Hypergraph Neural Networks

Yifan FengHaoxuan YouZizhao ZhangRongrong JiYue Gao

article2018AAAI2,177 citations

Modern data systems increasingly deal with complex, multi-modal information where relationships extend beyond simple pairwise links, such as social networks that combine text, visual content, and social ties. Standard graph convolutional networks are constrained by simple graphs that only connect two data points per edge, limiting their ability to model complex group interactions. Meanwhile, traditional hypergraph methods capture these multi-entity relationships but suffer from high computational complexity and memory demands, hindering practical use. The article addresses this gap by introducing Hypergraph Neural Networks, a framework that integrates high-order data correlation into deep learning while maintaining computational efficiency.

The main objective of the article is to formulate and demonstrate a deep learning architecture using hyperedge convolution that processes complex, multi-modal relationships for data classification. To achieve this, the authors designed a layer-wise propagation mechanism that aggregates data from nodes to hyperedges and back to nodes, approximating spectral convolutions on hypergraphs without needing expensive matrix inversions. The framework was evaluated across four standard benchmarks: two document citation networks (Cora and Pubmed) for graph-based semi-supervised classification, and two 3D visual object recognition datasets (ModelNet40 and NTU) using multiple visual feature extractors.

The evaluation produced several clear findings in order of importance. First, on visual object recognition using multi-modal feature structures, the proposed approach significantly outperformed standard graph convolutional networks, achieving accuracy gains between 8.1% and 10.4% on the NTU dataset and reaching 96.7% accuracy on ModelNet40. Second, the framework surpassed competitive 3D deep learning baselines on ModelNet40, beating point-cloud methods such as SO-Net by 3.3% and PointCNN by 4.9%. Third, when restricted to single-feature structures on visual datasets, it consistently maintained a modest performance advantage of 0.3% to 4.3% over graph convolutional networks. Finally, on standard citation networks where data relations are predominantly pairwise and simple, the model matched or slightly exceeded top baselines, achieving 81.6% accuracy on Cora and 80.1% on Pubmed (a 1.1% gain over graph convolutional networks).

These findings demonstrate that hyperedge convolutions provide a mathematically sound and scalable way to fuse heterogeneous, multi-modal data without manual feature alignment. Organizations deploying machine learning for complex data environments can achieve higher classification performance and richer representation learning, reducing errors in multi-modal retrieval and recognition tasks. The results also show that while the framework generalizes standard graph neural networks, its true performance advantages emerge in rich, multi-feature environments rather than simple pairwise network structures.

Stakeholders and engineering teams working with multi-modal recognition, classification, or retrieval should consider adopting hypergraph-based convolutional layers where standard graphs are currently used. When evaluating deployment, teams should focus implementation efforts on applications with multi-modal or complex group structures, as uniform pairwise datasets offer limited return on migration effort. Future efforts should assess performance on larger industrial-scale multi-modal graphs, refine automated hyperedge construction techniques, and evaluate runtime efficiency in latency-critical production environments.

Confidence in these findings is supported by consistent testing across multiple established benchmarks and comparisons against recent state-of-the-art models. However, limitations remain: the visual recognition experiments relied on nearest-neighbor distance metrics to construct the hypergraphs, which introduces sensitivity to metric choices, and the citation dataset evaluations showed minimal gains because the underlying data lacked complex multi-entity relationships. Readers should account for these boundary conditions when forecasting performance on simple or poorly structured datasets.

Fujian Key Laboratory of Sensing and Computing for Smart Cities KLISS Peng Cheng Laboratory Tsinghua University Xiamen University bipartite graphs network architecture recurrent neural networks hypergraph neural networks hyperedge convolution hypergraph learning

