Pith. sign in

REVIEW 4 major objections 6 minor 57 references

Towards Trustworthy Hypergraph Neural Networks under Label Noise

T0 review · 4 major / 6 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 The paper claims that noisy-label robustness in hypergraph node classification can be achieved by scoring hyperedges by label entropy and then editing the hypergraph structure around the trustworthy ones.

desk verdict A worthwhile benchmark plus a plausible robust hypergraph framework, but the theory assumes its own conclusion and the trustworthiness estimator is never validated against true labels. read the letter →

arxiv 2608.04377 v1 pith:BM52PRSJ submitted 2026-08-05 cs.LG cs.AI

Mengyao Zhou , Zhiheng Zhou , Xiao Han , Guiying Yan This is my paper · ORCID
classification cs.LGcs.AI
keywords hypergraphneuralnetworklabelnoisenodeclassificationhyperedgetrustworthinessentropyrobustlearningbenchmark
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

The paper tackles a specific gap: hypergraph neural networks trained on partly wrong labels are vulnerable because a single noisy node can corrupt many neighbors through high-order message passing. The authors argue that existing label-noise methods, designed for images or ordinary graphs, lose their effectiveness when adapted to hypergraphs, and they build a unified benchmark to show this. Their proposed framework, HyperTrust, estimates the trustworthiness of every hyperedge from the label entropy of its incident nodes, then uses that estimate to add reliable connections for unlabeled nodes and prune unreliable connections that could spread errors. The paper reports that this structure-editing strategy improves accuracy across several datasets, noise types, and backbones, and gives a margin-based theoretical explanation.

What carries the argument

The central object is the entropy-based hyperedge trustworthiness estimate: each hyperedge $e$ receives a label distribution $p^{(e)}_c = \frac{1}{|e|}\sum_{v_i\in e} \mathbb{I}(\tilde y_i = c)$ and an entropy $E(e) = -\sum_c p^{(e)}_c \log p^{(e)}_c$; low-entropy hyperedges are treated as clean and high-entropy ones as noisy. That single score drives both topology-augmenting modules and is also formalized in the theoretical analysis through the trustworthy message propagation margin $\Gamma_i(P)$, which quantifies the gap between class-consistent and class-inconsistent propagation weights and connects directly to the expected classification margin under label noise.

What would settle it

Run HyperTrust on a deliberately heterophilic hypergraph dataset, or construct one where low-entropy hyperedges are preferentially noisy, and compare against the unmodified backbone: if accuracy drops below the backbone or below a random-structure control, the entropy-trustworthiness premise is falsified. A direct check is to measure the precision of low-entropy hyperedges at identifying truly noise-free hyperedges using ground-truth labels on such a dataset.

Watch

Extended reading notes

Core claim

HyperTrust establishes that hypergraph topology itself can be the main lever for robustness under label noise. After pretraining an HGNN on noisy labels, it forms mixed labels for all nodes and computes each hyperedge's label distribution, treating low-entropy hyperedges as trustworthy and high-entropy ones as untrustworthy. The HyperedgeBoost module connects each unlabeled node to its K most similar trustworthy hyperedges, while the HyperedgePrune module removes incidence relations from untrustworthy hyperedges for labeled nodes whose prototype similarity is low; the final prediction is the average of the boosted and pruned views. The theory introduces the trustworthy message propagation margin $\Gamma_i(P) = S^+_i(P) - \max_{r\neq c} S^-_{i,r}(P)$, shows that under uniform noise the expected classification margin equals $\lambda \Gamma_i(P)$, and proves that boost, prune, and their fusion each enlarge this margin under stated purity and selectivity assumptions. Empirically, the method outperforms adapted robust-learning baselines and hypergraph backbones on co-citation, co-authorship, and 3D-vision datasets, with the largest gains under pair noise.

Load-bearing premise

The whole trustworthiness estimator rests on the homophily assumption that hyperedges tend to connect nodes of the same semantic class, so low label entropy reliably means clean; if hyperedges are heterophilic, that ranking is inverted and HyperTrust could reinforce the wrong structure.

Editorial extensions

If this is right

  • If HyperTrust is correct, existing LLN and GLN methods are not adequate for hypergraphs, and hypergraph-specific structure editing should be a standard component of robust hypergraph learning.
  • Label noise that propagates through high-order relations can be contained by editing node–hyperedge incidence: adding reliable supervision paths helps unlabeled nodes, while pruning untrustworthy incidences limits error spread.
  • The entropy-based trustworthiness score provides a simple, backbone-agnostic preprocessing step that can be applied to different hypergraph neural networks, not just the tested HGNN, UniGNN, and UniGAT variants.
  • The margin theory implies that the robustness benefit should grow when hyperedges are more pure in their latent class structure and when the noise rate is higher, which the paper observes in its noise-rate experiments.
  • The unified benchmark itself gives the field a common protocol for comparing future robust hypergraph methods under pair, uniform, and random label noise.

