Research map / Graph learning / Graph self-supervised learning
AutoGFM: Automated Graph Foundation Model with Adaptive Architecture Customization
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 AutoGFM, which searches graph neural architectures customized per dataset for GNN-based graph foundation models (built on GFT), using a disentangled contrastive encoder, invariant-guided architecture prediction, and a curriculum diversity constraint.
Problem
Existing GNN-based graph foundation models use hand-designed fixed GNN architectures, while optimal architectures vary across domains and tasks (architecture inconsistency), and differentiable GNAS suffers from conflicts and data domination when applied to GFMs.
Approach
Learn an invariant graph representation that predicts a per-graph architecture via prototype-based routing in a weight-sharing supernet, with an invariance loss separating invariant from variant patterns and a curriculum coefficient-of-variation constraint that keeps early architecture choices diverse across datasets.
Key results
- AutoGFM reaches the highest average accuracy of 80.86% across eight datasets in the pre-training and fine-tuning setting, above the best manual GNN (GraphSAGE, 79.84%).
- In few-shot learning, AutoGFM achieves the best accuracy on Cora 7-way, WN18RR 10-way, and CHEMHIV 2-way settings, e.g., 47.76% 5-shot on Cora.
- Ablations on five datasets show removing the disentangled encoder, invariant-guided customization, or curriculum constraint each reduces performance relative to the full model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AutoGFM instead of GIN | GFT · Cora | accuracy: 79.85±1.30 -> 80.32±1.12 | Helped |
| Use AutoGFM instead of GASSO | GFT · PubMed | accuracy: 77.82±1.68 -> 78.28±1.40 | Helped |
| Use AutoGFM instead of GraphSAGE | GFT · Wiki-CS | accuracy: 79.29±0.53 -> 79.45±0.69 | Helped |
| Use AutoGFM instead of GAT | GFT · Arxiv | accuracy: 72.82±0.33 -> 73.39±1.56 | Helped |
| Use AutoGFM instead of GCN | GFT · WN18RR | accuracy: 92.20±0.40 -> 93.17±0.88 | Helped |
| Use AutoGFM instead of GraphSAGE | GFT · FB15K237 | accuracy: 89.92±0.27 -> 90.27±1.64 | Helped |
| Use AutoGFM instead of GraphSAGE | GFT · HIV | accuracy: 72.85±2.47 -> 73.17±2.21 | Helped |
| Use AutoGFM instead of GraphSAGE | GFT · PCBA | accuracy: 78.32±0.24 -> 78.83±1.54 | Helped |
| Add the disentangled contrastive graph encoder (ablation) | AutoGFM · Cora, PubMed, Wiki-CS, FB15K237, PCBA | accuracy | Helped |
| Add invariant-guided architecture customization (ablation) | AutoGFM · Cora, PubMed, Wiki-CS, FB15K237, PCBA | accuracy | Helped |
| Add the curriculum architecture customization mechanism (ablation) | AutoGFM · Cora, PubMed, Wiki-CS, FB15K237, PCBA | accuracy | Helped |
Surprising findings
- Existing differentiable GNAS methods (DARTS, GraphNAS, GASSO, Graces) often underperform manually designed GNNs on several datasets, suggesting search conflicts under architecture inconsistency.
Limitations
- Focuses only on GNN-based graph foundation models; other GFM types are left for future work.
- Results are reported with a fixed GFT base model and a limited search space of two operations per super-network layer in the main experiments.
Experimental setup
Datasets: Cora PubMed WikiCS Arxiv WN18RR FB15K237 HIV PCBA ChEMBL
Benchmarks: Cora PubMed WikiCS Arxiv WN18RR FB15K237 HIV PCBA
Metrics: accuracy
Baselines: GCN GAT GIN GraphSAGE GraphConv BGRL GraphMAE GIANT DARTS GraphNAS GASSO Graces
Builds on: DARTS GFT OFA Graces
Closest papers by meaning
- Data-Augmented Curriculum Graph Neural Architecture Search under Distribution Shifts (AAAI 2024 · similarity 0.89)
- Large Language Models Enhanced Personalized Graph Neural Architecture Search in Federated Learning (AAAI 2025 · similarity 0.88)
- AutoSGNN: Automatic Propagation Mechanism Discovery for Spectral Graph Neural Networks (AAAI 2025 · similarity 0.87)
- Disentangled Continual Graph Neural Architecture Search with Invariant Modular Supernet (ICML 2024 · similarity 0.87)
- Towards Graph Foundation Models: Learning Generalities Across Graphs via Task-Trees (ICML 2025 · similarity 0.87)
- Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement (NeurIPS 2025 · similarity 0.86)
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.