arXiv: 1809.09401 iMoonLab/HGNN

Paper: Semi-Supervised Classification with Graph Convolutional Networks, Thomas N. Kipf et al. (2017). Provides the foundational spectral graph convolutional network formulation that Hypergraph Neural Networks adapts and generalizes from pairwise edges to higher-order hyperedge convolutions.
Paper: Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering, Michaël Defferrard et al. (2016). Introduces fast localized spectral graph filtering using Chebyshev polynomial approximations of the graph Laplacian, establishing the mathematical groundwork for spectral convolutions in hypergraphs.
Paper: Geometric Deep Learning: Going beyond Euclidean data, Michael M. Bronstein et al. (2016). Surveys the geometric deep learning paradigm for generalizing neural convolutions to non-Euclidean domains, which HGNN extends to high-order hypergraph structures.
Paper: Spectral Networks and Locally Connected Networks on Graphs, Joan Bruna et al. (2014). Pioneers spectral network constructions using graph Laplacians, providing the foundational theoretical roots for spectral operations in graph and hypergraph representation learning.
Paper: Inductive Representation Learning on Large Graphs, William L. Hamilton et al. (2017). Establishes spatial neighborhood aggregation mechanisms in graph neural networks, serving as key baseline and conceptual context for HGNN's hyperedge-based message passing.
Paper: The Graph Neural Network Model, Franco Scarselli et al. (2009). Introduces the classic Graph Neural Network model for iterative node state propagation, defining the core relational learning problem that hypergraph networks broaden to higher-order relations.

Paper: Weisfeiler and Leman Go Neural: Higher-Order Graph Neural Networks, Christopher Morris et al. (2019). Explores the theoretical expressiveness and limitations of higher-order graph neural networks by relating subgraphs and hyperedges to multidimensional Weisfeiler-Leman tests.
Paper: How Powerful are Graph Neural Networks?, Keyulu Xu et al. (2019). Analyzes the foundational expressive limits of message-passing architectures, offering theoretical guidance for evaluating advanced graph and hypergraph representation learning models.
Paper: A Comprehensive Survey on Graph Neural Networks, Zonghan Wu et al. (2019). Provides a comprehensive taxonomy and survey of advanced graph neural network architectures and learning schemes following early spectral and spatial models.
Paper: Heterogeneous Graph Transformer, Ziniu Hu et al. (2020). Extends complex relational data modeling to heterogeneous web-scale graphs through typed attention mechanisms, building beyond homogeneous and hypergraph convolution settings.
Paper: Simple and Deep Graph Convolutional Networks, Ming Chen et al. (2020). Addresses the over-smoothing problem inherent in multi-layer graph convolutions via initial residual connections, relevant for deepening hypergraph and graph neural networks.
Paper: Predict then Propagate: Graph Neural Networks meet Personalized PageRank, Johannes Gasteiger et al. (2019). Decouples prediction from propagation using Personalized PageRank to mitigate over-smoothing and scale information diffusion across distant connections.
Paper: Fast Graph Representation Learning with PyTorch Geometric, Matthias Fey et al. (2019). Supplies a standardized open-source library for implementing and scaling message-passing operators across diverse geometric, graph, and hypergraph data structures.

Abstract

In this paper, we present a hypergraph neural networks (HGNN) framework for data representation learning, which can encode high-order data correlation in a hypergraph structure. Confronting the challenges of learning representation for complex data in real practice, we propose to incorporate such data structure in a hypergraph, which is more flexible on data modeling, especially when dealing with complex data. In this method, a hyperedge convolution operation is designed to handle the data correlation during representation learning. In this way, traditional hypergraph learning procedure can be conducted using hyperedge convolution operations efficiently. HGNN is able to learn the hidden layer representation considering the high-order data structure, which is a general framework considering the complex data correlations. We have conducted experiments on citation network classification and visual object recognition tasks and compared HGNN with graph convolutional networks and other traditional methods. Experimental results demonstrate that the proposed HGNN method outperforms recent state-of-the-art methods. We can also reveal from the results that the proposed HGNN is superior when dealing with multi-modal data compared with existing methods.

Topics

bipartite graphs network architecture recurrent neural networks

Keywords

hypergraph neural networks hyperedge convolution hypergraph learning hypergraph construction node classification visual object classification semi-supervised learning citation network

Introduction
Related Work
Hypergraph learning
Neural networks on graph
Hypergraph Neural Networks
Hypergraph learning statement
Spectral convolution on hypergraph
Hypergraph neural networks analysis
Implementation
Experiments
Citation network classification
Visual object classification
Hypergraph structure construction on visual datasets
Conclusion
References