Reading between the lines

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

  • Beyond the paper: the entropy-trustworthiness criterion is essentially a homophily detector, so on heterophilic hypergraphs—where mixed-class hyperedges are informative—HyperTrust would likely invert its ranking and could hurt performance; the paper's own limitations appendix says exactly this, but no heterophilic experiment is run.
  • Beyond the paper: the same trustworthiness score could be applied to hyperedge-level noise or feature noise, not just node-label noise, since it only requires estimated labels for incident nodes.
  • Beyond the paper: the margin framework suggests a testable extension—if label noise is concentrated on low-degree nodes or on nodes at hyperedge boundaries, the expected gains from boost and prune should be larger, because those are the nodes where propagation-weight renormalization matters most.
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

4 major / 6 minor

Summary. The paper addresses hypergraph node classification under label noise. It first adapts representative LLN and GLN baselines to a unified HGNN backbone and evaluates them on seven datasets under pair, uniform, and random label noise. It then proposes HyperTrust, which pretrains an HGNN, estimates hyperedge trustworthiness from the entropy of mixed observed/pseudo labels, and uses this to (i) inject new incidence relations from unlabeled nodes to trustworthy hyperedges (HyperedgeBoost) and (ii) prune labeled-node incidences on untrustworthy hyperedges (HyperedgePrune), fusing the two views for the final prediction. A linear-propagation margin analysis is presented as theoretical support. Experiments report accuracy, fine-grained metrics, ablations, sensitivity, and results with three backbones.

Significance. If the trustworthiness estimator is reliable, HyperTrust is a well-motivated and carefully evaluated contribution: the benchmark and code are useful assets, the multi-backbone experiments over ten runs are thorough, and the ablation study confirms that each proposed module contributes to the reported accuracy. However, the theoretical analysis is conditional on assumptions that essentially assert the desired behavior, and the case study indicates that low-entropy hyperedges under noisy labels are not clean by true labels in a substantial fraction of cases. These gaps mean the central claim of a theoretically grounded and consistently superior method is only partially supported in the current form.

