Research map / Graph learning / GNN architectures
HeterGP: Bridging Heterogeneity in Graph Neural Networks with Multi-View Prompting
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 HeterGP, a graph prompt-tuning framework that uses random-walk-based BFS/DFS subgraphs and two prompt graphs (homophilic and heterophilic views) to adapt frozen pre-trained GNNs to node and graph classification in few-shot settings.
Problem
Existing graph prompt-tuning methods assume graph homophily in both pre-training and prompt design, so they perform poorly on highly heterophilic graphs for few-shot multi-task node and graph classification.
Approach
Use Node2Vec-style random walks with BFS-like (homophilic) and DFS-like (heterophilic) biases to build two subgraph views, attach a separately structured prompt graph to each view (similarity-threshold links for homophily vs. dissimilarity-threshold links for heterophily), encode both with a frozen pre-trained GNN, and fuse the representations with an attention-based classifier.
Key results
- On 5-shot Texas node classification, HeterGP reaches 66.67 ± 4.78 vs. 61.29 ± 3.99 for All-in-One and 65.00 ± 4.88 for GT.
- On 100-shot CiteSeer, HeterGP reaches 84.54 ± 0.65 vs. 80.62 ± 0.36 for All-in-One, while on 100-shot PubMed it trails GCN (78.90) with 66.71.
- The paper reports HeterGP achieves best performance on 19 of 25 NC/GC settings, though the headline GC gains are mostly small (e.g., 20-shot MUTAG 73.33 vs. GraphPrompt 74.67 is not best).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HeterGP's multi-view prompting instead of the compared methods | Graph classification models · ENZYMES, 5-shot | accuracy | Helped |
| Use HeterGP's multi-view prompting instead of the compared methods | Node classification models · heterophilic datasets, few-shot settings | accuracy | Helped |
| Add the homophilic-view prompt (ablation) | HeterGP · CiteSeer and Texas, 10-shot | accuracy | Mixed |
| Add the heterophilic-view prompt (ablation) | HeterGP · CiteSeer and Texas, 10-shot | accuracy | Mixed |
| Add both homophilic- and heterophilic-view prompts (ablation) | HeterGP · PROTEINS, 10-shot | accuracy | Helped |
| Train the multi-view prompt tokens instead of using random, non-trainable tokens (ablation) | HeterGP · CiteSeer, Texas, and PROTEINS, 10-shot | accuracy | Helped |
| Use random-walk subgraph extraction instead of K-hop subgraphs (ablation) | HeterGP · node classification and graph classification | accuracy | Helped |
| Increase the number of prompt tokens | HeterGP · Texas, Actor, and ENZYMES | accuracy | No clear effect |
| Increase the number of few-shot samples | HeterGP · node classification | accuracy | Helped |
| Increase the number of few-shot samples | HeterGP · graph classification | accuracy | Hurt |
| Use HeterGP with a GCN pre-trained backbone | GraphCL pre-training plus prompting · CiteSeer, Texas, and Actor, 5-shot | accuracy | Helped |
Limitations
- Gains on graph classification are small and often within standard deviations; on 20-shot MUTAG and PubMed 100-shot HeterGP is not best.
- Method is evaluated only on node and graph classification in few-shot settings; no evaluation on link prediction or larger-scale graphs.
- Hyperparameters such as random walk parameters, thresholds, and prompt count are fixed by hand with limited sensitivity analysis.
Experimental setup
Datasets: Cora CiteSeer PubMed Texas Wisconsin Cornell Actor Chameleon Squirrel ENZYMES PROTEINS MUTAG
Benchmarks: Cora CiteSeer PubMed Texas Wisconsin Cornell Actor Chameleon Squirrel ENZYMES PROTEINS MUTAG
Metrics: accuracy (mean ± std)
Baselines: GCN GAT GT GraphCL+FT simGRACE+FT GraphPrompt All-in-One
Builds on: All-in-One GraphPrompt GPPT RAWGNN Node2Vec
Closest papers by meaning
- HGPrompt: Bridging Homogeneous and Heterogeneous Graphs for Few-Shot Prompt Learning (AAAI 2024 · similarity 0.92)
- Edge Prompt Tuning for Graph Neural Networks (ICLR 2025 · similarity 0.92)
- HG-Adapter: Improving Pre-Trained Heterogeneous Graph Neural Networks with Dual Adapters (ICLR 2025 · similarity 0.91)
- HePa: Heterogeneous Graph Prompting for All-Level Classification Tasks (AAAI 2025 · similarity 0.91)
- GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks (NeurIPS 2025 · similarity 0.91)
- One Prompt Fits All: Universal Graph Adaptation for Pretrained Models (NeurIPS 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.