Knowls

Knowl 1 — Hypergraph Neural Network Layer Propagation Rule
model/method
The layer-wise propagation formula for a Hypergraph Neural Network (HGNN) updates vertex representations by aggregating information across hyperedges. Given an incidence matrix $H \in \mathbb{R}^{|V| \times |\mathcal{E}|}$ where entries satisfy $h(v, e) = 1$ if vertex $v$ belongs to hyperedge $e$ and $0$ otherwise, a diagonal hyperedge weight matrix $W \in \mathbb{R}^{|\mathcal{E}| \times |\mathcal{E}|}$ , diagonal vertex degree matrix $D_v \in \mathbb{R}^{|V| \times |V|}$ with entries $d(v) = \sum_{e \in \mathcal{E}} w(e) h(v, e)$ , and diagonal hyperedge degree matrix $D_e \in \mathbb{R}^{|\mathcal{E}| \times |\mathcal{E}|}$ with entries $\delta(e) = \sum_{v \in V} h(v, e)$ , the node representation at layer $l+1$ is computed as:

$X^{(l+1)} = \sigma\left(D_v^{-1/2} H W D_e^{-1} H^\top D_v^{-1/2} X^{(l)} \Theta^{(l)}\right)$

where:
- $X^{(l)} \in \mathbb{R}^{|V| \times C_l}$ is the matrix of vertex representations at layer $l$ , with $X^{(0)} = X$ denoting the input features of dimension $C_0$ .
- $\Theta^{(l)} \in \mathbb{R}^{C_l \times C_{l+1}}$ is the learnable filter weight matrix for layer $l$ .
- $\sigma(\cdot)$ denotes a non-linear activation function (such as ReLU).
- $D_v^{-1/2}$ and $D_e^{-1}$ provide vertex and hyperedge degree normalization, respectively.
Knowl 2 — Spectral Hypergraph Convolution and First-Order Polynomial Simplification
theoretical result
Spectral convolution on a hypergraph with $n$ vertices is formulated using the normalized hypergraph Laplacian $\Delta = I - D_v^{-1/2} H W D_e^{-1} H^\top D_v^{-1/2} \in \mathbb{R}^{n \times n}$ , which is symmetric positive semi-definite with eigendecomposition $\Delta = \Phi \Lambda \Phi^\top$ . Here, $\Phi$ contains orthonormal eigenvectors (Fourier basis) and $\Lambda = \text{diag}(\lambda_1, \dots, \lambda_n)$ contains non-negative eigenvalues.

For a vertex signal $x \in \mathbb{R}^n$ and spectral filter $g$ , hypergraph spectral convolution is defined as:

$g \star x = \Phi g(\Lambda) \Phi^\top x$

To eliminate the $\mathcal{O}(n^2)$ complexity of full eigendecomposition, $g(\Lambda)$ is approximated via truncated Chebyshev polynomials up to order $K=1$ parameterized on the scaled Laplacian $\tilde{\Delta} = \frac{2}{\lambda_{\max}}\Delta - I$ . Under the assumption $\lambda_{\max} \approx 2$ :

$g \star x \approx \theta_0 x - \theta_1 D_v^{-1/2} H W D_e^{-1} H^\top D_v^{-1/2} x$

To constrain parameter count and prevent overfitting, parameters are tied using a single scalar $\theta$ such that $\theta_1 = -\frac{1}{2}\theta$ and $\theta_0 = \frac{1}{2}\theta D_v^{-1/2} H D_e^{-1} H^\top D_v^{-1/2}$ , yielding with hyperedge weights $W$ :

$g \star x \approx \theta D_v^{-1/2} H W D_e^{-1} H^\top D_v^{-1/2} x$

For a multi-channel feature matrix $X \in \mathbb{R}^{n \times C_1}$ and filter parameter matrix $\Theta \in \mathbb{R}^{C_1 \times C_2}$ , the hyperedge convolution produces output $Y \in \mathbb{R}^{n \times C_2}$ via:

