Closer through commonality: Enhancing hypergraph contrastive learning with shared groups

Daehee Kim; Daeyoung Roh; Donghee Han; Keejun Han; Mun Yi

REVIEW 3 major objections 5 minor 45 references

Closer through commonality: Enhancing hypergraph contrastive learning with shared groups

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper introduces weak positive pairs—nodes that share hyperedges with the anchor—weighted by shared-group count, and shows they are the main driver of hypergraph contrastive learning performance.

desk verdict Weak-positive pairs are a real new idea with solid ablation support, but the paper overclaims performance because the controlled comparison is only against TriCL. read the letter →

arxiv 2502.08432 v1 pith:2AC6VPHO submitted 2025-02-12 cs.LG

Daeyoung Roh , Donghee Han , Daehee Kim , Keejun Han , Mun Yi This is my paper · ORCID

classification cs.LG

keywords hypergraphcontrastivelearningweakpositivepairsnodeclassificationself-supervisedneuralnetworkfeatureaugmentationcommonality

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that hypergraph contrastive learning should treat shared group membership as a graded form of similarity, not an all-or-nothing positive/negative decision. It introduces HyFi, which adds Gaussian noise to node features to create a positive view while leaving the hypergraph topology untouched, and defines weak positive pairs as nodes that share at least one hyperedge with the anchor, weighted by the number of shared hyperedges. The paper reports that this fine-grained relation is the main driver of performance: removing weak positives raises average rank from 1.4 to 4.2, a larger drop than removing the noise-view positive pairs. If correct, this shows that high-order commonality in hypergraphs carries contrastive signal that ordinary graph contrastive learning misses, and that topology-preserving augmentation is enough to learn strong embeddings.

What carries the argument

The load-bearing object is the weak-positive weight computed from the incidence matrix product $E = H H^T$, where $E_{ij}$ counts hyperedges shared by nodes $i$ and $j$; Eq. 5 defines $w^{pos}_{i,j} = |E_{ij}| \cdot |E_{ij}|/|E_{ii}|$, combining a group-unit contrast factor with a positive probability. This weight enters a node-level contrastive loss (Eqs. 6–7) computed per group, alongside an edge-level loss, so co-membership is contrasted repeatedly within each shared hyperedge. Complementing it is the augmentation function $X' = X + (-1)^X |\epsilon|$ with $\epsilon \sim N(0, \sigma^2)$, which creates the positive view by perturbing features only, leaving the hypergraph topology intact and keeping time and memory costs linear in $m \cdot n_d \cdot (|V| + |E|)$ per iteration.

What would settle it

Compute the correlation between the number of shared hyperedges and either label agreement or feature cosine similarity on the six datasets not shown in Figure 2; if the correlation is near zero or negative, and HyFi's accuracy advantage over its weak-positive-free ablation collapses on those datasets, the commonality assumption is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a three-way split of contrastive pairs—positive (the anchor's noise view), weak positive (nodes sharing at least one hyperedge, weighted by co-membership count), and negative (all others)—yields higher-quality node embeddings than the standard two-way split. The weak positive weight is $w^{pos}_{i,j} = |E_{i,j}| \cdot |E_{i,j}| / |E_{i,i}|$, the product of shared-hyperedge count and the conditional probability of positive given the anchor's group memberships, so the loss pulls an anchor toward co-members in proportion to how much group structure they share. The paper demonstrates this through node-classification linear evaluation on 10 datasets, where HyFi achieves the best average rank (1.9) among supervised and unsupervised baselines, and through an ablation isolating the weak-positive term as the largest single contributor. The claim is not that more positives are better—varying the number of noise views changes accuracy by under 1%—but that the graded, group-structured positives are what matter.

Load-bearing premise

The load-bearing premise is that the more hyperedges two nodes share, the more similar their features and labels are; the paper checks this on four datasets and then applies the weighting to all ten, so a dataset where shared groups do not imply similarity would break the method.

