Research map / Graph learning / GNN architectures
HG-Adapter: Improving Pre-Trained Heterogeneous Graph Neural Networks with Dual Adapters
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 HG-Adapter, a dual structure-aware adapter framework with label-propagated contrastive and self-supervised losses that tunes frozen pre-trained heterogeneous GNNs, with a generalization-bound analysis.
Problem
Prompt-tuning of pre-trained heterogeneous GNNs ignores graph structure and suffers from limited labeled data, causing high training error and a large generalization gap on downstream node tasks.
Approach
Insert two adapters (homogeneous and heterogeneous) that learn task-related graph structures adaptively (kNN cosine similarity for same-type edges, attention-style neighbor-type weights for cross-type edges), plus label propagation over unlabeled nodes and three losses (subgraph-similarity contrastive, feature reconstruction, margin) to treat unlabeled nodes as pseudo-labeled supervision.
Key results
- On node classification with HERO backbone, HG-Adapter reaches 92.7 Micro-F1 on ACM and 78.3 Macro-F1 on Aminer, the best among compared methods.
- Averaged over HDMI, HeCo and HERO, HG-Adapter improves node classification by 0.81% over the best prompt-tuning baseline HetGPT.
- For node clustering, HG-Adapter improves on average by 2.69% over HetGPT across pre-trained HGNN models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add HG-Adapter to HERO | HERO · ACM, Yelp, DBLP, and Aminer | Macro-F1: 92.4$\pm$0.4 -> 92.7$\pm$0.4; 92.6$\pm$0.5 -> 93.1$\pm$0.6; 93.8$\pm$0.4 -> 94.0$\pm$0.7; 75.7$\pm$0.6 -> 78.3$\pm$0.5 | Helped |
| Add HG-Adapter to HERO | HERO · ACM, Yelp, DBLP, and Aminer | Micro-F1: 92.2$\pm$0.3 -> 92.7$\pm$0.7; 92.4$\pm$0.7 -> 92.7$\pm$0.5; 94.5$\pm$0.3 -> 94.7$\pm$0.8; 85.2$\pm$0.8 -> 87.1$\pm$0.6 | Helped |
| Add L_con, L_rec, and L_mar together (ablation) | HG-Adapter · ACM | Macro-F1: 90.1$\pm$0.4 -> 92.7$\pm$0.4 | Helped |
| Add L_con (ablation) | HG-Adapter with L_rec and L_mar · ACM | Macro-F1: 32.3$\pm$0.5 -> 92.7$\pm$0.4 | Helped |
| Add L_rec (ablation) | HG-Adapter with L_con and L_mar · ACM | Macro-F1: 89.6$\pm$0.5 -> 92.7$\pm$0.4 | Helped |
| Add L_mar (ablation) | HG-Adapter with L_con and L_rec · ACM | Macro-F1: 90.1$\pm$0.4 -> 92.7$\pm$0.4 | Helped |
Surprising findings
- Removing the contrastive loss L_con causes a large drop (e.g., ACM Macro-F1 from 92.7 to 30.2), indicating label information is essential to the method.
Limitations
- Theoretical bounds rely on a finite hypothesis space argument and heuristic assumptions (e.g., parameter-count-based U-shape), so the bound comparison is largely qualitative.
- Evaluated only on four heterogeneous graph datasets with node-level tasks.
Experimental setup
Datasets: ACM Yelp DBLP Aminer
Benchmarks: ACM Yelp DBLP Aminer
Metrics: Macro-F1 Micro-F1 NMI ARI
Baselines: HAN HGT DMGI HDMI HeCo HGCML HGMAE HERO HGPrompt HetGPT
Builds on: HGPrompt HetGPT HERO HDMI GraphPrompt
Closest papers by meaning
- HeterGP: Bridging Heterogeneity in Graph Neural Networks with Multi-View Prompting (AAAI 2025 · similarity 0.91)
- AdapterGNN: Parameter-Efficient Fine-Tuning Improves Generalization in GNNs (AAAI 2024 · similarity 0.91)
- HGPrompt: Bridging Homogeneous and Heterogeneous Graphs for Few-Shot Prompt Learning (AAAI 2024 · similarity 0.91)
- Edge Prompt Tuning for Graph Neural Networks (ICLR 2025 · similarity 0.89)
- HGOT: Self-supervised Heterogeneous Graph Neural Network with Optimal Transport (ICML 2025 · similarity 0.88)
- HePa: Heterogeneous Graph Prompting for All-Level Classification Tasks (AAAI 2025 · 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.