Research map / Label-efficient and robust learning / Self-supervised and contrastive
HypeBoy: Generative Self-Supervised Representation Learning on Hypergraphs
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 hyperedge filling, a generative SSL task for hypergraphs (predict a missing node from the rest of a hyperedge), with theory linking it to node classification, and the HypeBoy method with a feature-reconstruction warm-up.
Problem
Designing generative self-supervised learning for hypergraphs is hard: the space of possible hyperedges grows as 2^|V|, and it is unclear what generative SSL task should be used and how it relates to downstream node classification.
Approach
Reformulate generative hypergraph SSL as hyperedge filling: predict the missing node given the other nodes of a hyperedge, with a softmax over all nodes (size |V| rather than 2^|V|), using cosine similarity between projected node and set embeddings, plus augmentation, projection heads, and a masked feature-reconstruction warm-up.
Key results
- Under fine-tuning, HypeBoy obtains the best average ranking (1.7) across 11 hypergraph datasets, outperforming the second-best TriCL by up to 5.0 accuracy points.
- Under linear evaluation, HypeBoy reaches 59.6 on Citeseer and 63.5 on Cora node classification and has the best average ranking (1.5) in node classification.
- In hyperedge prediction under linear evaluation, HypeBoy has the best average ranking (1.5), e.g., 95.5 AUROC on DBLP-P and 91.1 on Citeseer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| pre-train the UniGCNII encoder with hyperedge filling (HypeBoy) instead of using UniGCNII without SSL pre-training | UniGCNII · node classification, 10 of 11 datasets | accuracy: performance gains up to 12.5 points | Helped |
| use HypeBoy rather than TriCL for pre-training | hypergraph SSL node classification · node classification, 11 benchmark datasets | accuracy: accuracy gap is up to 5.0 points | Mixed |
| add feature reconstruction warm-up to hyperedge filling without projection heads (V1 to V4) (ablation) | HypeBoy · node classification, 11 benchmark datasets | average ranking: 4.3 -> 2.5 | Helped |
| add projection heads to hyperedge filling without feature reconstruction warm-up (V1 to V2) (ablation) | HypeBoy · node classification, 11 benchmark datasets | average ranking: 4.3 -> 3.3 | Helped |
| add projection heads to HypeBoy with feature reconstruction warm-up (V4 to HypeBoy) (ablation) | HypeBoy · node classification, 11 benchmark datasets | average ranking: 2.5 -> 1.4 | Helped |
| add hyperedge filling to feature reconstruction only (V3 to HypeBoy) (ablation) | HypeBoy · node classification, 11 benchmark datasets | average ranking: 3.6 -> 1.4 | Helped |
| add feature reconstruction warm-up to hyperedge filling with projection heads (V2 to HypeBoy) (ablation) | HypeBoy · Cora node classification | accuracy: 59.7 ({\scriptsize 9.2}) -> 62.3 ({\scriptsize 7.7}) | Mixed |
| use HypeBoy rather than TriCL for fixed-embedding node classification | hypergraph SSL representations · node classification, 11 benchmark datasets | accuracy: accuracy gap is up to 6.3 points | Mixed |
| use HypeBoy rather than other SSL methods to learn fixed embeddings | hypergraph representation learning · hyperedge prediction, 11 benchmark datasets | AUROC | Mixed |
Surprising findings
- Graph-oriented generative SSL methods (GraphMAE2, MaskGAE) underperform on hypergraphs, suggesting higher-order interactions matter.
- Without projection heads, the encoder suffers dimensional collapse, and with random initialization HypeBoy tends to over-rely on projection heads, motivating the feature-reconstruction warm-up.
Limitations
- Theoretical analysis relies on a simplified Gaussian data model with binary classes and identity covariance.
- Two-stage warm-up adds training complexity; without it, the encoder over-relies on projection heads.
- Some baselines time out (O.O.T.) on large datasets such as 20News and MaskGAE/HyperGRL runs are incomplete.
Experimental setup
Datasets: Citeseer Cora Pubmed Cora-CA DBLP-P DBLP-A AMiner IMDB MN-40 20News House
Benchmarks: Citeseer Cora Pubmed Cora-CA DBLP-P DBLP-A AMiner IMDB MN-40 20News House
Metrics: accuracy AUROC average ranking
Baselines: MLP HGNN HyperGCN HNHN UniGCN UniGIN UniGCNII AllSet ED-HNN PhenomNN GraphMAE2 MaskGAE
Builds on: UniGCNII GraphMAE2 TriCL Deep Sets alignment and uniformity
Closest papers by meaning
- Hyper-SAGNN: a self-attention based graph neural network for hypergraphs (ICLR 2020 · similarity 0.90)
- Augmentations in Hypergraph Contrastive Learning: Fabricated and Generative (NeurIPS 2022 · similarity 0.89)
- When Hypergraph Meets Heterophily: New Benchmark Datasets and Baseline (AAAI 2025 · similarity 0.88)
- Training-Free Message Passing for Learning on Hypergraphs (ICLR 2025 · similarity 0.88)
- HYGENE: A Diffusion-Based Hypergraph Generation Method (AAAI 2025 · similarity 0.88)
- SPHINX: Structural Prediction using Hypergraph Inference Network (ICML 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.