Editorial extensions

If this is right

HyFi ranks first on average across 10 node-classification datasets, ahead of the previous hypergraph contrastive method TriCL (average rank 1.9 vs. 2.5).
Removing weak positive pairs raises average rank from 1.4 to 4.2, a larger drop than removing noise-view positive pairs (3.2), so group commonality is the main source of contrastive signal.
Because topology is never augmented, training is faster and uses less GPU memory than TriCL, with up to roughly 8x speedups on the largest datasets and no out-of-memory failures.
Performance is nearly flat as the number of noise views goes from 1 to 5, indicating the method's gains do not come from adding more positive samples.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural testable extension is to replace the raw shared-hyperedge count with a normalized co-membership score such as the Jaccard index over hyperedge incidence; the paper does not explore whether this changes the optimal weighting.
The commonality assumption is verified on only four datasets in Figure 2, so the method's transfer to disassortative settings is untested; a dataset where co-membership anti-correlates with label agreement would be a stress test.
Because HyFi leaves topology untouched, it could be stacked with topology-augmenting contrastive methods rather than replacing them, potentially combining commonality signal with multi-view diversity.
The weak-positive idea is not hypergraph-specific in principle; ordinary graphs with overlapping motifs or community structure could define analogous weak positives from shared k-hop neighborhoods.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know. First, the weak-positive-pair idea is genuinely new and the ablation backs it up: dropping it hurts more than dropping the standard positive pair. Second, the paper's headline performance claim overreaches, because the only fully controlled comparison is against TriCL; most supervised baselines are cited from other papers with different protocols.

What's good: the mechanism is simple and plausible. Nodes sharing more hyperedges are pulled closer with weight proportional to shared count (Eq. 5). The group-unit contrast and the feature-noise-only augmentation are both clean departures from prior hypergraph contrastive work (TriCL, S2-HHGR). The efficiency numbers are real: HyFi is 1.4x–8x faster than TriCL and uses less memory, and that comparison is apples-to-apples since both use the same HGNN encoder. The ablation table is the paper's strongest evidence: average rank degrades from 1.4 to 4.2 without weak positives, a larger drop than removing the noise-view positive pairs. The code is public and the experiments are reproducible.

Soft spots. The complexity analysis in Section IV.E.1 is wrong as written. Equation 6 sums negatives over all non-neighbor nodes, which is O(|V|^2) per epoch; the paper claims O(m·nd·(|V|+|E|)). If the implementation samples negatives or batches, the paper needs to say so. The commonality-homophily assumption in Figure 2 is checked on four datasets only, and the method is applied to ten. That's a genuine gap, though the ablation mitigates it: weak positives improve results on most of the ten datasets, so the assumption is probably not badly violated anywhere. Still, the sentence 'our assumption is confirmed in Figure 2' overstates the evidence. The comparison with supervised baselines should be rerun under the evaluation protocol or clearly labeled as 'as reported in previous papers'; the average-rank table mixes protocols. Minor typos and notation collisions (E for both hyperedge set and the H·H^T matrix) don't affect the substance.

Bottom line: the core idea is worth a serious referee. The weak-positive mechanism is novel and the ablation supports it. I'd send it out and ask for a corrected complexity analysis, a rerun or caveat on the supervised baselines, and either the homophily plots for all datasets or a toned-down general claim.

Referee Report

3 major / 5 minor

Summary. The paper proposes HyFi, a hypergraph contrastive learning method that trains an HGNN encoder by contrasting an original view with node-feature noise views, and introduces a third class of pairs called "weak positives": nodes that share at least one hyperedge with the anchor, weighted by the (squared) number of shared hyperedges. The method is evaluated on node classification across ten datasets against ten supervised and six unsupervised baselines, with additional experiments on training time, GPU memory, ablations, augmentation type, and number of positive samples. The central claim is that weak positives, derived from common hyperedge membership, substantially improve hypergraph contrastive learning while the noise-only augmentation keeps the method efficient.

