Pith. sign in

REVIEW 3 major objections 4 minor 41 references

Wasserstein Hypergraph Neural Network

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that replacing mean and sum pooling in hypergraph neural networks with Sliced Wasserstein Pooling, which treats each hyperedge as a distribution of nodes, improves node classification and achieves top accuracy on several…

desk verdict Order-dependent pooling undermines an otherwise reasonable transfer of SWP to hypergraphs; the fix is simple but essential. read the letter →

arxiv 2506.09682 v1 pith:LILE2NOE submitted 2025-06-11 cs.LG

Iulia Duta , Pietro Liò This is my paper
classification cs.LG
keywords hypergraphneuralnetworksslicedWassersteinpoolingoptimaltransportnodeclassificationsetrepresentationlearningpermutation-invariantaggregationhigher-orderrelations
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

Wasserstein Hypergraph Neural Network (WHNN) claims that the mean and sum pooling operations used by most hypergraph neural networks discard geometric information about node neighborhoods, and that replacing them with Sliced Wasserstein Pooling recovers that information and improves node classification. The paper models each hyperedge as a probability distribution whose member nodes are samples, and represents a hyperedge by the cost of transporting its node distribution to a shared reference distribution. In experiments on seven real-world datasets, this aggregator outperforms Deep Sets and PMA inside the same two-stage architecture and beats published hypergraph baselines on several datasets. A sympathetic reading is that the paper is establishing optimal-transport geometry as a useful and practical inductive bias for higher-order relational learning.

What carries the argument

The central mechanism is Sliced Wasserstein Pooling (SWP), a permutation-invariant set embedding whose Euclidean distances approximate the Sliced Wasserstein distance between distributions. In WHNN it is used twice per layer, for nodes-to-hyperedges and hyperedges-to-nodes, always with one shared reference distribution q and one shared set of slice directions so that all neighborhoods are embedded into a comparable space. The pooling step includes linear interpolation to a fixed reference cardinality, so large and small hyperedges contribute feature vectors of equal length; the per-slice sorted differences are then combined by learnable scalars. This carries the argument by making both aggregation stages sensitive to the relative arrangement of points inside each neighborhood rather than to their mean alone.

What would settle it

Construct or identify a node-classification dataset in which nodes grouped into the same hyperedge are deliberately dissimilar in feature space while nodes from different hyperedges overlap, so that hyperedge membership is not predicted by feature proximity; if WHNN's advantage over mean and sum baselines disappears or reverses on such data, the distributional interpretation is the source of the gain. A more targeted check would be to ablate the interpolation step and measure accuracy on neighborhoods whose cardinality is informative.

Watch

Extended reading notes

Core claim

WHNN's central claim is that aggregating by Sliced Wasserstein Pooling instead of (weighted) summation preserves the geometry of each hyperedge as a distribution — its shape, spread, and density — and that this geometric information is what improves node classification. Concretely, after an MLP or Set Attention Block encoder, the model projects node embeddings onto L slices, interpolates each hyperedge's nodes to a fixed reference cardinality, sorts the projected values, subtracts sorted samples from a shared reference distribution, and combines the per-slice differences with learnable weights. The paper argues that the Euclidean distance between the resulting hyperedge vectors approximates the Sliced Wasserstein distance between the underlying distributions, so hyperedges that are similar in how their nodes are spread become close in embedding space. It reports top accuracy with WHNN on Cora, Citeseer, Cora-CA, DBLP-CA, ModelNet40, NTU2012, and 20News, with the learnable-reference variant usually at least matching the fixed-reference one.

Load-bearing premise

The load-bearing premise is that each hyperedge can be treated as a probability distribution whose observed nodes are representative samples, so that additional points drawn from the same distribution would also belong to the same hyperedge; the paper states this in Section 4 and acknowledges in Appendix A that real-world scenarios may violate it.