$Y = D_v^{-1/2} H W D_e^{-1} H^\top D_v^{-1/2} X \Theta$
Knowl 3 — Node-Edge-Node Feature Transformation Mechanism
model/method
The HGNN convolutional layer performs a two-stage spatial message-passing mechanism structured as a node-edge-node transformation:
1. Node Feature Transformation and Gathering (Node-to-Edge): Vertex features $X^{(l)} \in \mathbb{R}^{N \times C_l}$ are transformed by learnable parameters $\Theta^{(l)} \in \mathbb{R}^{C_l \times C_{l+1}}$ and normalized by vertex degrees $D_v^{-1/2}$ . Multiplying by the transposed incidence matrix $H^\top \in \mathbb{R}^{|\mathcal{E}| \times N}$ collects node embeddings belonging to each hyperedge into an edge feature tensor in $\mathbb{R}^{|\mathcal{E}| \times C_{l+1}}$ .
2. Hyperedge Feature Aggregation (Edge-to-Node): The hyperedge features are weighted by diagonal matrix $W$ , normalized by edge degrees $D_e^{-1}$ , and routed back to constituent vertices by multiplying with the incidence matrix $H \in \mathbb{R}^{N \times |\mathcal{E}|}$ and applying vertex normalization $D_v^{-1/2}$ .
This sequence allows HGNN to aggregate non-pairwise, high-order correlations across all nodes sharing hyperedges before redistributing the context back to individual node representations.
Knowl 4 — Hypergraph Construction and Multi-Modal Feature Fusion
model/method
When topology is not explicitly provided or when data includes multi-modal representations, hypergraphs are constructed as follows:
- Metric Space $K$ -NN Construction: For $N$ samples represented by feature vectors $\{x_1, \dots, x_N\}$ , pairwise Euclidean distances $d(x_i, x_j)$ are computed. For each vertex $v_i$ , a hyperedge $e_i$ is formed connecting $v_i$ and its $K$ nearest neighbors, yielding an incidence matrix $H \in \{0, 1\}^{N \times N}$ where each hyperedge contains $K+1$ vertices.
- Graph-Neighborhood Hypergraph Construction: For graph datasets, each vertex and its directly adjacent 1-hop neighbors form a hyperedge centered at that vertex.
- Multi-Modal Hypergraph Fusion: When data contains $M$ distinct modalities or feature sets, an incidence matrix $H_m \in \{0, 1\}^{N \times |\mathcal{E}_m|}$ is built independently for each modality $m \in \{1, \dots, M\}$ . The composite multi-modal incidence matrix $H$ is formed by concatenating all modality matrices along the hyperedge dimension:
$H = [H_1, H_2, \dots, H_M] \in \{0, 1\}^{N \times \sum_{m=1}^M |\mathcal{E}_m|}$

This concatenation allows HGNN to jointly learn from heterogeneous high-order relationships across diverse modalities.
Knowl 5 — Reduction of Hypergraph Neural Networks to Graph Convolutional Networks
theoretical result
Standard Graph Convolutional Networks (GCN) are a special case of the Hypergraph Neural Network (HGNN) framework.

When every hyperedge $e \in \mathcal{E}$ in a hypergraph connects exactly two vertices, the hyperedge degree is uniformly $\delta(e) = 2$ , reducing the hypergraph to a standard pairwise graph. Under this condition, the normalized hypergraph Laplacian $\Delta = I - D_v^{-1/2} H W D_e^{-1} H^\top D_v^{-1/2}$ is mathematically equivalent to the normalized graph Laplacian up to a constant scaling factor of $\frac{1}{2}$ . Consequently, the graph convolution propagation rule in GCN corresponds to a 2-uniform hypergraph restriction of the HGNN layer.
Knowl 6 — Hypergraph Regularization and Hypergraph Laplacian Definition
definition
Let $G = (V, \mathcal{E}, W)$ denote a hypergraph with vertex set $V$ , hyperedge set $\mathcal{E}$ , and diagonal hyperedge weight matrix $W \in \mathbb{R}^{|\mathcal{E}| \times |\mathcal{E}|}$ . The incidence matrix $H \in \{0, 1\}^{|V| \times |\mathcal{E}|}$ is defined by entries $h(v, e) = 1$ if $v \in e$ and $0$ otherwise. Vertex degrees $d(v)$ and hyperedge degrees $\delta(e)$ are defined by:

$d(v) = \sum_{e \in \mathcal{E}} w(e) h(v, e), \quad \delta(e) = \sum_{v \in V} h(v, e)$

with $D_v$ and $D_e$ denoting the diagonal vertex degree and hyperedge degree matrices, respectively.