Significance. If the reported results hold, the weak-positive mechanism is a simple and potentially useful contribution to hypergraph contrastive learning: Table V shows that removing weak positives raises the average rank from 1.4 to 4.2, a larger drop than removing positive pairs, and the code is publicly available. The noise-only augmentation is an appealing alternative to topology-destroying augmentations, and the efficiency comparison against TriCL is a real strength. However, the paper currently overclaims generality: the homophily-by-commonality assumption is only verified on four of the ten datasets, and the comparison with most baselines is not controlled, so the headline average-rank result is not yet convincing as a general statement about hypergraph contrastive learning.

major comments (3)

[Section I, Figure 2, Eq. (6)] The central assumption that nodes sharing more hyperedges are more similar is verified only for Cora-C, Cora-A, NTU2012, and ModelNet40, yet the weak-positive term is applied to all ten datasets in Table III. Because Eq. (6) with Eq. (5) actively attracts every co-hyperedged pair by a weight proportional to the squared number of shared hyperedges, a weak or negative correlation in any of the remaining six datasets (e.g., 20News, Mushroom, Zoo) would actively degrade the learned embeddings. Please report for each dataset the correlation between shared-hyperedge count and label agreement or feature-cosine similarity, or show via a per-dataset ablation that the weak-positive term helps on all ten datasets; otherwise restrict the claim to the four verified datasets.
[Sections V.A.3 and V.A.4, Table III] HyFi is evaluated under a 10%/10%/80% random split, while most supervised baselines and several unsupervised baselines (DGI, GRACE, etc.) are taken as "results from their respective references" with unspecified and likely different splits. The average-rank comparison in Table III is therefore not a controlled comparison; for example, standard citation-network evaluations often use a fixed number of labels per class, which is not the same as 10% of nodes. Please rerun all baselines under the identical split and evaluation protocol, or clearly state the source split for each cited number and recompute average ranks only over comparable settings. Also specify how the A.R. column handles missing entries marked '-'.
[Section IV.E.1, Eq. (6)] The claimed time complexity O(m * nd * (|V|+|E|)) omits the cost of the negative-pair term Lneg_n, which sums over all j not in N_zi for every anchor zi and is O(|V|^2 * nd) without negative sampling or batching. Since RQ2 and Table IV advertise training-time efficiency, this omission is material. Please either include the pairwise negative-similarity cost in the complexity analysis, or state the negative-sampling/batching strategy actually used and update the complexity accordingly.

minor comments (5)

[Algorithm 1 and Eq. (2)/(4)] There is a notation inconsistency: Eq. (2) states that f_theta returns (P, Q) with P as node embeddings and Q as hyperedge embeddings, but Algorithm 1 line 5 writes Q, P <- f_theta(G, X) and then uses Q for the node-level loss and P for the edge-level loss.
[Eq. (1) and Algorithm 1 line 8] Eq. (1) defines the perturbation as X' = X + (-1)^X * |epsilon|, while Algorithm 1 line 8 writes X' <- X + N(0, sigma^2); the two are not equivalent, and the text does not state whether clipping or clamping is applied to keep perturbed features in [0,1].
[Section V.A.4] The description of the evaluation protocol is ambiguous: "20 dataset splits with 5 random weight initializations" and "repeated a total of five times" need clarification regarding how many independent runs and splits were actually used.
[Table I] The entry for E contains a typo: "E = {e1, e2, ..., v|E|}" should be e_{|E|}.
[Section V.F] The sentence "This indicates that HyFi's performance is not driven by the relationship of weak positive pairs" appears to contradict the ablation result in Table V and should presumably read "not driven by positive pairs".

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: weak-positive weights are deterministic functions of the hypergraph incidence structure, and the claimed gains are established by held-out linear evaluation and ablations, not by construction.

full rationale

HyFi's weak-positive relationship is defined directly from the incidence matrix: the weight in Eq. 5 is w_pos(i,j) = |E_ij| * (|E_ij|/|E_ii|), a deterministic function of shared hyperedges, and the contrastive loss in Eq. 6 uses that weight without fitting it to labels or to the reported accuracies. The load-bearing assumption that more shared hyperedges imply greater similarity is an empirical premise, illustrated in Figure 2 for four datasets and applied to ten. That is a generalization and validation gap, not circularity: the paper does not define the result in terms of the assumption, nor does it fit a parameter to the downstream performance and then call it a prediction. Evaluation follows the standard linear-protocol setup (train encoder unsupervisedly, then train a linear classifier on fixed embeddings with 10% training, 10% validation, 80% test nodes), and the central claim is supported by Table V ablation results on held-out test accuracy. Benchmark results are compared against external baselines, and the method is code-released for reproduction. No step in the derivation chain reduces to its own inputs by construction, no fitted quantity is renamed as a prediction, and no load-bearing claim rests on a self-citation chain.