major comments (4)
  1. [Section VI-B, Lemmas 1 and 2, and Theorem 2] The margin-improvement theory is conditional on the very behavior the method is supposed to achieve. Lemma 1 assumes the injected hyperedges have true-class purity at least q* and dominant class equal to the target node's true class c, while Lemma 2 assumes pruning removes more propagation weight from the strongest wrong class than from the true class (beta_{i,r*} > alpha_i). These are exactly the desired consequences of correct trustworthiness estimation, and the proof in Appendix C derives the margin increase algebraically from these assumptions. The paper does not establish that the entropy- or prototype-based procedures satisfy these conditions from observed noisy labels. As a result, Theorem 2 shows that an oracle with perfect modules would improve the margin, but it does not explain why HyperTrust's actual modules work. The authors should either reframe this as an oracle-style conditional analysis and state that limitation explicitly in the main text, or provide a quantitative link between observable low entropy and true-class purity under the assumed noise models.
  2. [Section V-A and Table VII] There is a mismatch between the quantity the theory requires and the quantity the algorithm measures. The trustworthiness estimate (Eqs. 6-10) is computed from observed noisy labels and pseudo-labels, while Lemma 1 is stated in terms of true-class purity. Table VII directly exposes this gap: TN, the proportion of selected trustworthy hyperedges containing at least one noisy node, reaches 69.35% on ModelNet40 under pair noise and exceeds 45% on DBLP-CA across all noise types. The TN definition is not identical to dominant-class mismatch, but it clearly shows that low entropy under observed labels does not imply a clean hyperedge under true labels. No experiment verifies that the low-entropy set has higher true-class purity than random hyperedges, nor that the boost connections are aligned with the true class of the unlabeled node. Please add such a measurement (e.g., true-class purity of E_trust versus E, and true-label precision of boosted incidences) and, if it is unfavorable, temper the trustworthiness interpretation and the claims that the modules 'enhance reliable supervision' and 'suppress noisy propagation.'
  3. [Table II and abstract/conclusion wording] The claim that HyperTrust 'consistently' outperforms baselines is stronger than the statistical evidence reported. Several entries in Table II lack the p<0.05 asterisk, for example the Citeseer random-noise row (OURS 72.18±2.03 versus NRGNN 71.12±1.78), and Tables IV-V do not report significance tests at all. The wording should be adjusted to state how many settings show statistically significant improvement, how many show non-significant improvement, and whether any setting shows a loss. This matters because the abstract and conclusion assert consistent superiority, and the current presentation leaves the reader unable to verify that claim.
  4. [Section VI-B, Eq. (34)-(36)] The fusion analysis uses a surrogate propagation operator P_Fuse = (P_Boost + P_Prune)/2 and proves a margin result for this operator, but the actual model averages prediction logits from two separately trained classifiers (Eq. 20). The paper acknowledges this is a surrogate, yet it does not specify any condition under which the surrogate margin result transfers to the actual model's prediction-level fusion. Since the theoretical guarantee is central to the paper's narrative, the authors should either prove a transfer result (e.g., under a Lipschitz or linearization condition) or clearly state in the main text that the theory applies only to the linear surrogate, not to the implemented algorithm.
minor comments (6)
  1. [Figures 2, 4, 5, 6, and 7] In the manuscript version provided for review, the text inside these figures renders as escape sequences (e.g., '/uni00000026/uni00000052/...'), making the figures effectively unreadable. Please ensure fonts are fully embedded and labels are rendered as text or vector graphics in the submission.
  2. [Section IV-B] The text refers to 'missing purple bars' in Fig. 2(a), but the figure appears in grayscale in the provided version; please clarify the color convention or make the pattern differences visible in grayscale.
  3. [Section VII-A, implementation details] The hyperparameter search for K and rho is described, but it is not stated whether the selection is performed on the validation split or on the test set. Please state explicitly that all hyperparameters are chosen using the validation set to exclude selective reporting.
  4. [Table II] Some baselines listed in Section IV-A (e.g., SCE, Backward) do not appear in the main accuracy table, while they appear in later tables; please reconcile the presentation or state which baselines are in the main table and which are deferred to the appendix.
  5. [Section VII-D, Table VI] The 'All' variant presumably uses the tuned hyperparameters found in the full model, but it is unclear whether the w/o variants reuse the same K and rho values or retune them; please state this to make the ablation interpretable.
  6. [Eq. (13) and Algorithm 1] Minor notation issue: in Eq. (13), 'H i,e = 1' and 'H i,e = 0' should be written as matrix entries H_{i,e} to avoid confusion with the set E; the same applies in Eq. (18).

Circularity Check

2 steps flagged · score 6.0 of 10

The margin-improvement theory assumes the very effectiveness it claims to prove, making the theoretical support partially circular.

  1. self definitional [Section VI-B, Lemma 1 (p. 8)]
    "Lemma 1: For node vi with yi = c, suppose that the HyperedgeBoost module injects a propagation weight ηi ∈ (0, 1] from trustworthy hyperedges whose purity is at least q⋆ and whose dominant class is c. Furthermore, assume that 2q⋆ − 1 > Γi(P ), Then, we have Γi(P Boost) ≥ (1 − ηi)Γi(P ) + ηi(2q⋆ − 1) > Γi(P )."

    The theorem is presented as showing that HyperedgeBoost improves the classification margin, but its premise already states that the injected hyperedges have true-class purity q⋆ and dominant true class c. That is precisely the property the entropy-based trustworthiness estimator (Eqs. 6–8, computed from noisy observed labels and pseudo-labels) is supposed to certify, and the paper never proves or verifies that low-entropy hyperedges satisfy it. Table VII reports TN values up to 69.35% (ModelNet40 pair noise), meaning many selected 'trustworthy' hyperedges contain noisy nodes. The margin improvement is therefore assumed rather than derived from the actual selection criterion.

  2. self definitional [Section VI-B, Lemma 2 (p. 8)]
    "If the prototype-guided pruning is class-selective in the sense that it removes more propagation weight from the strongest competing class than from the true class, i.e., βi,r⋆ > αi, then, after renormalization, the trustworthy message propagation margin is strictly improved: Γi(P Prune) > Γi(P )."

