Research map / Graph learning / Graph self-supervised learning
Unified Graph Neural Networks Pre-training for Multi-domain Graphs
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 MDP-GNN, a multi-domain pre-training framework that trains one unified graph transformer over text-free graphs from several domains using feature integration, bi-level pivot-node graph connection learning, and Wasserstein distribution alignment.
Problem
GNN pre-training is limited to one domain per model ("one-domain-one-model"); graphs from different domains have heterogeneous feature spaces and distributions, so a single GNN generalizes poorly across text-free graph domains.
Approach
Integrates node features across domains by concatenating each domain's real features with MLP-converted features in a shared layout, learns a domain-synthesized network via bi-level optimization of trainable pivot nodes, and aligns domain distributions to a meta-domain graph by minimizing a Wasserstein-1 distance estimated with a discriminator.
Key results
- With 5% labels on seen domains, MDP-GNN with EdgePred and prompt-tuning reaches F1 43.23 / AUC 91.55 on Academic and 83.18 / 98.90 on Reddit, versus best supervised baseline GraphSAGE 38.78 / 90.17 and 74.99 / 97.94 on Reddit.
- On unseen domains with leave-one-out, MDP-GNN (AttrMask, fine-tuning) reports F1 47.98 / AUC 92.33 on Academic and 90.97 / 99.38 on Reddit at 10% labels, outperforming the listed baselines.
- Ablation removing feature integration (replaced by SVD) yields the largest drop, e.g. Academic F1 falls from 43.23 to 34.05 under prompt-tuning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MDP-GNN pre-training and prompt-tuning instead of the strongest supervised baseline | MDP-GNN · Academic, 5% labels known | F1: 36.38 -> 43.23 | Helped |
| Use MDP-GNN pre-training and prompt-tuning instead of the strongest supervised baseline | MDP-GNN · Product, 5% labels known | F1: 47.99 -> 52.14 | Helped |
| Use MDP-GNN pre-training and prompt-tuning with AttrMask instead of the strongest supervised baseline | MDP-GNN · Reddit, 5% labels known | F1: 68.64 -> 82.77 | Helped |
| Use MDP-GNN pre-training and fine-tuning instead of the strongest supervised baseline | MDP-GNN · Yelp, 5% labels known | F1: 57.91 -> 59.02 | Helped |
| Add feature integration (rather than using SVD) (ablation) | MDP-GNN prompt-tuning · Academic, Product, Reddit, and Yelp, 5% labels known | F1: 34.05 -> 43.23; 42.68 -> 52.14; 55.23 -> 74.24; 56.97 -> 59.00 | Helped |
| Add intra-domain loss (ablation) | MDP-GNN prompt-tuning · Academic, Product, Reddit, and Yelp, 5% labels known | F1: 39.71 -> 43.23; 46.16 -> 52.14; 45.37 -> 74.24; 56.88 -> 59.00 | Helped |
| Add inter-domain loss (ablation) | MDP-GNN prompt-tuning · Academic, Product, Reddit, and Yelp, 5% labels known | F1: 42.51 -> 43.23; 50.50 -> 52.14; 68.60 -> 74.24; 58.94 -> 59.00 | Helped |
| Add Wasserstein distribution alignment loss (ablation) | MDP-GNN prompt-tuning · Academic, Product, Reddit, and Yelp, 5% labels known | F1: 39.95 -> 43.23; 50.88 -> 52.14; 56.62 -> 74.24; 57.83 -> 59.00 | Helped |
| Add feature integration (rather than using SVD) (ablation) | MDP-GNN fine-tuning · Academic, Product, Reddit, and Yelp, 5% labels known | F1: 39.80 -> 46.16; 42.86 -> 50.79; 53.22 -> 62.45; 57.79 -> 59.02 | Helped |
| Add intra-domain loss (ablation) | MDP-GNN fine-tuning · Academic, Product, Reddit, and Yelp, 5% labels known | F1: 41.35 -> 46.16; 48.62 -> 50.79; 48.98 -> 62.45; 56.59 -> 59.02 | Helped |
| Add inter-domain loss (ablation) | MDP-GNN fine-tuning · Academic, Product, Reddit, and Yelp, 5% labels known | F1: 46.06 -> 46.16; 50.77 -> 50.79; 60.72 -> 62.45; 58.46 -> 59.02 | Helped |
| Add Wasserstein distribution alignment loss (ablation) | MDP-GNN fine-tuning · Academic, Product, Reddit, and Yelp, 5% labels known | F1: 46.10 -> 46.16; 50.43 -> 50.79; 51.36 -> 62.45; 59.00 -> 59.02 | Mixed |
Surprising findings
- Prompt-tuning is more effective than fine-tuning on seen domains, while fine-tuning is recommended for unseen domains.
- Self-supervised GT pre-trained without multi-domain alignment performs worse than vanilla supervised models.
Limitations
- Evaluated only on node classification with four text-free graph datasets and a two-layer GT backbone.
- Theoretical bounds are for a binary classification setting, which simplifies the multi-class node classification actually evaluated.
- Gains on some settings are mixed; removing some components slightly improves AUC on certain tasks (e.g. Yelp prompt-tuning).
Experimental setup
Datasets: Academic Product Reddit Yelp
Benchmarks: Academic Product Reddit Yelp
Metrics: F1 AUC-ROC
Baselines: GIN GAT GraphSAGE GT EdgePred AttrMask DGI SimGRACE GraphPrompt ProG GPF GCOPE
Builds on: GCOPE EdgePred AttrMask DGI SimGRACE
Closest papers by meaning
- Graph Metanetworks for Processing Diverse Neural Architectures (ICLR 2024 · similarity 0.89)
- How Much Can Transfer? BRIDGE: Bounded Multi-Domain Graph Foundation Model with Generalization Guarantees (ICML 2025 · similarity 0.88)
- Rethinking Propagation for Unsupervised Graph Domain Adaptation (AAAI 2024 · similarity 0.88)
- A Unified Graph Language Model for Multi-Domain Multi-Task Graph Alignment Instruction Tuning (NeurIPS 2026 · similarity 0.88)
- Multi-Domain Graph Foundation Models: Robust Knowledge Transfer via Topology Alignment (ICML 2025 · similarity 0.88)
- Domain Adaptive Unfolded Graph Neural Networks (AAAI 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.