Assumptions & free parameters 5 free parameters · 3 assumptions · 1 invented entities

The method introduces no new physical entities. It relies on a domain-specific homophily assumption and several unstated hyperparameters. The weak-positive relationship is a modeling construct whose evidence is internal ablation.

free parameters (5)

Noise standard deviation σ² = not reported
Controls the magnitude of feature perturbation in Eq. (1); selected per dataset but the value is not given in the paper.
Contrastive temperature τ_n = not reported
Temperature in node-level and edge-level contrastive losses (Eq. 6); not listed in the hyperparameter settings.
Edge loss weight α = not reported
Balances node-level and edge-level losses in Eq. (8); value is not specified.
Number of noise views M = 2 (per Figure 5 description)
Equal to the number of positive samples; stated as a hyperparameter and set to 2, with performance shown to be largely insensitive.
HGNN encoder hyperparameters (layers, hidden dimensions) = not reported
Architecture details affect results; not provided in the paper.

assumptions (3)

domain assumption Nodes sharing hyperedges are likely to have similar features (homophily via commonality).
Stated in the Introduction and Section IV.D; validated in Figure 2 on four datasets but assumed for all ten.
domain assumption Additive Gaussian noise with sign flip (Eq. 1) preserves label-relevant information while creating a valid positive view.
The augmentation is designed to keep features in [0,1], but there is no proof or external validation that it preserves semantic content; the ablation supports it indirectly.
standard math The HGNN backbone (Eq. 3) provides a fixed, unbiased feature extractor for contrastive pretraining.
Uses existing HGNN message passing, accepted from prior work [7].

invented entities (1)

Weak positive pair relationship
purpose: Defines a middle tier between positive and negative pairs based on shared hyperedge membership, weighted by shared hyperedge count.
Introduced by this paper; its benefit is shown only by the paper's internal ablation (Table V), with no external falsifiable prediction.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Closer through commonality: Enhancing hypergraph contrastive learning with shared groups." pith.science (2026). https://pith.science/paper/2AC6VPHO

@misc{pith2026250208432,
  author       = {Pith},
  title        = {Pith review of: Closer through commonality: Enhancing hypergraph contrastive learning with shared groups},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2AC6VPHO}},
  note         = {Machine review of arXiv:2502.08432}
}

read the original abstract

Hypergraphs provide a superior modeling framework for representing complex multidimensional relationships in the context of real-world interactions that often occur in groups, overcoming the limitations of traditional homogeneous graphs. However, there have been few studies on hypergraphbased contrastive learning, and existing graph-based contrastive learning methods have not been able to fully exploit the highorder correlation information in hypergraphs. Here, we propose a Hypergraph Fine-grained contrastive learning (HyFi) method designed to exploit the complex high-dimensional information inherent in hypergraphs. While avoiding traditional graph augmentation methods that corrupt the hypergraph topology, the proposed method provides a simple and efficient learning augmentation function by adding noise to node features. Furthermore, we expands beyond the traditional dichotomous relationship between positive and negative samples in contrastive learning by introducing a new relationship of weak positives. It demonstrates the importance of fine-graining positive samples in contrastive learning. Therefore, HyFi is able to produce highquality embeddings, and outperforms both supervised and unsupervised baselines in average rank on node classification across 10 datasets. Our approach effectively exploits high-dimensional hypergraph information, shows significant improvement over existing graph-based contrastive learning methods, and is efficient in terms of training speed and GPU memory cost. The source code is available at https://github.com/Noverse0/HyFi.git.