The hypergraph smoothness regularizer $\Omega(f)$ for a classification function $f: V \to \mathbb{R}$ is defined as:

$\Omega(f) = \frac{1}{2} \sum_{e \in \mathcal{E}} \sum_{\{u, v\} \subseteq V} \frac{w(e) h(u, e) h(v, e)}{\delta(e)} \left(\frac{f(u)}{\sqrt{d(u)}} - \frac{f(v)}{\sqrt{d(v)}}\right)^2 = f^\top \Delta f$

where $\Delta = I - D_v^{-1/2} H W D_e^{-1} H^\top D_v^{-1/2}$ is the positive semi-definite normalized hypergraph Laplacian matrix.

Knowl 7 — Citation Network Node Classification Performance on Cora and Pubmed

data/table

In semi-supervised node classification on citation benchmarks, HGNN is evaluated against graph-based methods. Cora contains 2,708 nodes, 5,429 edges, 1,433 features, and 7 classes (140 training, 500 validation, 1,000 test nodes). Pubmed contains 19,717 nodes, 44,338 edges, 500 features, and 3 classes (60 training, 500 validation, 1,000 test nodes). The table reports average classification accuracy over 100 runs:

HGNN achieves comparable performance to GCN on Cora (+0.1%) and an improvement on Pubmed (+1.1%). Because citation graph-derived hyperedges contain only direct neighborhoods without additional high-order multi-modal data, the hypergraph topology remains closely aligned with the pairwise graph topology.

Knowl 8 — Multi-Modal Structure Comparison between GCN and HGNN on ModelNet40 and NTU

data/table

HGNN is compared with GCN on 3D visual object recognition using the ModelNet40 dataset (12,311 3D shapes across 40 classes; 9,843 train, 2,468 test) and the NTU 3D model dataset (2,012 3D shapes across 67 classes; 1,639 train, 373 test). Features include MVCNN (4096-dimensional) and GVCNN (2048-dimensional). Graph adjacency matrices for GCN are constructed via probabilistic distances and averaged across modalities. For HGNN, hyperedges are generated via 10-nearest neighbors ( $K=10$ ) and concatenated across modalities.

ModelNet40 Classification Accuracy:

	Features for Structure
Feature	GVCNN		MVCNN		GVCNN+MVCNN
	GCN	HGNN	GCN	HGNN	GCN	HGNN
GVCNN	91.8%	92.6%	91.5%	91.8%	92.8%	96.6%
MVCNN	92.5%	92.9%	86.7%	91.0%	92.3%	96.6%
GVCNN+MVCNN	–	–	–	–	94.4%	96.7%

NTU Classification Accuracy:

	Features for Structure
Feature	GVCNN		MVCNN		GVCNN+MVCNN
	GCN	HGNN	GCN	HGNN	GCN	HGNN
GVCNN	78.8%	82.5%	78.8%	79.1%	75.9%	84.2%
MVCNN	74.0%	77.2%	71.3%	75.6%	73.2%	83.6%
GVCNN+MVCNN	–	–	–	–	76.1%	84.2%

HGNN outperforms GCN across all structural feature combinations, showing substantial performance gains when fusing multi-modal structures (GVCNN+MVCNN), improving accuracy over GCN by 2.3% on ModelNet40 and 8.1% on NTU.

Knowl 9 — ModelNet40 Benchmark Comparison Against 3D Deep Learning Methods

data/table

HGNN utilizing combined multi-modal GVCNN and MVCNN features and hypergraph structures is evaluated against specialized 3D deep learning architectures on the ModelNet40 dataset:

HGNN achieves a top accuracy of 96.7%, outperforming PointCNN by 4.9% and SO-Net by 3.3%.

Knowl 10 — Two-Layer HGNN Architecture and Hyperparameter Configuration
experimental setup
For node classification tasks, HGNN is configured as a two-layer hyperedge convolutional network:
- Architecture: Two stacked HGNN convolutional layers with a hidden layer feature dimension of 16, terminating in a softmax layer for class prediction.
- Regularization & Activation: Rectified Linear Unit (ReLU) is used as the activation function between layers. Dropout is applied to intermediate layer representations with a dropout rate of $p = 0.5$ .
- Optimization: The network parameters $\Theta$ are optimized by minimizing the cross-entropy loss function on training labels using the Adam optimizer with a learning rate of $0.001$ .