    The proof in Appendix C gives Γi(P Prune) = (Γi(P ) + βi,r⋆ − αi)/(1 − ωi) with 1 − ωi > 0 and Γi(P ) ≥ 0, so the conclusion Γi(P Prune) > Γi(P ) is algebraically equivalent to the assumption βi,r⋆ > αi. The assumption is exactly the 'class-selective' effectiveness that HyperedgePrune is supposed to have; nothing in the proof connects the prototype similarity computed from mixed labels (Eqs. 15–16) to this inequality. The lemma thus reduces to its own premise by construction.

full rationale

The empirical core of the paper is self-contained: HyperTrust is compared against adapted LLN/GLN baselines and hypergraph backbones under a unified benchmark, and those comparisons are external and falsifiable, so the measured accuracy gains are not circular. Self-citations such as [15], [42], and [46] are not load-bearing in the main argument; [15] is cited only as inspiration for adding incidence relations, and the empirical results do not depend on that citation. The circularity is confined to the theoretical analysis. Theorem 1 is a standard algebraic identity linking expected margin to propagation weights. However, the paper's claim that HyperTrust improves the margin (Theorem 2) rests on Lemma 1 and Lemma 2, whose assumptions are the desired behavior itself. Lemma 2's assumption βi,r⋆ > αi is equivalent to its conclusion by Eq. 51, and Lemma 1 assumes the injected hyperedges are true-class pure and aligned, a property the entropy estimator is never shown to deliver. The paper's own Table VII shows many 'trustworthy' hyperedges contain noisy nodes (TN up to 69.35%). Thus the theoretical analysis does not independently establish robustness; it restates the module's intended effect as premises. This is partial circularity, not total, because the benchmark and ablation studies provide independent empirical content.

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

The method adds no new physical or formal entities, but it relies on several assumptions the reader pays for: the homophily prior on hyperedges, the reliability of noisy-label-pretrained pseudo-labels, the linear-propagation surrogate, and the class-selectivity of pruning in the theory. The free parameters are the entropy threshold, the boost count K, the pruning threshold rho, and backbone hyperparameters.

free parameters (4)
  • delta = 1e-10
    Entropy threshold that separates trustworthy from untrustworthy hyperedges; set to 1e-10 with no sensitivity search, so it selects only near-zero-entropy hyperedges. It directly controls which hyperedges get boosted and which get pruned.
  • K = searched over {5, 10, 20, 30, 40, 50, 60}; best often around 50
    Number of most similar trustworthy hyperedges added per unlabeled node in HyperedgeBoost; tuned on noisy validation labels.
  • rho = searched over {0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9}; best around 0.8 on ModelNet40
    Normalized prototype-consistency threshold for removing node-hyperedge incidences in HyperedgePrune; tuned on noisy validation labels.
  • hidden_dim, layers, learning_rate, weight_decay = searched over {16, 32, 64, 128, 256}, {2, 3, 4, 5}, {1e-1 to 5e-4}, {5e-2 to 5e-5}
    Standard backbone hyperparameters selected on validation accuracy; they affect the reported results but are common to all methods.
assumptions (4)
  • domain assumption Homophily: nodes incident to the same hyperedge are assumed to share the true class more often than not, so low-entropy label distributions indicate trustworthy hyperedges.
    Section V-A states "Under the homophily assumption, hyperedges are more likely to connect nodes belonging to the same semantic category." The entire trustworthiness estimation and both subsequent modules depend on this.
  • domain assumption Mixed labels combining noisy observed labels for labeled nodes and pseudo-labels for unlabeled nodes are reliable enough to estimate hyperedge purity.
    Equation (6) constructs mixed labels from a pretrained classifier trained on noisy labels; if pseudo-labels are wrong, the entropy ranking of hyperedges is corrupted.
  • domain assumption Linear propagation approximation: the analysis models label message passing with a row-stochastic matrix P rather than nonlinear HGNN layers.
    Section VI opens with "Under the linear propagation assumption"; the theorem applies to the surrogate, not the actual network, as Appendix E acknowledges.
  • ad hoc to paper Class-selectivity of pruning: beta_i,r* > alpha_i, meaning pruning removes more propagation weight from the strongest competing class than from the true class.
    Lemma 2 assumes this condition; it is exactly the property that the paper claims to demonstrate, and no derivation from prototype similarity is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Trustworthy Hypergraph Neural Networks under Label Noise." pith.science (2026). https://pith.science/paper/BM52PRSJ

@misc{pith2026260804377,
  author       = {Pith},
  title        = {Pith review of: Towards Trustworthy Hypergraph Neural Networks under Label Noise},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BM52PRSJ}},
  note         = {Machine review of arXiv:2608.04377}
}
read the original abstract

