02.52}$ | Helped |
| Use retrieval augmentation rather than Vanilla/FT | RAGraph · COX2 graph classification (5-shot) | accuracy: 64.04$\scriptstyle{\pm 08.20}$ -> 76.60$\scriptstyle{\pm 02.30}$ | Helped |
| Use retrieval augmentation rather than Vanilla/FT | RAGraph · ENZYMES graph classification (5-shot) | accuracy: 40.00$\scriptstyle{\pm 04.36}$ -> 49.17$\scriptstyle{\pm 04.64}$ | Helped |
| Use retrieval augmentation rather than Vanilla/FT | RAGraph · BZR graph classification (5-shot) | accuracy: 69.01$\scriptstyle{\pm 02.21}$ -> 77.53$\scriptstyle{\pm 05.26}$ | Helped |
| Use retrieval augmentation rather than Vanilla/FT | RAGraph · TAOBAO link prediction | Recall@20: 23.99$\scriptstyle{\pm 02.11}$ -> 24.89$\scriptstyle{\pm 02.01}$ | Helped |
| Use retrieval augmentation rather than Vanilla/FT | RAGraph · KOUBEI link prediction | Recall@20: 33.96$\scriptstyle{\pm 04.13}$ -> 34.27$\scriptstyle{\pm 03.93}$ | Helped |
| Use retrieval augmentation rather than Vanilla/FT | RAGraph · AMAZON link prediction | Recall@20: 18.14$\scriptstyle{\pm 07.55}$ -> 18.69$\scriptstyle{\pm 07.45}$ | Helped |
| Add noise prompt tuning instead of prompt tuning (ablation) | RAGraph · PROTEINS node classification (5-shot) | accuracy: 58.74$\scriptstyle{\pm 00.87}$ -> 59.83$\scriptstyle{\pm 00.40}$ | Helped |
| Add noise prompt tuning instead of prompt tuning (ablation) | RAGraph · ENZYMES node classification (5-shot) | accuracy: 75.74$\scriptstyle{\pm 01.92}$ -> 76.23$\scriptstyle{\pm 01.63}$ | Helped |
| Add noise prompt tuning instead of prompt tuning (ablation) | RAGraph · PROTEINS graph classification (5-shot) | accuracy: 62.33$\scriptstyle{\pm 02.52}$ -> 59.08$\scriptstyle{\pm 03.50}$ | Hurt |
Surprising findings
- The tuning-free RAGraph/NF variant often matches or beats fine-tuned baselines, e.g. ENZYMES 75.92% for NF vs. 57.98% for PRODIGY/FT.
- Performance first rises then falls as the number of hops k and retrieved toy graphs topK increase, since excess retrieved knowledge becomes noise.
Limitations
- Currently limited to retrieving subgraphs; nodes, edges, and trees not explored.
- Sensitive to hyper-parameters k and topK; too much retrieved knowledge adds noise.
- Relies on frozen pre-trained GNNs for embeddings; a theoretical justification via mutual information is informal.
Experimental setup
Datasets: PROTEINS COX2 ENZYMES BZR TAOBAO KOUBEI AMAZON
Benchmarks: PROTEINS COX2 ENZYMES BZR TAOBAO KOUBEI AMAZON
Metrics: accuracy Recall@20 nDCG@20
Baselines: GCN GraphSAGE GAT GIN GraphPrompt PRODIGY LightGCN SGL MixGCF SimGCL GraphPro
Builds on: GraphPrompt PRODIGY GraphPro RAG P-GNN
Closest papers by meaning
- GFM-RAG: Graph Foundation Model for Retrieval Augmented Generation (NeurIPS 2025 · similarity 0.91)
- G-Retriever: Retrieval-Augmented Generation for Textual Graph Understanding and Question Answering (NeurIPS 2024 · similarity 0.89)
- Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation (ICLR 2025 · similarity 0.88)
- Can Knowledge-Graph-based Retrieval Augmented Generation Really Retrieve What You Need? (NeurIPS 2025 · similarity 0.87)
- Iteratively Refined Early Interaction Alignment for Subgraph Matching based Graph Retrieval (NeurIPS 2024 · similarity 0.86)
- Beyond the Golden Teacher: Enhancing Graph Learning through LLM-GNN Co-teaching (NeurIPS 2026 · 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.