Coverage note — None was omitted; all core mathematical formulations, spectral convolution derivations, hypergraph construction mechanisms, and empirical classification experiments are covered.

References

1.Atwood, J., and Towsley, D. 2016. Diffusion-Convolutional Neural Networks. In NIPS, 1993–2001.
2.Bruna, J.; Zaremba, W.; Szlam, A.; and LeCun, Y. 2014. Spectral Networks and Locally Connected Networks on Graphs. In Proc. ICLR.
3.Chen, D.-Y.; Tian, X.-P.; Shen, Y.-T.; and Ouhyoung, M. 2003. On Visual Similarity Based 3D Model Retrieval. In Computer Graphics Forum, volume 22, 223–232. Wiley Online Library.
4.Defferrard, M.; Bresson, X.; and Vandergheynst, P. 2016. Convolutional Neural Networks on Graphs with Gast Localized Spectral Filtering. In NIPS, 3844–3852.
5.Feng, Y.; Zhang, Z.; Zhao, X.; Ji, R.; and Gao, Y. 2018. Gvcnn: Group-View Convolutional Neural Networks for 3D Shape Recognition. In Proc. CVPR, 264–272.
6.Gao, Y.; Wang, M.; Tao, D.; Ji, R.; and Dai, Q. 2012. 3-D Object Retrieval and Recognition with Hypergraph Analysis. IEEE Transactions on Image Processing 21(9):4290–4303.
7.Gao, Y.; Wang, M.; Zha, Z.-J.; Shen, J.; Li, X.; and Wu, X. 2013. Visual-Textual Joint Relevance Learning for Tagbased Social Image Search. IEEE Transactions on Image Processing 22(1):363–376.
8.Gori, M.; Monfardini, G.; and Scarselli, F. 2005. A New Model for Learning in Graph Domains. In Proc. IJCNN, volume 2, 729–734. IEEE.
9.Henaff, M.; Bruna, J.; and LeCun, Y. 2015. Deep Convolutional Networks on Graph-Structured Data. arXiv preprint arXiv:1506.05163.
10.Huang, Y.; Liu, Q.; Zhang, S.; and Metaxas, D. N. 2010. Image Retrieval via Probabilistic Hypergraph Ranking. In Proc. CVPR, 3376–3383. IEEE.
11.Huang, Y.; Liu, Q.; and Metaxas, D. 2009. ] Video Object Segmentation by Hypergraph Cut. In Proc. CVPR, 1738–1745. IEEE.
12.Hwang, T.; Tian, Z.; Kuangy, R.; and Kocher, J.-P. 2008. Learning on Weighted Hypergraphs to Integrate Protein Interactions and Gene Expressions for Cancer Outcome Prediction. In Proc. ICDM, 293–302. IEEE.
13.Kingma, D. P., and Ba, J. 2014. Adam: A Method for Stochastic Optimization. In Proc. ICLR.
14.Kipf, T. N., and Welling, M. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In Proc. ICLR.
15.Li, Y.; Bu, R.; Sun, M.; and Chen, B. 2018. PointCNN. In NIPS.
16.Li, J.; Chen, B. M.; and Lee, G. H. 2018. SO-Net: Self-Organizing Network for Point Cloud Analysis. In Proc. CVPR, 9397–9406.
17.Lu, Q., and Getoor, L. 2003. Link-based Classification. In Proc. ICML, 496–503.
18.Monti, F.; Boscaini, D.; Masci, J.; Rodola, E.; Svoboda, J.; and Bronstein, M. M. 2017. Geometric Deep Learning on Graphs and Manifolds Using Mixture Model CNNs. In Proc. CVPR, volume 1, 3.
19.Perozzi, B.; Al-Rfou, R.; and Skiena, S. 2014. Deepwalk: Online Learning of Social Representations. In Proc. SIGKDD, 701–710. ACM.
20.Qi, C. R.; Su, H.; Mo, K.; and Guibas, L. J. 2017a. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. Proc. CVPR 1(2):4.
21.Qi, C. R.; Yi, L.; Su, H.; and Guibas, L. J. 2017b. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In NIPS, 5105–5114.
22.Scarselli, F.; Gori, M.; Tsoi, A. C.; Hagenbuchner, M.; and Monfardini, G. 2009. The Graph Neural Network Model. IEEE Transactions on Neural Networks 20(1):61–80.
23.Sen, P.; Namata, G.; Bilgic, M.; Getoor, L.; Galligher, B.; and Eliassi-Rad, T. 2008. Collective Classification in Network Data. AI magazine 29(3):93.
24.Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; and Salakhutdinov, R. 2014. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. The Journal of Machine Learning Research 15(1):1929–1958.
25.Su, H.; Maji, S.; Kalogerakis, E.; and Learned-Miller, E. 2015. Multi-View Convolutional Neural Networks for 3D Shape Recognition. In Proc. ICCV, 945–953.
26.Velickovic, P.; Cucurull, G.; Casanova, A.; Romero, A.; Lio, P.; and Bengio, Y. 2018. Graph Attention networks. In Proc. ICLR, volume 1.
27.Wu, Z.; Song, S.; Khosla, A.; Yu, F.; Zhang, L.; Tang, X.; and Xiao, J. 2015. 3D ShapeNets: A Deep Representation for Volumetric Shapes. In Proc. CVPR, 1912–1920.
28.Yang, Z.; Cohen, W. W.; and Salakhutdinov, R. 2016. Revisiting Semi-Supervised Learning with Graph Embeddings. Proc. ICML.
29.Zhou, D.; Huang, J.; and Schölkopf, B. 2007. Learning with Hypergraphs: Clustering, Classification, and Embedding. In NIPS, 1601–1608.