Editorial extensions

If this is right

  • Any two-stage hypergraph architecture that currently uses Deep Sets or PMA-style aggregation could swap in SWP; the paper's ablation shows the Wasserstein aggregator wins regardless of encoder.
  • The edge-dependent SAB encoder is slightly stronger than an MLP encoder but costs more memory, and the low-rank ISAB variant makes it feasible on large datasets.
  • Under the distributional view, homophilic hyperedges behave like low-variance distributions and diverse hyperedges like high-variance ones, giving a geometric reading of group cohesion.
  • Because interpolation erases exact neighborhood cardinality, models should add neighborhood size as an explicit feature whenever group size matters.
  • The same pooling logic can be applied to graph neural network message passing, a direction the paper lists as future work.

Reading between the lines

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

  • Not tested in the paper, a class-conditional reference distribution would let the model align hyperedge geometry with the classification target and could sharpen the learned embeddings.
  • A natural extension, suggested by the paper's own future-work sentence about GNNs, is to apply SWP to ordinary graph message passing, where the distributional assumption is weaker but geometric pooling may still beat sum and mean.
  • An implication the authors leave implicit is that the sorted-difference operation behaves like a soft, differentiable rank statistic, so WHNN may be especially sensitive to ordinal structure in features while remaining invariant to node permutation.
  • If the results transfer, optimal-transport pooling could serve as a general replacement for readout layers in set-and-graph encoders, not only within the specific two-stage hypergraph pipeline tested here.
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.

Referee Report

3 major / 4 minor

Summary. WHNN applies Sliced Wasserstein Pooling (SWP) as the aggregation operator in both stages of a two-stage hypergraph message-passing framework. The paper models each hyperedge and each node neighborhood as a distribution, projects node features along random slices, compares sorted projected values with a shared reference sample, and uses the resulting vector as the pooled representation. Two encoder variants are considered (MLP and SAB/ISAB), and the method is evaluated on seven node-classification datasets against Deep Sets, PMA, and published hypergraph baselines. The experiments include ablations that fix the architecture and change only the aggregator, hyperparameter search, and 10 runs with random splits. The appendix acknowledges the distributional assumption and the loss of exact cardinality information due to interpolation.

Significance. If the aggregation were genuinely permutation-invariant, the paper would offer a useful and timely bridge between optimal transport and hypergraph learning. The ablation design is a real strength: it isolates the aggregator from the encoder, and the paper studies both fixed and learnable reference distributions. The distributional assumption is honestly acknowledged in Appendix A, and I treat it as a scope limitation rather than a correctness error. However, the load-bearing correctness issue in Algorithm 2 and Section 4 means the current manuscript does not yet establish the central claim that SWP is a valid permutation-invariant set aggregator. No code is released, and the reference-projection step is underspecified, so reproducibility is currently incomplete.

