Research map / Graph learning / Graph self-supervised learning
Unsupervised Episode Generation for Graph Meta-learning
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 NaQ, an unsupervised episode generation method (random support nodes plus similar-node queries by feature or diffusion similarity) that lets any supervised graph meta-learner train on unlabeled graphs for few-shot node classification.
Problem
Few-shot node classification with graph meta-learning requires abundant labeled base-class nodes, which causes poor performance under label scarcity, label noise, and class imbalance, while GCL pretraining ignores the downstream task format.
Approach
Generate training episodes without labels: randomly sample N 1-shot support nodes from the whole graph, then take each support node's top-Q most similar nodes (by raw feature cosine/Euclidean similarity, NaQ-Feat, or by personalized-PageRank graph diffusion, NaQ-Diff) as its pseudo-labeled queries, so any supervised meta-learner can be trained unsupervised.
Key results
- On Amazon-Clothing 5-way 1-shot, NaQ-Feat-Best reaches 86.58% vs. 82.24% for the best supervised baseline (COSMIC).
- On Cora-Full 5-way 1-shot, NaQ-Feat-Best reaches 66.30% vs. 62.24% for COSMIC and 62.59% for TLP-BGRL.
- On ogbn-arxiv 5-way 1-shot, NaQ-Feat reaches 54.09% vs. 50.32% for COSMIC (supervised best).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train episodes with NaQ-Feat instead of supervised episodes | Graph meta-learning for FSNC · Amazon-Clothing, 5-way 1-shot | accuracy: 82.24± -> 86.58± | Helped |
| Train episodes with NaQ-Feat instead of supervised episodes | Graph meta-learning for FSNC · Amazon-Electronics, 5-way 1-shot | accuracy: 72.61± -> 76.46± | Helped |
| Train episodes with NaQ-Feat instead of supervised episodes | Graph meta-learning for FSNC · Cora-Full, 5-way 1-shot | accuracy: 62.59± -> 66.30± | Helped |
| Use graph diffusion neighbors as queries rather than the feature-based NaQ variant | NaQ on DBLP · DBLP, 5-way 1-shot | accuracy: 73.55± -> 76.58± | Helped |
| Use feature similarity for query selection rather than graph diffusion | NaQ on Amazon-Clothing · Amazon-Clothing, 5-way 1-shot | accuracy: 84.40± -> 86.58± | Helped |
| Use cosine similarity instead of negative Euclidean distance for query selection (ablation) | NaQ-Feat · Cora-Full, 5-way 1-shot | accuracy: 45.60% -> 64.20% | Helped |
| Use cosine similarity instead of negative Euclidean distance for query selection (ablation) | NaQ-Feat · DBLP, 5-way 1-shot | accuracy: 67.53% -> 71.38% | Helped |
| Use Jaccard similarity instead of cosine similarity for query selection (ablation) | NaQ-Feat · Amazon-Electronics, 5-way 1-shot | accuracy: 76.46% -> 76.63% | Helped |
Surprising findings
- Training with 1-shot unlabeled support sets outperforms supervised baselines that use 5-shot support sets.
- NaQ-Diff degrades on Amazon-Electronics as the number of queries grows beyond 5, since its queries have low class-level similarity there.
Limitations
- NaQ-Diff cannot compute the truncated diffusion on large graphs with many edges such as ogbn-products.
- Random support set sampling may place nodes of the same class in different support sets, introducing label noise.
- Performance under feature noise drops for NaQ-Feat because it relies on raw feature similarity.
Experimental setup
Datasets: Amazon-Clothing Amazon-Electronics Cora-Full DBLP ogbn-arxiv
Benchmarks: Amazon-Clothing Amazon-Electronics Cora-Full DBLP ogbn-arxiv
Metrics: accuracy 95% confidence interval
Baselines: MAML ProtoNet TENT G-Meta GLITTER COSMIC TLP-BGRL TLP-SUGRL TLP-AFGRL VNT
Builds on: ProtoNet MAML TENT G-Meta generalized graph diffusion
Closest papers by meaning
- Learning to Propagate for Graph Meta-Learning (NeurIPS 2019 · similarity 0.86)
- Node Classification on Graphs with Few-Shot Novel Labels via Meta Transformed Network Embedding (NeurIPS 2020 · similarity 0.86)
- Graph Meta Learning via Local Subgraphs (NeurIPS 2020 · similarity 0.85)
- UNR-Explainer: Counterfactual Explanations for Unsupervised Node Representation Learning Models (ICLR 2024 · similarity 0.84)
- A Hierarchical Bayesian Model for Few-Shot Meta Learning (ICLR 2024 · similarity 0.83)
- Unsupervised Meta-Learning via In-Context Learning (ICLR 2025 · similarity 0.83)
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.