Research map / Graph learning / GNN architectures
Hypergraph-enhanced Dual Semi-supervised Graph Classification
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 HEAL, a dual-branch semi-supervised graph classification framework that pairs a learnable hypergraph branch with a line-graph branch and aligns them via relational consistency learning on unlabeled graphs.
Problem
Semi-supervised graph classification with few labeled and many unlabeled graphs, where GNNs capture only low-order local neighborhood information and underuse unlabeled graphs.
Approach
Learns a low-rank hypergraph structure jointly with the GNN encoder to model higher-order node dependencies, builds a line graph over the learned hyperedges to capture hyperedge interactions, and aligns the two branches by matching similarity distributions over anchor labeled graphs held in a FIFO memory bank.
Key results
- HEAL reaches 73.4% on PROTEINS, 72.1% on DD, 73.5% on IMDB-B, 45.9% on REDDIT-M-5k and 68.3% on COLLAB, the best on five of six datasets in the table.
- On IMDB-M, HEAL (44.3%) trails DualGraph (44.8%), which the authors attribute to small graphs where hypergraph construction is unnecessary.
- Ablations show the full dual model beats hypergraph-only, line-graph-only, dual-supervised and ensemble variants on PROTEINS, REDDIT-M-5k and COLLAB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add HEAL's hypergraph and line-graph branches with relational consistency learning | HEAL · PROTEINS | accuracy (%): 71.0 -> 73.4 | Helped |
| add HEAL's hypergraph and line-graph branches with relational consistency learning | HEAL · DD | accuracy (%): 70.8 -> 72.1 | Helped |
| add HEAL's hypergraph and line-graph branches with relational consistency learning | HEAL · IMDB-B | accuracy (%): 72.8 -> 73.5 | Helped |
| add HEAL's hypergraph and line-graph branches with relational consistency learning | HEAL · IMDB-M | accuracy (%): 44.8 -> 44.3 | Hurt |
| add HEAL's hypergraph and line-graph branches with relational consistency learning | HEAL · REDDIT-M-5k | accuracy (%): 44.8 -> 45.9 | Helped |
| add HEAL's hypergraph and line-graph branches with relational consistency learning | HEAL · COLLAB | accuracy (%): 67.7 -> 68.3 | Helped |
| add the line-graph branch to the supervised hypergraph branch (ablation) | HEAL · PROTEINS | accuracy (%): 69.1\pm1.0 -> 70.1\pm0.9 | Helped |
| add the line-graph branch to the supervised hypergraph branch (ablation) | HEAL · REDDIT-M-5K | accuracy (%): 41.2\pm1.2 -> 42.4\pm1.2 | Helped |
| add the line-graph branch to the supervised hypergraph branch (ablation) | HEAL · COLLAB | accuracy (%): 64.4\pm0.7 -> 65.3\pm0.8 | Helped |
| replace the line-graph branch with a second hypergraph branch (ablation) | HEAL · PROTEINS | accuracy (%): 73.4 -> 71.8\pm1.1 | Hurt |
| replace the line-graph branch with a second hypergraph branch (ablation) | HEAL · REDDIT-M-5K | accuracy (%): 45.9 -> 43.6\pm1.3 | Hurt |
| replace the line-graph branch with a second hypergraph branch (ablation) | HEAL · COLLAB | accuracy (%): 68.3 -> 66.7\pm0.7 | Hurt |
Surprising findings
- On IMDB-M, the proposed method underperforms DualGraph and several baselines, which the authors attribute to small graph size making hypergraph convolution less useful.
Limitations
- Underperforms on IMDB-M, a small-graph dataset, where hypergraph modeling is less beneficial.
- Relies on a fixed hyperedge count k (set to 32), with performance peaking and then declining for larger k.
- Evaluated only on six small benchmark datasets with a single 10% label setting for the main table.
Experimental setup
Datasets: PROTEINS DD IMDB-B IMDB-M REDDIT-M-5k COLLAB
Benchmarks: PROTEINS DD IMDB-B IMDB-M REDDIT-M-5k COLLAB
Metrics: classification accuracy
Baselines: WL Graph2Vec Sub2Vec EntMin Mean-Teacher VAT InfoGraph ASGN GraphCL JOAO DualGraph KGNN
Builds on: DualGraph GIN hypergraph convolution InfoNCE
Closest papers by meaning
- DARLING: Dual Hypergraph-Enhanced Curriculum-Guided Graph Structure Learning for Node Classification (AAAI 2026 · similarity 0.90)
- HyperGCN: A New Method For Training Graph Convolutional Networks on Hypergraphs (NeurIPS 2019 · similarity 0.90)
- Nonlinear Feature Diffusion on Hypergraphs (ICML 2022 · similarity 0.90)
- Hypergraph Learning for Unsupervised Graph Alignment via Optimal Transport (AAAI 2025 · similarity 0.89)
- Fast Algorithms for Hypergraph PageRank with Applications to Semi-Supervised Learning (ICML 2024 · similarity 0.88)
- Core Knowledge Learning Framework for Graph (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.