Figures

Figures reproduced from arXiv: 2502.08432 by the authors.

**Figure 2.** The average of the cosine similarity of the node features [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Illustration of the node-level contrastive loss calculation process in HyFi. HyFi uses both the origin view and the noise [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Illustration of the GPU memory usage comparison between HyFi and TriCL, focusing on how they differ in memory [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Illustration of the performance of node classification as the number of positive samples. The number of noise views is [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 34 canonical work pages

[1]

Uci machine learning repository,

A. Asuncion, D. Newman et al. , “Uci machine learning repository,” 2007

work page 2007
[2]

Hypergraph convolution and hyper- graph attention,

S. Bai, F. Zhang, and P. H. Torr, “Hypergraph convolution and hyper- graph attention,” Pattern Recognition, vol. 110, p. 107637, 2021

work page 2021
[3]

On visual similarity based 3d model retrieval,

D.-Y . Chen, X.-P. Tian, Y .-T. Shen, and M. Ouhyoung, “On visual similarity based 3d model retrieval,” in Computer graphics forum , vol. 22, no. 3. Wiley Online Library, 2003, pp. 223–232

work page 2003
[4]

You are allset: A multiset function framework for hypergraph neural networks,

E. Chien, C. Pan, J. Peng, and O. Milenkovic, “You are allset: A multiset function framework for hypergraph neural networks,” in International Conference on Learning Representations , 2021

work page 2021
[5]

Fast and accurate deep network learning by exponential linear units (elus),

D.-A. Clevert, T. Unterthiner, and S. Hochreiter, “Fast and accurate deep network learning by exponential linear units (elus),” in International Conference on Learning Representations , 2015

work page 2015
[6]

Hnhn: Hypergraph networks with hyperedge neurons,

Y . Dong, W. Sawin, and Y . Bengio, “Hnhn: Hypergraph networks with hyperedge neurons,” in International Conference on Machine Learning Graph Representation Learning and Beyond Workshop , 2020

work page 2020
[7]

Hypergraph neural net- works,

Y . Feng, H. You, Z. Zhang, R. Ji, and Y . Gao, “Hypergraph neural net- works,” in Proceedings of the AAAI conference on artificial intelligence , vol. 33, no. 01, 2019, pp. 3558–3565

work page 2019
[8]

Quantum experiments and hypergraphs: Multiphoton sources for quantum interference, quantum computation, and quantum entanglement,

X. Gu, L. Chen, and M. Krenn, “Quantum experiments and hypergraphs: Multiphoton sources for quantum interference, quantum computation, and quantum entanglement,” Physical Review A , vol. 101, no. 3, p. 033816, 2020

work page 2020

Show all 45 references

[9]

Contrastive multi-view represen- tation learning on graphs,

K. Hassani and A. H. Khasahmadi, “Contrastive multi-view represen- tation learning on graphs,” in International conference on machine learning. PMLR, 2020, pp. 4116–4126

2020
[10]

Robust basket recommendation via noise- tolerated graph contrastive learning,

X. He, T. Wei, and J. He, “Robust basket recommendation via noise- tolerated graph contrastive learning,” in Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , 2023, pp. 709–719

2023
[11]

Unignn: a unified framework for graph and hypergraph neural networks,

J. Huang and J. Yang, “Unignn: a unified framework for graph and hypergraph neural networks,” in International Joint Conference on Artificial Intelligence, 2021

2021
[12]

I’m me, we’re us, and i’m us: Tri-directional contrastive learning on hypergraphs,

D. Lee and K. Shin, “I’m me, we’re us, and i’m us: Tri-directional contrastive learning on hypergraphs,” in Proceedings of the AAAI Con- ference on Artificial Intelligence , vol. 37, no. 7, 2023, pp. 8456–8464

2023
[13]

Coclep: Contrastive learning-based semi-supervised community search,

L. Li, S. Luo, Y . Zhao, C. Shan, Z. Wang, and L. Qin, “Coclep: Contrastive learning-based semi-supervised community search,” IEEE 39th ICDE, 2023

2023
[14]

