Research map / Graph learning / Graph self-supervised learning
Zero-Shot Generalization of GNNs over Distinct Attribute Domains
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 STAGE, which encodes pairwise statistical (conditional probability) dependencies between node attributes as edge embeddings via STAGE-edge-graphs, enabling zero-shot GNN generalization to graphs with unseen attribute domains.
Problem
Traditional GNNs cannot generalize zero-shot to new graphs whose node attributes differ in type, name, semantics, or number from the training graphs.
Approach
Replace absolute node attribute values with empirical pairwise statistical dependencies (conditional probabilities of attribute order/value) computed per edge, forming a STAGE-edge-graph per edge; an intra-edge GNN embeds each, and an inter-edge GNN runs on the attribute-free graph with these edge embeddings. Invariance to order-preserving attribute transforms and attribute/node permutations is pr
Key results
- NBFNet-STAGE reaches 0.46 average zero-shot Hits@1 on held-out e-commerce categories, 41% above the strongest baseline (normalized, 0.33).
- On H&M zero-shot link prediction, STAGE gets 0.466 Hits@1 vs 0.230 for the best parametric baseline (llm), a 103% relative gain.
- On zero-shot Friendster-to-Pokec gender classification, GINE-STAGE reaches 0.652 accuracy vs 0.591 for GraphAny (10.3% gain).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STAGE attribute-dependency edge embeddings instead of normalized node attributes | NBFNet · E-Commerce Stores, held-out domain | Hits@1: 0.3269 ± 0.0213 -> 0.4606 ± 0.0123 | Helped |
| Use STAGE attribute-dependency edge embeddings instead of normalized node attributes | NBFNet · E-Commerce Stores, held-out domain | MRR: 0.3844 ± 0.0159 -> 0.4971 ± 0.0073 | Helped |
| Use STAGE attribute-dependency edge embeddings instead of the strongest parametric baseline | NBFNet · H&M | Hits@1: 0.2302 vs. 0.4666 | Helped |
| Use STAGE attribute-dependency edge embeddings instead of Adamic Adar | NBFNet · H&M | MRR: 0.3184 vs. 0.4703 | Helped |
| Use STAGE attribute-dependency edge embeddings instead of GraphAny | GINE · Pokec, gender classification | accuracy: 0.591 ± 0.0083 -> 0.652 ± 0.0042 | Helped |
| Train with more graph domains (ablation) | NBFNet-STAGE · E-Commerce Stores, held-out domain | Hits@1 | Helped |
| Train with more graph domains (ablation) | NBFNet-STAGE · E-Commerce Stores, held-out domain | MRR | Helped |
Surprising findings
- Adamic Adar and Common Neighbors, non-parametric heuristics, outperform several parametric GNN baselines on H&M zero-shot link prediction.
Limitations
- Complexity is quadratic in the number of attributes, limiting scale to small and medium datasets.
- Only pairwise attribute dependencies are modeled; higher-order relations are left to future work.
- Theoretical results are restricted to domains with a fixed number of attributes.
Experimental setup
Datasets: E-Commerce Stores (Kechinov, 2020) H&M Personalized Fashion Recommendations (Kaggle, 2021) Friendster (Teixeira et al., 2019) Pokec (SNAP, 2012)
Benchmarks: E-Commerce Stores H&M Pokec
Metrics: Hits@1 MRR accuracy
Baselines: raw gaussian structural llm normalized GraphAny GCOPE Common Neighbors Adamic Adar Personalized PageRank
Builds on: NBFNet GINE GraphAny PRODIGY
Closest papers by meaning
- On the Benefits of Attribute-Driven Graph Domain Adaptation (ICLR 2025 · similarity 0.87)
- Towards Bridging Generalization and Expressivity of Graph Neural Networks (ICLR 2025 · similarity 0.87)
- Making Classic GNNs Strong Baselines Across Varying Homophily: A Smoothness–Generalization Perspective (NeurIPS 2025 · similarity 0.87)
- Preference-driven Knowledge Distillation for Few-shot Node Classification (NeurIPS 2025 · similarity 0.87)
- LLMs as Zero-shot Graph Learners: Alignment of GNN Representations with LLM Token Embeddings (NeurIPS 2024 · similarity 0.87)
- A Manifold Perspective on the Statistical Generalization of Graph Neural Networks (ICML 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.