Citation

MLA

Feng, Y., et al. “Hypergraph Neural Networks”. arXiv, 2018, http://arxiv.org/abs/1809.09401v3.

APA

Feng, Y., You, H., Zhang, Z., Ji, R., & Gao, Y. (2018). Hypergraph Neural Networks. arXiv. http://arxiv.org/abs/1809.09401v3

Chicago

Feng, Y., H. You, Z. Zhang, R. Ji, and Y. Gao. 2018. “Hypergraph Neural Networks”. arXiv. http://arxiv.org/abs/1809.09401v3.

Harvard

Feng, Y. et al. (2018) “Hypergraph Neural Networks”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1809.09401v3.

Vancouver

1. Feng Y, You H, Zhang Z, Ji R, Gao Y (2018) Hypergraph Neural Networks. arXiv

BibTeX

@article{feng2018hypergraph,
  title = {Hypergraph Neural Networks},
  author = {Feng, Yifan and You, Haoxuan and Zhang, Zizhao and Ji, Rongrong and Gao, Yue},
  year = {2018},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1809.09401v3},
  eprint = {1809.09401}
}

Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

Method	Cora	Pubmed
DeepWalk	67.2%	65.3%
ICA	75.1%	73.9%
Planetoid	75.7%	77.2%
Chebyshev	81.2%	74.4%
GCN	81.5%	79.0%
HGNN	81.6%	80.1%

Method	Classification Accuracy
PointNet	89.2%
PointNet++	90.7%
PointCNN	91.8%
SO-Net	93.4%
HGNN	96.7%

Hypergraph Neural Networks

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Hypergraph Neural Network Layer Propagation Rule

Knowl 2 — Spectral Hypergraph Convolution and First-Order Polynomial Simplification

Knowl 3 — Node-Edge-Node Feature Transformation Mechanism

Knowl 4 — Hypergraph Construction and Multi-Modal Feature Fusion

Knowl 5 — Reduction of Hypergraph Neural Networks to Graph Convolutional Networks

Knowl 6 — Hypergraph Regularization and Hypergraph Laplacian Definition

Knowl 7 — Citation Network Node Classification Performance on Cora and Pubmed

Knowl 8 — Multi-Modal Structure Comparison between GCN and HGNN on ModelNet40 and NTU

Knowl 9 — ModelNet40 Benchmark Comparison Against 3D Deep Learning Methods

Knowl 10 — Two-Layer HGNN Architecture and Hyperparameter Configuration

References

Citation

BibTeX

Source Code

Access the Paper