Hyperbolic hypergraphs for sequential recommendation,

Y . Li, H. Chen, X. Sun, Z. Sun, L. Li, L. Cui, P. S. Yu, and G. Xu, “Hyperbolic hypergraphs for sequential recommendation,” in Proceedings of the 30th ACM international conference on information & knowledge management , 2021, pp. 988–997

2021
[15]

Generating 3d molecules for target protein binding,

M. Liu, Y . Luo, K. Uchino, K. Maruhashi, and S. Ji, “Generating 3d molecules for target protein binding,” arXiv preprint arXiv:2204.09410, 2022

2022 arXiv
[16]

Self-supervised dynamic hypergraph recommendation based on hyper-relational knowl- edge graph,

Y . Liu, H. Xuan, B. Li, M. Wang, T. Chen, and H. Yin, “Self-supervised dynamic hypergraph recommendation based on hyper-relational knowl- edge graph,” in Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , 2023, pp. 1617–1626

2023
[17]

Decoupled weight decay regularization,

I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in International Conference on Learning Representations , 2019. [Online]. Available: https://openreview.net/forum?id=Bkg6RiCqY7

2019
[18]

Dualgraph: Improving semi-supervised graph classification via dual contrastive learning,

X. Luo, W. Ju, M. Qu, C. Chen, M. Deng, X.-S. Hua, and M. Zhang, “Dualgraph: Improving semi-supervised graph classification via dual contrastive learning,” in 2022 IEEE 38th International Conference on Data Engineering (ICDE) . IEEE, 2022, pp. 699–712

2022
[19]

Graph representation learning via graphical mutual information maxi- mization,

Z. Peng, W. Huang, M. Luo, Q. Zheng, Y . Rong, T. Xu, and J. Huang, “Graph representation learning via graphical mutual information maxi- mization,” in Proceedings of The Web Conference 2020, 2020, pp. 259– 270

2020
[20]

The network data repository with interactive graph analytics and visualization,

R. Rossi and N. Ahmed, “The network data repository with interactive graph analytics and visualization,” in Proceedings of the AAAI confer- ence on artificial intelligence , vol. 29, no. 1, 2015

2015
[21]

Hygnn: Drug-drug interaction prediction via hypergraph neural network,

K. M. Saifuddin, B. Bumgardner, F. Tanvir, and E. Akbas, “Hygnn: Drug-drug interaction prediction via hypergraph neural network,” in 2023 IEEE 39th International Conference on Data Engineering (ICDE) . IEEE, 2023, pp. 1503–1516

2023
[22]

Seq- hygan: Sequence classification via hypergraph attention network,

K. M. Saifuddin, C. May, F. Tanvir, M. I. K. Islam, and E. Akbas, “Seq- hygan: Sequence classification via hypergraph attention network,” in Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , 2023, pp. 2167–2177

2023
[23]

Collective classification in network data,

P. Sen, G. Namata, M. Bilgic, L. Getoor, B. Galligher, and T. Eliassi- Rad, “Collective classification in network data,” AI magazine , vol. 29, no. 3, pp. 93–93, 2008

2008
[24]

Infograph: Unsupervised and semi-supervised graph-level representation learning via mutual information maximization,

F.-Y . Sun, J. Hoffmann, V . Verma, and J. Tang, “Infograph: Unsupervised and semi-supervised graph-level representation learning via mutual information maximization,” arXiv preprint arXiv:1908.01000 , 2019

1908 arXiv
[25]

Ms-hgat: memory- enhanced sequential hypergraph attention network for information dif- fusion prediction,

L. Sun, Y . Rao, X. Zhang, Y . Lan, and S. Yu, “Ms-hgat: memory- enhanced sequential hypergraph attention network for information dif- fusion prediction,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 4, 2022, pp. 4156–4164

2022
[26]

Self- supervised hypergraph representation learning for sociological analysis,

X. Sun, H. Cheng, B. Liu, J. Li, H. Chen, G. Xu, and H. Yin, “Self- supervised hypergraph representation learning for sociological analysis,” IEEE Transactions on Knowledge and Data Engineering , 2023

2023
[27]

