Research map / Graph learning / GNN architectures
When Hypergraph Meets Heterophily: New Benchmark Datasets and Baseline
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.
Introduces hypergraph homophily measures, four new heterophilic hypergraph benchmarks (Actor co-occurrence, Amazon-ratings, Twitch-gamers, Pokec), and HyperUFG, a framelet-based hypergraph neural network with low- and high-pass filters as a baseline for heterophilic hypergraph learning.
Problem
Heterophilic hypergraph learning is understudied: existing hypergraph neural networks often underperform MLPs on low-homophily hypergraphs, existing homophily measures are ill-suited for hypergraphs, and few suitable heterophilic hypergraph benchmarks exist.
Approach
Defines hyperedge- and node-level homophily ratios for hypergraphs, builds new large heterophilic hypergraph datasets, and proposes HyperUFG, a spectral framelet hypergraph convolution combining low-pass and high-pass framelet filters with initial residual and identity mapping.
Key results
- On heterophilic hypergraphs, HyperUFG reaches 40.53 ± 2.25% on Twitch-gamers versus 26.70% for MLP and 26.93% for SheafHyperGNN, and ranks first on the heterophilic benchmark average.
- On Actor, HyperUFG scores 89.32 ± 0.75% versus 91.86% for ED-HNN and 92.52% for HyperND, so it is not best on every dataset.
- Most existing HNNs (e.g., HGNN, HyperGCN, AllDeepSets) underperform MLP on heterophilic datasets such as Amazon-ratings and Twitch-gamers.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HyperUFG instead of SheafHyperGNN | Cora node classification · Cora | accuracy: 81.30 ± 1.70 -> 81.51 ± 0.99 | Helped |
| Use HyperUFG instead of SheafHyperGNN | Citeseer node classification · Citeseer | accuracy: 74.71 ± 1.23 -> 74.72 ± 2.10 | No clear effect |
| Use HyperUFG instead of ED-HNN | Pubmed node classification · Pubmed | accuracy: 89.03 ± 0.53 -> 88.73 ± 0.42 | Hurt |
| Use HyperUFG instead of SheafHyperGNN | Cora-CA node classification · Cora-CA | accuracy: 85.52 ± 1.28 -> 85.18 ± 0.69 | Hurt |
| Use HyperUFG instead of ED-HNN | DBLP-CA node classification · DBLP-CA | accuracy: 91.90 ± 0.19 -> 91.67 ± 0.31 | Hurt |
| Use HyperUFG instead of HyperND | Actor node classification · Actor | accuracy: 92.52 ± 0.81 -> 89.32 ± 0.75 | Hurt |
| Use HyperUFG instead of SheafHyperGNN | Amazon-ratings node classification · Amazon-ratings | accuracy: 26.93 ± 3.04 -> 40.53 ± 2.25 | Helped |
| Use HyperUFG instead of MLP | Twitch-gamers node classification · Twitch-gamers | accuracy: 52.77 ± 1.81 -> 52.35 ± 0.04 | Hurt |
| Use HyperUFG instead of ED-HNN | Pokec node classification · Pokec | accuracy: 59.11 ± 0.57 -> 62.30 ± 0.12 | Helped |
| Use HyperUFG instead of SheafHyperGNN | Senate node classification · Senate | accuracy: 68.73 ± 4.68 -> 67.61 ± 7.00 | Hurt |
| Use HyperUFG instead of SheafHyperGNN | House node classification · House | accuracy: 73.84 ± 2.30 -> 72.82 ± 2.22 | Hurt |
| Increase the hyperedge homophily ratio | HNNs on synthetic hypergraphs · Synthetic hypergraphs | accuracy | Helped |
Surprising findings
- Most classic HNNs underperform a graph-agnostic MLP on low-homophily hypergraphs, including on the Twitch-gamers dataset where MLP beats all baselines.
- Several existing benchmarks (Congress, Senate, House, Walmart) that prior work presumed heterophilic have CE homophily above 0.5 under the authors' measures.
Limitations
- Synthetic HSBM generation with r=3 and two classes only reaches a minimum homophily of 0.3.
- HyperUFG does not win on every dataset; ED-HNN and HyperND are better on Actor, and the ranking relies on a single-GPU grid-searched setup with baseline hyperparameters taken from original papers.
- Heterophily-specific baselines are limited; many advanced heterophilic GNN ideas were not tested on hypergraphs.
Experimental setup
Datasets: Actor (co-occurrence) Amazon-ratings Twitch-gamers Pokec Cora Citeseer Pubmed Cora-CA DBLP-CA Senate House Congress
Benchmarks: Actor Amazon-ratings Twitch-gamers Pokec Senate House Cora Citeseer Pubmed Cora-CA DBLP-CA synthetic HSBM hypergraphs
Metrics: node classification accuracy (%) hyperedge homophily node homophily
Baselines: MLP HGNN HyperGCN UniGCNII HyperND AllDeepSets AllSetTransformer ED-HNN SheafHyperGNN
Builds on: HGNN HyperGCN UniGNN ED-HNN GNNs with initial residual and identity mapping (GCNII)
Closest papers by meaning
- Graph homophily booster: Reimagining the role of discrete features in heterophilic graph learning (ICLR 2026 · similarity 0.91)
- Deep Hypergraph Neural Networks with Tight Framelets (AAAI 2025 · similarity 0.90)
- What Is Missing For Graph Homophily? Disentangling Graph Homophily For Graph Neural Networks (NeurIPS 2024 · similarity 0.89)
- On the Impact of Feature Heterophily on Link Prediction with Graph Neural Networks (NeurIPS 2024 · similarity 0.89)
- Training-Free Message Passing for Learning on Hypergraphs (ICLR 2025 · similarity 0.89)
- HeroFilter: Adaptive Spectral Graph Filter for Varying Heterophilic Relations (NeurIPS 2025 · similarity 0.89)
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.