Hypergraph neural networks (HGNNs) have demonstrated remarkable capabilities in processing complex higher-order relationships. However, their performance is highly dependent on labeled data, making them vulnerable to label noise. Despite advances in learning with label noise (LLN) and graph learning with label noise (GLN), noisy-label learning on hypergraphs remains underexplored. In this paper, we present a systematic study of hypergraph node classification under label noise. First, we adapt representative LLN and GLN methods to hypergraphs and evaluate them under a unified benchmark, revealing the limitations of existing robust learning strategies for hypergraphs. Building on this, we propose a new hypergraph robust framework, HyperTrust, which first estimates hyperedge trustworthiness through a pretraining-based, entropy-aware strategy, and then incorporates the HyperedgeBoost module to enhance reliable supervision by connecting unlabeled nodes to trustworthy hyperedges, as well as the HyperedgePrune module to suppress noisy propagation by removing untrustworthy node-hyperedge incidences. Finally, two modules work collaboratively to adjust the hypergraph structure and generate final predictions. Extensive experiments and theoretical analysis demonstrate the effectiveness and robustness of HyperTrust on multiple hypergraph datasets under various noisy settings. Our work provides a unified benchmark and an effective solution for hypergraph learning with label noise and lays a foundation for future research in this direction.

Figures

Figures reproduced from arXiv: 2608.04377 by the authors.

Figure 1
Figure 1. Benchmark design for hypergraph learning under [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. (a)–(c) illustrate the performance differences of different methods relative to HGNN across multiple datasets [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The framework of HyperTrust. Specifically, we first pre-train a hypergraph classifier on the hypergraph to obtain node representations and pseudo-labels, which serve as the basis for hyperedge trustworthiness estimation. The formulation is as follows: Z = fpre(X, H), (2) followed by a classifier that generates predicted logits: Y pre = fcls(Z). (3) The pseudo-label of node vi is obtained by: yˆi = arg max c Y pre i,… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: (a)(b)(c) show multi-metric results on the NTU2012 dataset under pair noise. (d)(e)(f) show multi-metric [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: (a)(b)(c) show accuracy on Citeseer with various levels of label noise. (d)(e)(f) shows accuracy on Pubmed [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: (a)–(c) show accuracy and runtime on Cora-CA. (d)–(f) show accuracy and runtime on NTU2012. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: (a)–(c) show accuracy on Cora,ModelNet40, DBLP with various hyper-parameters [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 44 canonical work pages

  1. [17]

    Noise-robust classification with hypergraph neural network

    N. T. V. Dang, L. Tran, and L. Tran, “Noise-robust clas- sification with hypergraph neural network,” arXiv preprint arXiv:2102.01934, 2021. 16

    work page Pith review arXiv 2021

  2. [1]

    Learning with hyper- graphs: Clustering, classification, and embedding,

    D. Zhou, J. Huang, and B. Schölkopf, “Learning with hyper- graphs: Clustering, classification, and embedding,” Advances in neural information processing systems, vol. 19, 2006

    work page 2006

  3. [2]

    A survey on hypergraph representation learning,

    A. Antelmi, G. Cordasco, M. Polato, V. Scarano, C. Spagnuolo, and D. Yang, “A survey on hypergraph representation learning,” ACM Computing Surveys, vol. 56, no. 1, pp. 1–38, 2023

    work page 2023

  4. [3]

    Berge, Hypergraphs: combinatorics of finite sets

    C. Berge, Hypergraphs: combinatorics of finite sets. Elsevier, 1984, vol. 45

    work page 1984

  5. [4]

    Hypergraph topolog- ical quantities for tagged social networks,

    V. Zlatić, G. Ghoshal, and G. Caldarelli, “Hypergraph topolog- ical quantities for tagged social networks,” Physical Review E— Statistical, Nonlinear, and Soft Matter Physics, vol. 80, no. 3, p. 036118, 2009

    work page 2009

  6. [5]

    Social influence maximization in hypergraph in social networks,

    J. Zhu, J. Zhu, S. Ghosh, W. Wu, and J. Yuan, “Social influence maximization in hypergraph in social networks,” IEEE Transactions on Network Science and Engineering, vol. 6, no. 4, pp. 801–811, 2018

    work page 2018

  7. [6]

    Self-supervised hypergraph transformer for recommender systems,

    L. Xia, C. Huang, and C. Zhang, “Self-supervised hypergraph transformer for recommender systems,” in Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, 2022, pp. 2100–2109

    work page 2022

  8. [7]

    Next-item recommendation with sequential hypergraphs,

    J. Wang, K. Ding, L. Hong, H. Liu, and J. Caverlee, “Next-item recommendation with sequential hypergraphs,” in Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval, 2020, pp. 1101–1110

    work page 2020

Show all 57 references
  1. [8]

    Hypergraph models of biological networks to identify genes critical to pathogenic viral response,

    S. Feng, E. Heath, B. Jefferson, C. Joslyn, H. Kvinge, H. D. Mitchell, B. Praggastis, A. J. Eisfeld, A. C. Sims, L. B. Thackray et al., “Hypergraph models of biological networks to identify genes critical to pathogenic viral response,” BMC bioinformatics, vol. 22, no. 1, p. 287, 2021