major comments (3)
  1. [Section 4, Algorithm 2 (lines 10-16) and Step 2 of Section 4] Algorithm 2 interpolates each neighborhood to size N before sorting it (lines 11-14), and Section 4 Step 2-3 does the same. For a hyperedge represented as a set, there is no canonical ordering of its nodes, and linear interpolation over list positions is not permutation-equivariant; sorting after interpolation does not restore invariance. For example, the scalar set {0,5,10} linear-interpolated to two points can yield different sorted sequences depending on whether the input order is [0,5,10] or [5,10,0]. The same hyperedge would therefore produce different pooled vectors under different node orderings, so WHNN is not a function of the multiset of node features. This invalidates the comparison with permutation-invariant aggregators such as Deep Sets and PMA in Figure 2 and Table 2, since the reported gains could be artifacts of the node ordering in the benchmark splits. The fix is local: sort the projected values first, or use a quantile-based interpolation in sorted space, and then interpolate to the reference size.
  2. [Algorithm 2 (lines 5-7, 16) and Section 4 Step 3] Algorithm 2 projects the entity features X through the slice matrix Theta (line 5) but sorts Q without projecting it (line 7), and then forms Z_s as Q - X'_s (line 16). Section 4 Step 3 likewise compares projected node values z^{e,theta_l}_i with reference points y_i but does not state whether the y_i are projected through the same slice. Unless Q is defined in the projected space for each slice, the subtraction is dimensionally inconsistent and the algorithm is underspecified. Please specify exactly how the reference distribution is sampled, projected, and aligned with the L slices, since the pooled representation is defined relative to this reference.
  3. [Section 5, Table 2 and the Comparison with baselines paragraph] The text states that 'regardless of the encoder, our model consistently obtain[s] top results, outperforming the other methods on all datasets.' Table 2 does not fully support this: on Cora, WHNN_MLP (79.84 +/- 1.56) is below ED-HNN (80.31 +/- 1.35), and on several datasets the best WHNN variant is within one standard deviation of the runner-up. Please either add paired significance tests over the 10 splits or qualify the 'consistently outperforms' claim, so that the empirical conclusion matches the reported uncertainty.
minor comments (4)
  1. [Table 2 header] The header 'WHNN_(I)SAB' appears to be a typo for 'WHNN_SAB'.
  2. [Appendix D] The complexity expression for the hyperedge-to-node stage has an unbalanced parenthesis: 'O(N x (R log R) .'
  3. [General reproducibility] No code or implementation link is provided; given the under-specification of the reference projection and interpolation, a public implementation would materially help reproducibility.
  4. [Section 4 Step 2] The phrase 'increase/decrease the number of nodes using linear interpolation' should define the interpolation grid and explicitly state that interpolation is applied to the sorted projected values; as written it conflicts with Algorithm 2 and with the claimed permutation invariance.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: WHNN's accuracy gains are measured on held-out test splits against external baselines, and SWP is imported from prior work rather than fitted to the target.

full rationale

The paper's central claim is empirical: replacing mean/sum pooling with Sliced Wasserstein Pooling improves hypergraph node classification. That claim is supported by test-set accuracy on seven real-world datasets, with ablations that keep the encoder fixed and vary only the aggregator (SWP vs. Deep Sets vs. PMA), so the comparison does not reduce to a fitted parameter renamed as a prediction. SWP is cited from Naderializadeh et al. (NeurIPS 2021), an external source, not from the authors' own prior work. The 'hyperedges as probability distributions' assumption is stated explicitly in Section 4 and its limitations are acknowledged in Appendix A; it is a modeling choice, not a conclusion made true by definition. The paper's self-citations (e.g., reference [2] on graph representation learning in medicine and reference [19] on MultiSetMixer) appear only as background and related work and are not load-bearing for the main claim. No equation defines the output in terms of the training labels or a fitted quantity; the learnable reference distribution and projection slices are optimized on training data, and performance is reported on held-out test nodes. The Appendix even flags the interpolation-related loss of cardinality information, which is an honest limitation rather than a hidden reuse of the conclusion. The strongest non-circularity concern is the order-dependence of Algorithm 2, where interpolation is applied before sorting; that is a correctness/permutation-invariance issue, not a circularity, because it does not make the reported result equivalent to the input by construction. Accordingly, the appropriate finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No new particles, forces, or physical entities are postulated. The method's machinery consists of SWP from prior work plus learnable reference and projection parameters; the main assumptions are the distributional interpretation of hyperedges and the preservation of geometry under interpolation.

