Research map / Graph learning / GNN architectures
Edge Prompt Tuning for Graph Neural Networks
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 EdgePrompt and EdgePrompt+, graph prompt tuning that learns prompt vectors on edges (shared or attention-weighted over anchor prompts) inserted via message passing into frozen pre-trained GNNs, evaluated on node and graph classification under four pre-training strategies.
Problem
Objective gaps between self-supervised GNN pre-training and downstream tasks; existing graph prompt tuning methods place prompts on node features or representations and ignore graph structure/edges.
Approach
Learn prompt vectors on edges at each GNN layer that are aggregated during message passing; EdgePrompt uses one shared edge prompt, while EdgePrompt+ computes per-edge prompts as attention-weighted combinations of a set of anchor prompts, keeping the pre-trained GNN frozen.
Key results
- EdgePrompt+ achieves best accuracy on 5-shot node classification under GraphCL on Cora (62.88 vs 58.52 for GPF) and CiteSeer (46.20 vs 43.55 for GPF).
- Under EP-GPPT pre-training, EdgePrompt+ reaches 56.41 on Cora 5-shot versus 41.28 for the best baseline GPPT.
- On 50-shot graph classification, EdgePrompt+ is best or runner-up in every setting and reports gains over GPF-plus, e.g. NCI1 66.30 vs 64.56 under GraphCL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use customized edge prompts (EdgePrompt+) instead of shared global edge prompts (EdgePrompt) | EdgePrompt node classification · 5-shot node classification across five datasets and four pre-training strategies | accuracy | Mixed |
| Use customized edge prompts (EdgePrompt+) instead of shared global edge prompts (EdgePrompt) | EdgePrompt graph classification · 50-shot graph classification across five datasets and four pre-training strategies | accuracy | Helped |
| Restrict edge prompts to the first GNN layer instead of learning them at each layer (ablation) | EdgePrompt+ · CiteSeer 5-shot node classification with GraphCL | accuracy: 46.96_{\pm 2.63} -> 46.20_{\pm 0.99} | Hurt |
| Restrict edge prompts to the first GNN layer instead of learning them at each layer (ablation) | EdgePrompt+ · ENZYMES 50-shot graph classification with SimGRACE | accuracy: 28.58_{\pm 2.45} -> 32.67_{\pm 2.53} | Hurt |
| Use one anchor prompt per layer, reducing EdgePrompt+ to EdgePrompt (ablation) | EdgePrompt+ · Node classification and graph classification | accuracy | Hurt |
| Increase the number of anchor prompts to 50 (ablation) | EdgePrompt+ · Node classification and graph classification | accuracy | No clear effect |
Surprising findings
- The shared-prompt EdgePrompt gives only small gains over GPF, while the per-edge EdgePrompt+ is needed for most improvements.
- Learning edge prompts only at the first layer degrades performance, especially for EdgePrompt+ (e.g. 56.41 to 56.10 vs 37.26 to 36.74 on Cora under EP-GPPT).
Limitations
- Score function phi is limited to one attention form; other variants are left for future work.
- Theoretical node-classification analysis relies on a contextual stochastic block model with simplifying assumptions.
- Edge prompts for graph classification assume the GNN has no edge attribute support; experiments on graphs with edge features are limited to two datasets.
Experimental setup
Datasets: Cora CiteSeer Pubmed ogbn-arxiv Flickr ENZYMES DD NCI1 NCI109 Mutagenicity BACE BBBP
Benchmarks: Cora CiteSeer Pubmed ogbn-arxiv Flickr ENZYMES DD NCI1 NCI109 Mutagenicity BACE BBBP
Metrics: accuracy
Baselines: Classifier Only GPPT GraphPrompt ALL-in-one GPF GPF-plus
Builds on: GPF GPF-plus GraphPrompt GPPT All-in-one
Closest papers by meaning
- GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks (NeurIPS 2025 · similarity 0.93)
- HeterGP: Bridging Heterogeneity in Graph Neural Networks with Multi-View Prompting (AAAI 2025 · similarity 0.92)
- ProG: A Graph Prompt Learning Benchmark (NeurIPS 2024 · similarity 0.91)
- HGPrompt: Bridging Homogeneous and Heterogeneous Graphs for Few-Shot Prompt Learning (AAAI 2024 · similarity 0.91)
- One Prompt Fits All: Universal Graph Adaptation for Pretrained Models (NeurIPS 2025 · similarity 0.90)
- HG-Adapter: Improving Pre-Trained Heterogeneous Graph Neural Networks with Dual Adapters (ICLR 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.