Large-scale representation learning on graphs via bootstrapping,

S. Thakoor, C. Tallec, M. G. Azar, M. Azabou, E. L. Dyer, R. Munos, P. Veliˇckovi´c, and M. Valko, “Large-scale representation learning on graphs via bootstrapping,” in International Conference on Learning Representations, 2022

2022
[28]

Deep graph infomax,

P. Veli ˇckovi´c, W. Fedus, W. L. Hamilton, P. Li `o, Y . Bengio, and R. D. Hjelm, “Deep graph infomax,” in International Conference on Learning Representations, 2018

2018
[29]

Hypertenet: Hypergraph and transformer-based neural network for personalized list continuation,

M. Vijaikumar, D. Hada, and S. Shevade, “Hypertenet: Hypergraph and transformer-based neural network for personalized list continuation,” in 2021 IEEE International Conference on Data Mining (ICDM) . IEEE, 2021, pp. 1210–1215

2021
[30]

3d shapenets: A deep representation for volumetric shapes,

Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang, and J. Xiao, “3d shapenets: A deep representation for volumetric shapes,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2015, pp. 1912–1920

2015
[31]

Self-supervised hypergraph convolutional networks for session-based recommendation,

X. Xia, H. Yin, J. Yu, Q. Wang, L. Cui, and X. Zhang, “Self-supervised hypergraph convolutional networks for session-based recommendation,” in Proceedings of the AAAI conference on artificial intelligence , vol. 35, no. 5, 2021, pp. 4503–4511

2021
[32]

Contrastive learning for sequential recommendation,

X. Xie, F. Sun, Z. Liu, S. Wu, J. Gao, J. Zhang, B. Ding, and B. Cui, “Contrastive learning for sequential recommendation,” in 2022 IEEE 38th international conference on data engineering (ICDE). IEEE, 2022, pp. 1259–1273

2022
[33]

Hypergcn: A new method for training graph convolutional networks on hypergraphs,

N. Yadati, M. Nimishakavi, P. Yadav, V . Nitin, A. Louis, and P. Talukdar, “Hypergcn: A new method for training graph convolutional networks on hypergraphs,” Advances in neural information processing systems , vol. 32, 2019

2019
[34]

Semi-supervised hyper- graph node classification on hypergraph line expansion,

C. Yang, R. Wang, S. Yao, and T. Abdelzaher, “Semi-supervised hyper- graph node classification on hypergraph line expansion,” in Proceedings of the 31st ACM International Conference on Information & Knowledge Management, 2022, pp. 2352–2361

2022
[35]

Revisiting user mobility and social relationships in lbsns: a hypergraph embedding approach,

D. Yang, B. Qu, J. Yang, and P. Cudre-Mauroux, “Revisiting user mobility and social relationships in lbsns: a hypergraph embedding approach,” in The world wide web conference , 2019, pp. 2147–2157

2019
[36]

Group identification via transitional hypergraph convolution with cross-view self-supervised learning,

M. Yang, Z. Liu, L. Yang, X. Liu, C. Wang, H. Peng, and P. S. Yu, “Group identification via transitional hypergraph convolution with cross-view self-supervised learning,” in Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , 2023, pp...

2023
[37]

Dynamic hypergraph convolutional network,

N. Yin, F. Feng, Z. Luo, X. Zhang, W. Wang, X. Luo, C. Chen, and X.-S. Hua, “Dynamic hypergraph convolutional network,” in 2022 IEEE 38th International Conference on Data Engineering (ICDE) . IEEE, 2022, pp. 1621–1634

2022
[38]

Graph con- trastive learning with augmentations,

Y . You, T. Chen, Y . Sui, T. Chen, Z. Wang, and Y . Shen, “Graph con- trastive learning with augmentations,” Advances in neural information processing systems, vol. 33, pp. 5812–5823, 2020

2020
[39]

Cross-modality and self-supervised protein em- bedding for compound–protein affinity and contact prediction,

Y . You and Y . Shen, “Cross-modality and self-supervised protein em- bedding for compound–protein affinity and contact prediction,” Bioin- formatics, vol. 38, no. Supplement 2, pp. ii68–ii74, 2022