free parameters (4)
  • Reference distribution q = Learnable or fixed; num_ref in {5, 10, 25, 50} selected per dataset
    Each hyperedge is embedded as a distance to q, and the reference and its sample count are chosen by hyperparameter search (Appendix C).
  • Projection slices Theta = Learnable; number of slices L equals MLP_hid (128, 256, or 512)
    Directions used to project node features into 1D before sorting; learned during training.
  • Slice combination weights w_l = Learnable scalars
    A weighted mean combines per-slice distance vectors into the final hyperedge and node embeddings; learned during training.
  • Residual coefficient alpha = 0.5 (fixed)
    Mixes input features with updated features after each aggregation stage; chosen by hand in Appendix C.
assumptions (4)
  • domain assumption Nodes of a hyperedge are samples from one underlying probability distribution, and unobserved samples would also belong to that hyperedge.
    Section 4, "Hyperedges as probability distributions," and Appendix A acknowledge this may fail in real-world scenarios.
  • standard math Euclidean distance between SWP embeddings approximates Sliced Wasserstein distance between the source sets.
    Imported from Naderializadeh et al. [14]; the paper relies on this to claim hyperedge representations reflect transformation cost (Sections 3.2 and 4).
  • ad hoc to paper Linear interpolation between sorted projected values preserves distribution geometry when hyperedge cardinality differs from the reference sample count.
    Algorithm 2 line 12 and Section 4 Step 2; no justification or error analysis is provided.
  • domain assumption The MLP or SAB encoder projects features into a space where the distributional assumption holds approximately.
    Appendix A: "Our model relies solely on the node encoder to project features into a space where the assumption is approximately valid."

how reviews work

0 comments
Cite this review

Pith. "Pith review of Wasserstein Hypergraph Neural Network." pith.science (2026). https://pith.science/paper/LILE2NOE

@misc{pith2026250609682,
  author       = {Pith},
  title        = {Pith review of: Wasserstein Hypergraph Neural Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LILE2NOE}},
  note         = {Machine review of arXiv:2506.09682}
}
read the original abstract

The ability to model relational information using machine learning has driven advancements across various domains, from medicine to social science. While graph representation learning has become mainstream over the past decade, representing higher-order relationships through hypergraphs is rapidly gaining momentum. In the last few years, numerous hypergraph neural networks have emerged, most of them falling under a two-stage, set-based framework. The messages are sent from nodes to edges and then from edges to nodes. However, most of the advancement still takes inspiration from the graph counterpart, often simplifying the aggregations to basic pooling operations. In this paper we are introducing Wasserstein Hypergraph Neural Network, a model that treats the nodes and hyperedge neighbourhood as distributions and aggregate the information using Sliced Wasserstein Pooling. Unlike conventional aggregators such as mean or sum, which only capture first-order statistics, our approach has the ability to preserve geometric properties like the shape and spread of distributions. This enables the learned embeddings to reflect how easily one hyperedge distribution can be transformed into another, following principles of optimal transport. Experimental results demonstrate that applying Wasserstein pooling in a hypergraph setting significantly benefits node classification tasks, achieving top performance on several real-world datasets.

Figures

Figures reproduced from arXiv: 2506.09682 by the authors.

Figure 1
Figure 1. (A) One stage (node-to-hyperedge) of Wasserstein Hypergraph Neural Network pipeline designed to be more sensitive to the geometric structure of the hyperedge compared to the traditional aggregators. First, a node encoder processes the nodes using a simple MLP or an edge-dependent self-attention block (SAB). The hypergraphs is than viewed as a collection of probability distributions {pi}, one for each hyperedge, with… view at source ↗
Figure 2
Figure 2. Ablation study on the importance of Wasserstein aggregator for hypergraph repre￾sentation learning on Citeseer and NTU2012 datasets. We are testing two versions of the Sliced Wasserstein Pooling: with fixed (FPSWE) or learnable (LPSWE) reference distribution. Regardless of the encoder used to project the nodes and hyperedges, the Wasserstein aggregators outperform both the Deep Sets and PMA commonly used inside hype… view at source ↗
Figure 3
Figure 3. Additional results for the ablation study on the importance of Wasserstein aggregator for hypergraph representation learning Cora_CA and ModelNet datasets. FPSWE denotes the Wasserstein aggregator with fixed reference while LPSWE denotes the Wasserstein aggergator with learnable reference distribution. Regardless of the encoder used to project the nodes and hyperedges, the Wasserstein aggregators outperform both the… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 36 canonical work pages