    2021

  2. [9]

    Hypergraphs and cellu- lar networks,

    S. Klamt, U.-U. Haus, and F. Theis, “Hypergraphs and cellu- lar networks,” PLoS computational biology, vol. 5, no. 5, p. e1000385, 2009

    2009

  3. [10]

    A survey on hypergraph neural networks: an in-depth and step-by-step guide,

    S. Kim, S. Y. Lee, Y. Gao, A. Antelmi, M. Polato, and K. Shin, “A survey on hypergraph neural networks: an in-depth and step-by-step guide,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 6534–6544

    2024

  4. [11]

    Hypergraph convolution on nodes- hyperedges network for semi-supervised node classification,

    H. Wu and M. K. Ng, “Hypergraph convolution on nodes- hyperedges network for semi-supervised node classification,” ACM Transactions on Knowledge Discovery from Data (TKDD), vol. 16, no. 4, pp. 1–19, 2022

    2022

  5. [12]

    Nhp: Neural hypergraph link prediction,

    N. Yadati, V. Nitin, M. Nimishakavi, P. Yadav, A. Louis, and P. Talukdar, “Nhp: Neural hypergraph link prediction,” in Proceedings of the 29th ACM international conference on information & knowledge management, 2020, pp. 1705–1714

    2020

  6. [13]

    Link prediction in social networks based on hypergraph,

    D. Li, Z. Xu, S. Li, and X. Sun, “Link prediction in social networks based on hypergraph,” in Proceedings of the 22nd international conference on world wide web, 2013, pp. 41–42

    2013

  7. [14]

    Hypersynergyx: Synergistic drug combination prediction via hypergraph modeling and knowledge graph-enhanced retrieval- augmented generation,

    Q. Wang, B. Wu, M. Xu, X. Liu, Y. Mao, Z. Zhou, and G. Yan, “Hypersynergyx: Synergistic drug combination prediction via hypergraph modeling and knowledge graph-enhanced retrieval- augmented generation,” IEEE Journal of Biomedical and Health Informatics, 2026

    2026

  8. [15]

    Graph topology adaptive judgment against node label noise,

    M. Zhou, X. Han, W. Wei, and G. Yan, “Graph topology adaptive judgment against node label noise,” Knowledge-Based Systems, p. 114162, 2025

    2025

  9. [16]

    Classification in the presence of label noise: a survey,

    B. Frénay and M. Verleysen, “Classification in the presence of label noise: a survey,” IEEE transactions on neural networks and learning systems, vol. 25, no. 5, pp. 845–869, 2013

    2013

  10. [18]

    Co-teaching: Robust training of deep neural networks with extremely noisy labels,

    B. Han, Q. Yao, X. Yu, G. Niu, M. Xu, W. Hu, I. Tsang, and M. Sugiyama, “Co-teaching: Robust training of deep neural networks with extremely noisy labels,” Neural Information Pro- cessing Systems,Neural Information Processing Systems, Jan 2018

    2018

  11. [19]

    How does disagreement help generalization against label corruption,

    X. Yu, B. Han, J. Yao, G. Niu, I. Tsang, and M. Sugiyama, “How does disagreement help generalization against label corruption,” International Conference on Machine Learning,International Conference on Machine Learning, Jan 2019

    2019

  12. [20]

    Decoupling “when to up- date

    E. Malach and S. Shalev-Shwartz, “Decoupling “when to up- date” from “how to update”,” Neural Information Processing Systems,Neural Information Processing Systems, Jun 2017

    2017

  13. [21]

    Dividemix: Learning with noisy labels as semi-supervised learning,

    J. Li, R. Socher, and S. Hoi, “Dividemix: Learning with noisy labels as semi-supervised learning,” arXiv: Computer Vision and Pattern Recognition,arXiv: Computer Vision and Pattern Recognition, Feb 2020

    2020

  14. [22]

    Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels,

    L. Jiang, Z. Zhou, T. Leung, L. Li, and F. Li, “Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels,” Cornell University - arXiv,Cornell Univer- sity - arXiv, Dec 2017

    2017

  15. [23]

    Training deep neural- networks using a noise adaptation layer,

    J. Goldberger and E. Ben-Reuven, “Training deep neural- networks using a noise adaptation layer,” in International conference on learning representations, 2017

    2017

  16. [24]

    Dimensionality-driven learn- ing with noisy labels,

    X. Ma, Y. Wang, M. Houle, S. Zhou, S. Erfani, S.-T. Xia, S. Wijewickrema, and J. Bailey, “Dimensionality-driven learn- ing with noisy labels,” International Conference on Machine Learning,International Conference on Machine Learning, Jul 2018

