Research map / Graph learning / GNN architectures
ProG: A Graph Prompt Learning Benchmark
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.
Introduces ProG, the first benchmark and open-source library for graph prompt learning, comparing 6 pre-training methods and 5 graph prompt methods across 15 node and graph classification datasets in few-shot settings.
Problem
Graph prompt learning lacks a unified framework, standardized evaluation, and an easy-to-use toolkit, so existing methods are hard to compare under few-shot settings against supervised and pre-train-and-fine-tune baselines.
Approach
Decomposes graph prompts into prompt tokens, token structure, and insert patterns, grouping methods into "prompt as graph" and "prompt as token"; implements all methods in one library (ProG) and benchmarks them under a standardized few-shot protocol with random hyperparameter search.
Key results
- GPF-plus reduced the negative transfer rate for node classification from 43% to 0% with GraphMAE pre-training, per the paper's 1-shot tables.
- All-in-one reduced graph-classification negative transfer rate from 38% to 0% and achieved best 1-shot results on 7 of 8 graph datasets.
- All-in-one achieved a RED% of 95.06 in the flexibility analysis, the lowest error bound among methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use graph prompting instead of supervised training | GCN-based graph learning · 1-shot node and graph classification across 15 datasets | accuracy | Helped |
| Use graph prompting instead of pre-training followed by fine-tuning | Pre-trained GCN · 1-shot node and graph classification across 15 datasets | accuracy | Helped |
| Add GPF-plus prompting to pre-training | Pre-trained models on 1-shot node classification · Seven node datasets | negative transfer rate: 43% to 0% | Helped |
| Add All-in-one prompting to pre-training | Pre-trained models on 1-shot graph classification · Eight graph datasets | negative transfer rate: 38% to 0% | Helped |
| Use node-level pre-training with GPF-plus rather than traditional fine-tuning | Pre-trained model on node classification · Node classification tasks | accuracy | Helped |
| Use graph-level pre-training with All-in-one rather than traditional fine-tuning | Pre-trained model on graph classification · Graph classification tasks | accuracy | Helped |
| Use GPPT prompting | Pre-trained GCN on heterophilic node datasets · Wisconsin, Texas, and Actor | accuracy | Hurt |
| Use All-in-one prompting | Pre-trained model on graph classification · Pure structural graph datasets IMDB-B and COLLAB | accuracy | Mixed |
Surprising findings
- GPPT, a prompt-as-token method, underperforms classical pre-train & fine-tune on heterophilic and most node datasets, attributed to clustering instability.
- Graph prompt performance generally decreases as shot number grows, sometimes falling below supervised baselines.
Limitations
- Evaluation limited to node and graph classification tasks
- Results focus on few-shot settings; prompt advantage diminishes with more shots
- Some methods (GPPT) unstable and require task adaptations to apply across node/graph tasks
Experimental setup
Datasets: Cora Citeseer PubMed ogbn-arxiv Wisconsin Texas Actor IMDB-BINARY COLLAB PROTEINS MUTAG ENZYMES
Benchmarks: Cora Citeseer PubMed ogbn-arxiv Wisconsin Texas Actor IMDB-BINARY COLLAB PROTEINS MUTAG ENZYMES
Metrics: Accuracy Macro F1 AUROC error bound (RED%) training time tunable parameters
Baselines: Supervised GCN Pre-train & Fine-tune GPPT All-in-one Gprompt GPF GPF-plus
Builds on: GPPT All-in-one Gprompt GPF GPF-plus GraphMAE DGI GraphCL SimGRACE
Closest papers by meaning
- Does Graph Prompt Work? A Data Operation Perspective with Theoretical Analysis (ICML 2025 · similarity 0.92)
- Edge Prompt Tuning for Graph Neural Networks (ICLR 2025 · similarity 0.91)
- One Prompt Fits All: Universal Graph Adaptation for Pretrained Models (NeurIPS 2025 · similarity 0.91)
- GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks (NeurIPS 2025 · similarity 0.91)
- HeterGP: Bridging Heterogeneity in Graph Neural Networks with Multi-View Prompting (AAAI 2025 · similarity 0.89)
- HGPrompt: Bridging Homogeneous and Heterogeneous Graphs for Few-Shot Prompt Learning (AAAI 2024 · similarity 0.88)
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.