  1. [1]

    Graph AI in Medicine

    Ruth Johnson, Michelle M. Li, Ayush Noori, Owen Queen, and Marinka Zitnik. Graph ai in medicine.CoRR, abs/2310.13767, 2023

    work page Pith review arXiv 2023

  2. [2]

    Catherine Tong, Emma Rocheteau, Petar Veliˇckovi´c, Nicholas Lane, and Pietro Lio.Predicting Patient Outcomes with Graph Representation Learning, pages 281–293. 01 2022

    work page 2022

  3. [3]

    Learning to simulate complex physics with graph networks

    Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. In Hal Daumé III and Aarti Singh, editors,Proceedings of the 37th International Conference on Machine Learning, volume 119, pages 8459–8468, 2020

    work page 2020

  4. [4]

    Graphcast: Learning skillful medium-range global weather forecasting, 2023

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, Alexander Merose, Stephan Hoyer, George Holland, Oriol Vinyals, Jacklynn Stott, Alexander Pritzel, Shakir Mohamed, and Peter Battaglia. Graphcast: Learning skillful medium-range global weather forecas...

    work page 2023

  5. [5]

    Fake news detection on social media using geometric deep learning.arXiv preprint arXiv:1902.06673, 2019

    Federico Monti, Fabrizio Frasca, Davide Eynard, Damon Mannion, and Michael M Bron- stein. Fake news detection on social media using geometric deep learning.arXiv preprint arXiv:1902.06673, 2019

    arXiv 1902

  6. [6]

    Schoenholz, Patrick F

    Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry. In Doina Precup and Yee Whye Teh, edi- tors,Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 1263–1272, 2017

    work page 2017

  7. [7]

    Skipgnn: predicting molecular interactions with skip-graph networks.Scientific reports, 10(1):1–16, 2020

    Kexin Huang, Cao Xiao, Lucas M Glass, Marinka Zitnik, and Jimeng Sun. Skipgnn: predicting molecular interactions with skip-graph networks.Scientific reports, 10(1):1–16, 2020

    work page 2020

  8. [8]

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

    Eli Chien, Chao Pan, Jianhao Peng, and Olgica Milenkovic. You are allset: A multiset function framework for hypergraph neural networks. InInternational Conference on Learning Representations, 2022

    work page 2022

Show all 41 references
  1. [9]

    Equivariant hyper- graph diffusion neural operators.arXiv preprint arXiv:2207.06680, 2022

    Peihao Wang, Shenghao Yang, Yunyu Liu, Zhangyang Wang, and Pan Li. Equivariant hyper- graph diffusion neural operators.arXiv preprint arXiv:2207.06680, 2022

    2022 arXiv

  2. [11]

    Advances in set function learning: A survey of techniques and applications, 2025

    Jiahao Xie and Guangmo Tong. Advances in set function learning: A survey of techniques and applications, 2025

    2025

  3. [12]

    Deep sets

    Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhutdinov, and Alexander J Smola. Deep sets. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing Systems...

    2017

  4. [13]

    Set transformer: A framework for attention-based permutation-invariant neural networks

    Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors,Proceedings of the 36th International Conference ...

    2019

  5. [14]

    Pooling by sliced-wasserstein embedding

    Navid Naderializadeh, Joseph F Comer, Reed Andrews, Heiko Hoffmann, and Soheil Kolouri. Pooling by sliced-wasserstein embedding. In M. Ranzato, A. Beygelzimer, Y . Dauphin, P.S. Liang, and J. Wortman Vaughan, editors,Advances in Neural Information Processing Systems, volume 34...