2022
[40]

Double-scale self-supervised hypergraph learning for group recommendation,

J. Zhang, M. Gao, J. Yu, L. Guo, J. Li, and H. Yin, “Double-scale self-supervised hypergraph learning for group recommendation,” in Proceedings of the 30th ACM international conference on information & knowledge management , 2021, pp. 2557–2567

2021
[41]

Multiscale and integrative single-cell hi-c analysis with higashi,

R. Zhang, T. Zhou, and J. Ma, “Multiscale and integrative single-cell hi-c analysis with higashi,” Nature biotechnology , vol. 40, no. 2, pp. 254–261, 2022

2022
[42]

Graph debiased contrastive learning with joint representation clustering

H. Zhao, X. Yang, Z. Wang, E. Yang, and C. Deng, “Graph debiased contrastive learning with joint representation clustering.” in IJCAI, 2021, pp. 3434–3440

2021
[43]

Deep graph con- trastive representation learning,

Y . Zhu, Y . Xu, F. Yu, Q. Liu, S. Wu, and L. Wang, “Deep graph con- trastive representation learning,” inInternational Conference on Machine Learning Graph Representation Learning and Beyond Workshop , 2020

2020
[44]

Graph contrastive learning with adaptive augmentation,

——, “Graph contrastive learning with adaptive augmentation,” in Proceedings of the Web Conference 2021 , 2021, pp. 2069–2080

2021
[45]

Inhomogeneous social recommendation with hypergraph convolutional networks,

Z. Zhu, C. Gao, X. Chen, N. Li, D. Jin, and Y . Li, “Inhomogeneous social recommendation with hypergraph convolutional networks,” in 2022 IEEE 38th International Conference on Data Engineering (ICDE) , 2022

2022

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Uci machine learning repository,

A. Asuncion, D. Newman et al. , “Uci machine learning repository,” 2007

work page 2007

[2] [2]

Hypergraph convolution and hyper- graph attention,

S. Bai, F. Zhang, and P. H. Torr, “Hypergraph convolution and hyper- graph attention,” Pattern Recognition, vol. 110, p. 107637, 2021

work page 2021

[3] [3]

On visual similarity based 3d model retrieval,

D.-Y . Chen, X.-P. Tian, Y .-T. Shen, and M. Ouhyoung, “On visual similarity based 3d model retrieval,” in Computer graphics forum , vol. 22, no. 3. Wiley Online Library, 2003, pp. 223–232

work page 2003

[4] [4]

You are allset: A multiset function framework for hypergraph neural networks,

E. Chien, C. Pan, J. Peng, and O. Milenkovic, “You are allset: A multiset function framework for hypergraph neural networks,” in International Conference on Learning Representations , 2021

work page 2021

[5] [5]

Fast and accurate deep network learning by exponential linear units (elus),

D.-A. Clevert, T. Unterthiner, and S. Hochreiter, “Fast and accurate deep network learning by exponential linear units (elus),” in International Conference on Learning Representations , 2015

work page 2015

[6] [6]

Hnhn: Hypergraph networks with hyperedge neurons,

Y . Dong, W. Sawin, and Y . Bengio, “Hnhn: Hypergraph networks with hyperedge neurons,” in International Conference on Machine Learning Graph Representation Learning and Beyond Workshop , 2020

work page 2020

[7] [7]

Hypergraph neural net- works,

Y . Feng, H. You, Z. Zhang, R. Ji, and Y . Gao, “Hypergraph neural net- works,” in Proceedings of the AAAI conference on artificial intelligence , vol. 33, no. 01, 2019, pp. 3558–3565

work page 2019

[8] [8]

Quantum experiments and hypergraphs: Multiphoton sources for quantum interference, quantum computation, and quantum entanglement,

X. Gu, L. Chen, and M. Krenn, “Quantum experiments and hypergraphs: Multiphoton sources for quantum interference, quantum computation, and quantum entanglement,” Physical Review A , vol. 101, no. 3, p. 033816, 2020

work page 2020