    2018

  17. [25]

    Making deep neural networks robust to label noise: a loss correction approach,

    G. Patrini, A. Rozza, A. K. Menon, R. Nock, and L. Qu, “Making deep neural networks robust to label noise: a loss correction approach,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jul 2017. [Online]. A vailable:http://dx.doi.org/10.1109/cvpr.2017.240

    2017 doi

  18. [26]

    Training deep neural networks on noisy labels with bootstrapping,

    S. Reed, H. Lee, D. Anguelov, C. Szegedy, D. Erhan, and A. Ra- binovich, “Training deep neural networks on noisy labels with bootstrapping,” Cornell University - arXiv,Cornell University - arXiv, Dec 2014

    2014

  19. [27]

    Nrgnn: Learning a label noise resistant graph neural network on sparsely and noisily labeled graphs,

    E. Dai, C. Aggarwal, and S. Wang, “Nrgnn: Learning a label noise resistant graph neural network on sparsely and noisily labeled graphs,” in Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, 2021, pp. 227–236

    2021

  20. [28]

    Robust training of graph neural networks via noise governance,

    S. Qian, H. Ying, R. Hu, J. Zhou, J. Chen, D. Z. Chen, and J. Wu, “Robust training of graph neural networks via noise governance,” in Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, 2023, pp. 607–615

    2023

  21. [29]

    Unified robust training for graph neural networks against label noise,

    Y. Li, J. Yin, and L. Chen, “Unified robust training for graph neural networks against label noise,” in Pacific-Asia Conference on Knowledge Discovery and Data Mining. Springer, 2021, pp. 528–540

    2021

  22. [30]

    Noise-robust graph learning by esti- mating and leveraging pairwise interactions,

    X. Du, T. Bian, Y. Rong, B. Han, T. Liu, T. Xu, W. Huang, Y. Li, and J. Huang, “Noise-robust graph learning by esti- mating and leveraging pairwise interactions,” arXiv preprint arXiv:2106.07451, 2021

    2021 arXiv

  23. [31]

    Adversarial label- flipping attack and defense for graph neural networks,

    M. Zhang, L. Hu, C. Shi, and X. Wang, “Adversarial label- flipping attack and defense for graph neural networks,” in 2020 IEEE International Conference on Data Mining (ICDM). IEEE, 2020, pp. 791–800

    2020

  24. [32]

    Learning on graphs under label noise,

    J. Yuan, X. Luo, Y. Qin, Y. Zhao, W. Ju, and M. Zhang, “Learning on graphs under label noise,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  25. [33]

    Robust loss functions under label noise for deep neural networks,

    A. Ghosh, H. Kumar, and P. S. Sastry, “Robust loss functions under label noise for deep neural networks,” Proceedings of the AAAI Conference on Artificial Intelligence, Jun 2022. [Online]. A vailable:http://dx.doi.org/10.1609/aaai.v31i1.10894

    2022 doi

  26. [34]

    Noisygl: A comprehensive benchmark for graph neural networks under label noise,

    Z. Wang, D. Sun, S. Zhou, H. Wang, J. Fan, L. Huang, and J. Bu, “Noisygl: A comprehensive benchmark for graph neural networks under label noise,” arXiv preprint arXiv:2406.04299, 2024

    2024 arXiv

  27. [35]

    Hypergraph neural networks,

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

    2019

  28. [36]

    Hnhn: Hyper- graph networks with hyperedge neurons,

    Y. Dong, W. Sawin, and Y. Bengio, “Hnhn: Hyper- graph networks with hyperedge neurons,” arXiv preprint arXiv:2006.12278, 2020

    2006 arXiv

  29. [37]

    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

  30. [38]

    Hypersage: Generalizing inductive representation learning on hypergraphs,

    D. Arya, D. K. Gupta, S. Rudinac, and M. Worring, “Hypersage: Generalizing inductive representation learning on hypergraphs,” arXiv preprint arXiv:2010.04558, 2020

    2010 arXiv

  31. [39]

    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,” arXiv preprint arXiv:2105.00956, 2021

    2021 arXiv

  32. [40]

    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,” arXiv preprint arXiv:2106.13264, 2021

    2021 arXiv

  33. [41]

    Hypergraph dynamic system,

    J. Yan, Y. Feng, S. Ying, and Y. Gao, “Hypergraph dynamic system,” in The twelfth international conference on learning representations, 2024

    2024

  34. [42]

    Hypergraph neural diffusion: A pde-inspired framework for hypergraph message passing,