    2021

  6. [15]

    Hypergraph neural networks.Proc

    Yifan Feng, Haoxuan You, Zizhao Zhang, Rongrong Ji, and Yue Gao. Hypergraph neural networks.Proc. Conf. AAAI Artif. Intell., 33(01):3558–3565, July 2019

    2019

  7. [16]

    Hypergraph node classification with graph neural networks.CoRR, abs/2402.05569, 2024

    Bohan Tang, Zexi Liu, Keyue Jiang, Siheng Chen, and Xiaowen Dong. Hypergraph node classification with graph neural networks.CoRR, abs/2402.05569, 2024

    2024 arXiv

  8. [17]

    Song Bai, Feihu Zhang, and Philip H.S. Torr. Hypergraph convolution and hypergraph attention. Pattern Recognition, 110:107637, 2021

    2021

  9. [18]

    Learnable hypergraph laplacian for hypergraph learning

    Jiying Zhang, Yuzhao Chen, Xiong Xiao, Runiu Lu, and Shutao Xia. Learnable hypergraph laplacian for hypergraph learning. InICASSP, 2022

    2022

  10. [19]

    Hypergraph neural networks through the lens of message passing: A common perspective to homophily and architecture design.Transactions on Machine Learning Research, 2025

    Lev Telyatnikov, Maria Sofia Bucarelli, Guillermo Bernardez, Olga Zaghen, Simone Scardapane, and Pietro Lio. Hypergraph neural networks through the lens of message passing: A common perspective to homophily and architecture design.Transactions on Machine Learning Research, 2025

    2025

  11. [20]

    MLP-mixer: An all-MLP architecture for vision

    Ilya Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Peter Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. MLP-mixer: An all-MLP architecture for vision. In A. Beygelzimer, Y . ...

    2021

  12. [21]

    Co-representation neural hypergraph diffusion for edge- dependent node classification, 2025

    Yijia Zheng and Marcel Worring. Co-representation neural hypergraph diffusion for edge- dependent node classification, 2025

    2025

  13. [22]

    Classification of edge-dependent labels of nodes in hypergraphs

    Minyoung Choe, Sunwoo Kim, Jaemin Yoo, and Kijung Shin. Classification of edge-dependent labels of nodes in hypergraphs. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, page 298–309. ACM, August 2023

    2023

  14. [23]

    Tensorized hypergraph neural networks, 2024

    Maolin Wang, Yaoming Zhen, Yu Pan, Yao Zhao, Chenyi Zhuang, Zenglin Xu, Ruocheng Guo, and Xiangyu Zhao. Tensorized hypergraph neural networks, 2024

    2024

  15. [24]

    Murphy, Balasubramaniam Srinivasan, Vinayak Rao, and Bruno Ribeiro

    Ryan L. Murphy, Balasubramaniam Srinivasan, Vinayak Rao, and Bruno Ribeiro. Janossy pooling: Learning deep permutation-invariant functions for variable-size inputs. InInternational Conference on Learning Representations, 2019

    2019

  16. [25]

    Rep the set: Neural networks for learning set representations.ArXiv, abs/1904.01962, 2019

    Konstantinos Skianis, Giannis Nikolentzos, Stratis Limnios, and Michalis Vazirgiannis. Rep the set: Neural networks for learning set representations.ArXiv, abs/1904.01962, 2019

    1904 arXiv

  17. [26]

    FSPool: Learning set representations with featurewise sort pooling

    Yan Zhang, Jonathon Hare, and Adam Prügel-Bennett. FSPool: Learning set representations with featurewise sort pooling. 2019

    2019

  18. [27]

    Equivari- ant vs

    Abihith Kothapalli, Ashkan Shahbazi, Xinran Liu, Robert Sheng, and Soheil Kolouri. Equivari- ant vs. invariant layers: A comparison of backbone and pooling for point cloud classification, 2024

    2024

  19. [28]

    Wasserstein generative adversarial networks

    Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In Doina Precup and Yee Whye Teh, editors,Proceedings of the 34th International Conference on Machine Learning, volume 70 ofProceedings of Machine Learning Research, pages 214–223....

