Research map / Architectures and efficiency / Neural architecture search
Hypergraph Neural Architecture Search
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 HyperNAS, a differentiable neural architecture search over a hypergraph-specific search space with a hypergraph structure-aware distance for leave-one-out architecture selection, evaluated on node classification.
Problem
Designing effective hypergraph neural networks (HGNNs) requires heavy manual architecture tuning, and graph NAS search spaces do not transfer directly to hypergraphs.
Approach
A hypergraph-specific search space over vertex, hyperedge, and skip-connection aggregators, optimized with DARTS-style one-shot differentiable search, and a final architecture chosen by leave-one-out pruning guided by a hypergraph structure-aware distance between predicted and ground-truth label-structure matrices.
Key results
- On standard citation splits HyperNAS reaches 83.9% on Cora, 74.1% on Citeseer and 81.3% on Pubmed, above the best baseline SANE (83.6, 73.9, 81.0).
- On Cora with 2% label rate HyperNAS achieves 79.1% vs 76.9% for DHGNN, a 2.2-point gain.
- On co-authorship hypergraphs HyperNAS reports 77.6% on DBLP and 72.6% on Cora-CA, above HGNN+ (77.3, 72.2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HyperNAS instead of the strongest listed baseline | node classification · Cora | accuracy: 83.6 -> 83.9 | Helped |
| Use HyperNAS instead of the strongest listed baseline | node classification · Citeseer | accuracy: 73.9 -> 74.1 | Helped |
| Use HyperNAS instead of the strongest listed baseline | node classification · Pubmed | accuracy: 81.1 -> 81.3 | Helped |
| Use HyperNAS instead of HGNN+ | hypergraph node classification · DBLP | accuracy: 77.3 -> 77.6 | Helped |
| Use HyperNAS instead of HGNN+ | hypergraph node classification · Cora-CA | accuracy: 72.2 -> 72.6 | Helped |
| Transfer the HyperNAS architecture designed on Citation Network | node classification · Coauthor-CS | accuracy: 89.4 -> 89.7 | Helped |
| Transfer the HyperNAS architecture designed on Citation Network | node classification · Photos | accuracy: 92.5 -> 92.7 | Helped |
| Transfer the HyperNAS architecture designed on Citation Network | node classification · Computers | accuracy: 82.1 -> 82.5 | Helped |
| Replace HyperNAS search with random architecture selection (ablation) | HyperNAS · Cora, Citeseer, and Pubmed | accuracy: Cora: 83.9 -> 82.8; Citeseer: 74.1 -> 73.6; Pubmed: 81.3 -> 80.7 | Hurt |
| Replace differentiable search with reinforcement learning (ablation) | HyperNAS · Cora, Citeseer, and Pubmed | accuracy: Cora: 83.9 -> 83.3; Citeseer: 74.1 -> 73.7; Pubmed: 81.3 -> 80.9 | Hurt |
| Replace hypergraph-guided selection with weight size-based selection (ablation) | HyperNAS · Cora, Citeseer, and Pubmed | accuracy: Cora: 83.9 -> 83.5; Citeseer: 74.1 -> 73.8; Pubmed: 81.3 -> 80.9 | Hurt |
| Replace hypergraph-guided selection with validation loss-based selection (ablation) | HyperNAS · Cora, Citeseer, and Pubmed | accuracy: Cora: 83.9 -> 83.3; Citeseer: 74.1 -> 73.6; Pubmed: 81.3 -> 80.8 | Hurt |
Limitations
- Scalability to large-scale graphs is identified as an open issue.
- Hypergraphs are constructed from graph datasets via k-means clustering rather than native hypergraph data for most experiments.
Experimental setup
Datasets: Cora Citeseer Pubmed Coauthor-CS Amazon-Photos Amazon-Computers DBLP Cora-CA
Benchmarks: Cora Citeseer Pubmed Coauthor-CS Amazon-Photos Amazon-Computers DBLP Cora-CA
Metrics: node classification accuracy
Baselines: GCN GAT HGNN DHGNN HCHA HGNN+ GraphNAS SANE HGNAS++ GMI MVGRL GCA
Builds on: DARTS GraphNAS SANE HGNN HGNN+
Closest papers by meaning
- Training-Free Message Passing for Learning on Hypergraphs (ICLR 2025 · similarity 0.89)
- K-hop Hypergraph Neural Network: A Comprehensive Aggregation Approach (AAAI 2025 · similarity 0.89)
- When Hypergraph Meets Heterophily: New Benchmark Datasets and Baseline (AAAI 2025 · similarity 0.88)
- LightHGNN: Distilling Hypergraph Neural Networks into MLPs for 100x Faster Inference (ICLR 2024 · similarity 0.88)
- Parameter-Free Hypergraph Neural Network for Few-Shot Node Classification (NeurIPS 2025 · similarity 0.88)
- Improved Expressivity of Hypergraph Neural Networks through High-Dimensional Generalized Weisfeiler-Leman Algorithms (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.