Research map / Graph learning / GNN architectures
GraphTOP: Graph Topology-Oriented Prompting 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 GraphTOP, a topology-oriented graph prompting method that learns edge rewiring between each target node and its 2-hop subgraph via Gumbel-Softmax reparameterization, keeping a pre-trained GCN frozen.
Problem
Existing graph prompting for adapting pre-trained GNNs is feature-oriented and ignores graph topology; the paper asks how to design topology-oriented prompts for node classification.
Approach
Reformulate topology prompting as learnable edge rewiring between each target node and nodes in its multi-hop subgraph, relaxed to Bernoulli probabilities produced by a shared projector over frozen node representations and sampled with Gumbel-Softmax, with entropy and sparsity regularizers for tight, sparse relaxation.
Key results
- GraphTOP wins 17 of 20 pre-training/dataset settings in 5-shot node classification, consistently best on Amazon and Flickr across all four pre-training strategies.
- Under LP-GPPT on Amazon, GraphTOP reaches 32.41% vs 28.90% for the best baseline GPPT (runner-up).
- Efficiency: GraphTOP on Amazon with rho=2 takes 218.61s vs 651.22s for the all-node-pair variant, which runs out of GPU memory at rho=3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use topology-oriented edge rewiring with GraphTOP instead of the strongest reported baseline | pre-trained GCN node classification · 5-shot node classification across five datasets and four pre-training strategies | accuracy | Helped |
| use GraphTOP rather than GraphPrompt+ | pre-trained GCN · Cora, SimGRACE, 5-shot | accuracy: 51.26_{\pm 4.90} -> 50.57_{\pm 2.91} | Hurt |
| remove the entropy regularizer L_E (ablation) | GraphTOP · probability distribution of rewired edges | probability concentration near 0 or 1 | Hurt |
| add the entropy regularizer L_E (ablation) | GraphTOP · probability distribution of rewired edges | probability concentration near 0 or 1 | Helped |
| remove the sparsity regularizer L_S (ablation) | GraphTOP · target-node prompted subgraphs | edge density | Hurt |
| add the sparsity regularizer L_S (ablation) | GraphTOP · target-node prompted subgraphs | edge density | Helped |
| rewire edges only between each target node and other nodes, rather than all pairs in the subgraph (ablation) | GraphTOP · Cora, rho=2 | running time (seconds): 101.70 -> 36.52 | Helped |
| rewire edges only between each target node and other nodes, rather than all pairs in the subgraph (ablation) | GraphTOP · Amazon, rho=2 | running time (seconds): 651.22 -> 218.61 | Helped |
| rewire edges between every pair of nodes in the local subgraph (ablation) | GraphTOP · Cora and Amazon, rho=3 | GPU memory availability: OOM | Hurt |
| set lambda_2 to 0, removing the sparsity penalty (ablation) | GraphTOP · PubMed under GraphCL and Amazon under LP-GPPT | accuracy | Hurt |
Surprising findings
- Linear Probe is competitive and beats several graph prompting baselines on Flickr across all four pre-training strategies.
- Without the sparsity regularizer, the learned prompted graphs become nearly fully dense within the local subgraphs.
Limitations
- Theoretical analysis is limited to linear 2-layer GCN on CSBM random graphs; nonlinear separability is not analyzed.
- Empirical evaluation is restricted to node classification; graph-level extension is left as future work.
- Hyperparameters lambda1 and lambda2 require grid search, and the sparsity target gamma is fixed at 0.5.
Experimental setup
Datasets: Cora PubMed Amazon Minesweeper Flickr
Benchmarks: Cora PubMed Amazon Minesweeper Flickr
Metrics: accuracy
Baselines: Linear Probe GPPT All-in-one GraphPrompt GraphPrompt+ ProNoG
Builds on: GPPT GraphPrompt All-in-one ProNoG Gumbel-Softmax
Closest papers by meaning
- Edge Prompt Tuning for Graph Neural Networks (ICLR 2025 · similarity 0.93)
- One Prompt Fits All: Universal Graph Adaptation for Pretrained Models (NeurIPS 2025 · similarity 0.91)
- ProG: A Graph Prompt Learning Benchmark (NeurIPS 2024 · similarity 0.91)
- HeterGP: Bridging Heterogeneity in Graph Neural Networks with Multi-View Prompting (AAAI 2025 · similarity 0.91)
- Does Graph Prompt Work? A Data Operation Perspective with Theoretical Analysis (ICML 2025 · similarity 0.90)
- Spectral Reversal: Counteracting Singular Value Bias for Graph Prompting (NeurIPS 2026 · 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.