    2017

  20. [29]

    Distributional sliced-wasserstein and applications to generative modeling

    Khai Nguyen, Nhat Ho, Tung Pham, and Hung Bui. Distributional sliced-wasserstein and applications to generative modeling. InInternational Conference on Learning Representations, 2021

    2021

  21. [30]

    Learning embeddings into entropic wasserstein spaces, 2019

    Charlie Frogner, Farzaneh Mirzazadeh, and Justin Solomon. Learning embeddings into entropic wasserstein spaces, 2019

    2019

  22. [31]

    Point-set distances for learning representations of 3d point clouds, 2021

    Trung Nguyen, Quang-Hieu Pham, Tam Le, Tung Pham, Nhat Ho, and Binh-Son Hua. Point-set distances for learning representations of 3d point clouds, 2021. 11

    2021

  23. [32]

    Wasserstein weisfeiler–lehman graph kernels

    Matteo Togninalli, Elisabetta Ghisu, Felipe Llinares-López, Bastian Rieck, and Karsten Borg- wardt. Wasserstein weisfeiler–lehman graph kernels. In H. Wallach, H. Larochelle, A. Beygelz- imer, F. d’Alché-Buc, E. Fox, and R. Garnett, editors,Advances in Neural Information Proce...

    2019

  24. [33]

    Rohde, and Heiko Hoffmann

    Soheil Kolouri, Navid Naderializadeh, Gustavo K. Rohde, and Heiko Hoffmann. Wasserstein embedding for graph learning. InInternational Conference on Learning Representations, 2021

    2021

  25. [34]

    A trainable optimal transport embedding for feature aggregation and its relationship to attention, 2021

    Grégoire Mialon, Dexiong Chen, Alexandre d’Aspremont, and Julien Mairal. A trainable optimal transport embedding for feature aggregation and its relationship to attention, 2021

    2021

  26. [35]

    Learning wasserstein embeddings

    Nicolas Courty, Rémi Flamary, and Mélanie Ducoffe. Learning wasserstein embeddings. In International Conference on Learning Representations, 2018

    2018

  27. [36]

    Unignn: a unified framework for graph and hypergraph neural net- works

    Jing Huang and Jie Yang. Unignn: a unified framework for graph and hypergraph neural net- works. InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, 2021

    2021

  28. [37]

    Sliced and radon wasser- stein barycenters of measures

    Nicolas Bonneel, Julien Rabin, Gabriel Peyré, and Hanspeter Pfister. Sliced and radon wasser- stein barycenters of measures. 51(1):22–45, January 2015

    2015

  29. [38]

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

    Naganand Yadati, Madhav Nimishakavi, Prateek Yadav, Vikram Nitin, Anand Louis, and Partha Talukdar. Hypergcn: A new method for training graph convolutional networks on hypergraphs. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, A...

    2019

  30. [39]

    3d shapenets: A deep representation for volumetric shapes, 2015

    Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaoou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes, 2015

    2015

  31. [40]

    On visual similarity based 3d model retrieval.Comput

    Ding-Yun Chen, Xiao-Pei Tian, Yu-Te Shen, and Ming Ouhyoung. On visual similarity based 3d model retrieval.Comput. Graph. Forum, 22:223–232, 09 2003

    2003

  32. [41]

    Twenty Newsgroups

    Tom Mitchell. Twenty Newsgroups. UCI Machine Learning Repository, 1997. DOI: https://doi.org/10.24432/C5C323

    1997 doi

  33. [42]

    Hnhn: Hypergraph networks with hyperedge neurons

    Yihe Dong, Will Sawin, and Yoshua Bengio. Hnhn: Hypergraph networks with hyperedge neurons. InGraph Representation Learning and Beyond Workshop at ICML 2020, June 2020. Code available: https://github.com/twistedcubic/HNHN. 12 Appendix: Wasserstein Hypergraph Neural Network Thi...

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.