    Z. Zhou, M. Zhou, X. Lin, X. Qi, and G. Yan, “Hypergraph neural diffusion: A pde-inspired framework for hypergraph message passing,” arXiv preprint arXiv:2604.10955, 2026

    2026 arXiv

  35. [43]

    K-hop hypergraph neural network: A comprehensive aggregation approach,

    L. Xie, S. Gao, J. Liu, M. Yin, and T. Jin, “K-hop hypergraph neural network: A comprehensive aggregation approach,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 20, 2025, pp. 21 679–21 687

    2025

  36. [44]

    Un- derstanding deep learning requires rethinking generalization,

    C. Zhang, S. Bengio, M. Hardt, B. Recht, and O. Vinyals, “Un- derstanding deep learning requires rethinking generalization,” Cornell University - arXiv,Cornell University - arXiv, Nov 2016

    2016

  37. [45]

    Contrastive learning of graphs under label noise,

    X. Li, Q. Li, H. Qian, J. Wang et al., “Contrastive learning of graphs under label noise,” Neural Networks, vol. 172, p. 106113, 2024

    2024

  38. [46]

    Tackling over-smoothing on hypergraphs: A ricci flow-guided neural diffusion approach,

    M. Zhou, Z. Zhou, X. Han, X. Qi, G. Wang, and G. Yan, “Tackling over-smoothing on hypergraphs: A ricci flow-guided neural diffusion approach,” arXiv preprint arXiv:2603.15696, 2026

    2026

  39. [47]

    Combating noisy labels by agreement: A joint training method with co-regularization,

    H. Wei, L. Feng, X. Chen, and B. An, “Combating noisy labels by agreement: A joint training method with co-regularization,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 13 726–13 735

    2020

  40. [48]

    Normalized loss functions for deep learning with noisy labels,

    X. Ma, H. Huang, Y. Wang, S. Romano, S. Erfani, and J. Bailey, “Normalized loss functions for deep learning with noisy labels,” in International conference on machine learning. PMLR, 2020, pp. 6543–6553

    2020

  41. [49]

    Symmetric cross entropy for robust learning with noisy labels,

    Y. Wang, X. Ma, Z. Chen, Y. Luo, J. Yi, and J. Bailey, “Symmetric cross entropy for robust learning with noisy labels,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 322–330

    2019

  42. [50]

    Clnode: Curriculum learning for node classification,

    X. Wei, X. Gong, Y. Zhan, B. Du, Y. Luo, and W. Hu, “Clnode: Curriculum learning for node classification,” in Proceedings of the sixteenth ACM international conference on web search and data mining, 2023, pp. 670–678

    2023

  43. [51]

    Learning graph neural networks with noisy labels,

    H. NT, C. J. Jin, and T. Murata, “Learning graph neural networks with noisy labels,” arXiv preprint arXiv:1905.01591, 2019

    1905 arXiv

  44. [52]

    Node similarity preserving graph convolutional networks,

    W. Jin, T. Derr, Y. Wang, Y. Ma, Z. Liu, and J. Tang, “Node similarity preserving graph convolutional networks,” in Proceedings of the 14th ACM international conference on web search and data mining, 2021, pp. 148–156

    2021

  45. [53]

    Inferring anchor links across multiple heterogeneous social networks,

    X. Kong, J. Zhang, and P. S. Yu, “Inferring anchor links across multiple heterogeneous social networks,” in Proceedings of the 22nd ACM international conference on Information & Knowledge Management, 2013, pp. 179–188

    2013

  46. [54]

    node2vec: Scalable feature learning for networks,

    A. Grover and J. Leskovec, “node2vec: Scalable feature learning for networks,” in Proceedings of the 22nd ACM SIGKDD inter- national conference on Knowledge discovery and data mining, 2016, pp. 855–864

    2016

  47. [55]

    Current and future directions in network biology,

    M. Zitnik, M. M. Li, A. Wells, K. Glass, D. Morselli Gysi, A. Krishnan, T. Murali, P. Radivojac, S. Roy, A. Baudot et al., “Current and future directions in network biology,” p. vbae099, 2024

    2024

  48. [56]

    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

  49. [57]

    On vi- sual similarity based 3d model retrieval,

    D.-Y. Chen, X.-P. Tian, Y.-T. Shen, and M. Ouhyoung, “On vi- sual similarity based 3d model retrieval,” in Computer graphics forum. Wiley Online Library, 2003, pp. 223–232. 1 Appendix A The pseudo-code of the HyperTrust In this section, we present the pseudo-code of Hyper- Tru...

    2003

Pith tools

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