Research map / Graph learning / GNN architectures
HePa: Heterogeneous Graph Prompting for All-Level Classification Tasks
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 HePa, a heterogeneous graph prompting framework that uses a unified prompt template-graph with prototype nodes and a learnable pre-prompt token to enable in-context learning for node, edge, and graph classification in few-shot settings.
Problem
Heterogeneous graph prompt learning is limited to the "prompt as tokens" paradigm, cannot handle all levels (node, edge, graph) of classification downstream, and randomly initialized downstream prompts give incomplete task representations.
Approach
Builds HGNN-compatible prompt template-graphs with a virtual prompt node and class prototype nodes (computed from k-shot support subgraphs), pretrains jointly on all-level tasks, and optimizes a pre-prompt token per task level during upstream training so downstream tasks get task information via in-context learning without fine-tuning.
Key results
- On 5-shot node classification, HePa reaches 54.53% on IMDB, 78.81% on DBLP and 74.62% on ACM, outperforming HGPrompt (51.73, 75.53, 69.73).
- On few-shot edge classification, HePa achieves 53.39% (IMDB), 67.66% (DBLP), 57.21% (ACM), above HAN and MAGNN.
- Ablation shows removing the pre-prompt token and prototype computation each lowers accuracy (e.g., IMDB node-level 54.53 vs 51.82 without prototype strategy).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the full HePa prompt mechanism, including a separate prompt node (ablation) | HePa · IMDB node classification | accuracy: 47.16 -> 54.53 | Helped |
| Add task information using the Pre-prompt token (ablation) | HePa · IMDB node classification | accuracy: 50.72 -> 54.53 | Helped |
| Use the computed prototype feature instead of average labeled-node features (ablation) | HePa · IMDB node classification | accuracy: 51.82 -> 54.53 | Helped |
| Add the full HePa prompt mechanism, including a separate prompt node (ablation) | HePa · ACM graph classification | accuracy: 30.81 -> 40.51 | Helped |
| Add task information using the Pre-prompt token (ablation) | HePa · ACM edge classification | accuracy: 42.38 -> 57.21 | Helped |
| Use the computed prototype feature instead of average labeled-node features (ablation) | HePa · ACM edge classification | accuracy: 52.19 -> 57.21 | Helped |
| Use HePa rather than HGPrompt | Few-shot node classification · IMDB | accuracy: 51.73± 7.86 -> 54.53± 6.82 | Helped |
| Use HePa rather than HGPrompt | Few-shot graph classification · DBLP | accuracy: 56.49±11.68 -> 61.08± 8.13 | Helped |
| Use HePa rather than MAGNN | Few-shot edge classification · ACM | accuracy: 45.36± 1.78 -> 57.21± 6.05 | Helped |
Surprising findings
- Training only on graph-level data still yields strong edge classification on IMDB (51.85%), close to joint training results.
Limitations
- Evaluated on only three heterogeneous benchmark datasets (IMDB, DBLP, ACM).
- Absolute accuracies remain modest (e.g., ~54% on IMDB node classification) with high variance across tasks.
- Pre-prompt tokens are optimized per task level, so downstream task type must be inferred from the number of target nodes.
Experimental setup
Datasets: IMDB DBLP ACM
Benchmarks: IMDB DBLP ACM
Metrics: accuracy
Baselines: GAT GCN HAN MAGNN GraphCL HeCo GPPT GraphPrompt HGPrompt
Builds on: GraphPrompt HGPrompt PRODIGY One-for-All R-GCN
Closest papers by meaning
- HGPrompt: Bridging Homogeneous and Heterogeneous Graphs for Few-Shot Prompt Learning (AAAI 2024 · similarity 0.92)
- HeterGP: Bridging Heterogeneity in Graph Neural Networks with Multi-View Prompting (AAAI 2025 · similarity 0.91)
- GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks (NeurIPS 2025 · similarity 0.88)
- HG-Adapter: Improving Pre-Trained Heterogeneous Graph Neural Networks with Dual Adapters (ICLR 2025 · similarity 0.88)
- Edge Prompt Tuning for Graph Neural Networks (ICLR 2025 · similarity 0.88)
- HGOT: Self-supervised Heterogeneous Graph Neural Network with Optimal Transport (ICML 2025 · similarity 0.87)
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.