Research map / Graph learning / Graph self-supervised learning
One Prompt Fits All: Universal Graph Adaptation for Pretrained Models
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.
Argues representation-level graph prompts are equivalent to a linear probe and proposes UniPrompt, a learnable kNN-based prompt graph with bootstrapped topology integration that adapts frozen pretrained GNNs across homophilic and heterophilic graphs.
Problem
Graph prompt learning (GPL) methods show unstable performance across pretrained models and fail to adapt under distribution shifts such as homophilic-to-heterophilic graphs, and their mechanisms are unclear.
Approach
Theoretically shows representation-level prompts reduce to a linear classifier (linear probe), and proposes an input-level prompt: a kNN graph with learnable edge gates, integrated into the original topology via an iterative bootstrapped update, while a lightweight classifier is tuned jointly.
Key results
- In 1-shot DGI-pretrained in-domain node classification, UniPrompt reaches 51.13 on Cornell and 58.75 on Wisconsin versus best GPL baselines around 33-36 on those datasets.
- In 1-shot cross-domain setting with FUG, UniPrompt achieves 55.01 on PubMed and 51.58 on Cornell, outperforming MDGFM (46.84 and 40.77).
- UniPrompt is best or competitive across most heterophilic datasets, though it underperforms on Chameleon under DGI and GRACE where baselines are comparable or better.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UniPrompt instead of MDGFM | Cross-domain pretrained graph adaptation · PubMed 1-shot | accuracy: 46.84_{\pm 7.31} -> 55.01_{\pm 3.36} | Helped |
| Use UniPrompt instead of MDGFM | Cross-domain pretrained graph adaptation · Cornell 1-shot | accuracy: 40.77_{\pm 5.96} -> 51.58_{\pm 9.91} | Helped |
| Use UniPrompt instead of MDGFM | Cross-domain pretrained graph adaptation · Chameleon 1-shot | accuracy: 28.36_{\pm 3.65} -> 25.14_{\pm 5.65} | Hurt |
| Use UniPrompt instead of MDGFM | Cross-domain pretrained graph adaptation · CiteSeer 5-shot | accuracy: 61.24_{\pm 4.82} -> 59.37_{\pm 2.78} | Hurt |
| Inject a prompt graph rather than set τ=1.0 (no prompt graph injected) (ablation) | UniPrompt with DGI · 1-shot node classification; datasets vary | accuracy | Mixed |
| Vary k in the kNN prompt graph (ablation) | UniPrompt with DGI · PubMed 1-shot | accuracy | Hurt |
Surprising findings
- Representation-level graph prompts are functionally equivalent to a linear probe, so they offer little beyond fine-tuning a classifier.
- Linear probe often matches or beats existing GPL methods across pretrained models, and GPPT shows significant incompatibility when the pretrained model changes.
Limitations
- Gains are small or absent on some datasets, e.g., Chameleon under DGI and GRACE where baselines match or beat UniPrompt.
- The theoretical equivalence relies on simplifying assumptions (orthogonality and full-rank conditions) that are not fully justified in practice.
- Few-shot evaluation only; the paper does not analyze scalability to large graphs.
Experimental setup
Datasets: Cora CiteSeer PubMed Cornell Texas Wisconsin Chameleon Actor Squirrel
Benchmarks: Cora CiteSeer PubMed Cornell Texas Wisconsin Chameleon Actor Squirrel
Metrics: accuracy
Baselines: Fine-tune Linear-probe GPPT GraphPrompt All-in-one GPF GPF-plus EdgePrompt EdgePrompt-plus GCN GAT DGI
Builds on: GPPT GraphPrompt GPF EdgePrompt ProG
Closest papers by meaning
- ProG: A Graph Prompt Learning Benchmark (NeurIPS 2024 · similarity 0.91)
- GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks (NeurIPS 2025 · similarity 0.91)
- Does Graph Prompt Work? A Data Operation Perspective with Theoretical Analysis (ICML 2025 · similarity 0.90)
- Edge Prompt Tuning for Graph Neural Networks (ICLR 2025 · similarity 0.90)
- HGPrompt: Bridging Homogeneous and Heterogeneous Graphs for Few-Shot Prompt Learning (AAAI 2024 · similarity 0.90)
- HeterGP: Bridging Heterogeneity in Graph Neural Networks with Multi-View Prompting (AAAI 2025 · similarity 0.89)
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.