Pith. sign in

REVIEW 5 major objections 5 minor 61 references

Scalable and Effective Negative Sample Generation for Hyperedge Prediction

T0 review · 5 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read SEHP shows that negative hyperedges for training can be manufactured by a diffusion model that denoises a positive hyperedge embedding while increasing the discriminator score, outperforming all evaluated baselines and achieving a 20-71x…

desk verdict Plausible and fast latent-space negative sampling for hyperedge prediction, but the 'diffusion' framing is a stretch and generated negatives' validity is unverified — worth reviewing. read the letter →

arxiv 2411.12354 v1 pith:GZH4VLGM submitted 2024-11-19 cs.IR

Shilin Qu , Weiqing Wang , Yuan-Fang Li , Quoc Viet Hung Nguyen , Hongzhi Yin This is my paper · ORCID
classification cs.IR
keywords hyperedgepredictionnegativesamplegenerationconditionaldiffusionlatent-spaceboundary-awarelosssub-hypergraphbatchinghypergraphneuralnetworksscalability
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 argues that hyperedge prediction is held back less by the classifier than by the quality of negative training samples, and that a diffusion model can manufacture the hard negatives that rule-based and adversarial samplers miss. SEHP starts from a positive hyperedge's embedding and iteratively denoises it while pushing the discriminator's score upward, so each generated candidate lands near the decision boundary; the denoising is conditioned on the sub-hypergraph's pooled node embeddings to capture global structure. The paper reports that SEHP outperforms all compared baselines on AUROC and Precision across six real-world datasets, and that a latent-space variant that skips node-ID decoding is 20 to 71 times faster per training epoch with under one percent average metric loss. A sympathetic reader would take away that generative, boundary-aware negative sampling can be both more accurate and more scalable than the fixed schemes currently used.

What carries the argument

The central mechanism is a boundary-aware conditional diffusion loop $h_{t+1}^e = \mathrm{Denoise}(h_t^e, h_{\mathrm{sub}})$ with generator loss $\mathcal{L}_{\mathrm{gen}} = \mathcal{L}_{\mathrm{neg}} + \mathcal{L}_{\mathrm{diff}}$, where $\mathcal{L}_{\mathrm{neg}} = -s_i$ asks the generated samples to be hard for the discriminator and $\mathcal{L}_{\mathrm{diff}}$ asks successive steps to raise the discriminator score. Starting from $h_0^e$, the embedding of a positive hyperedge, the loop iteratively moves the sample toward the decision boundary while the condition $h_{\mathrm{sub}}$ (the average-pooled node embeddings of the sub-hypergraph) keeps the sample aligned with the batch's global structure. Two supporting pieces carry the scalability claim: neighbor-expansion sub-hypergraph sampling, which allows batch training on large hypergraphs, and an optional node-ID extraction MLP that maps a continuous representation to the top-$k$ most probable nodes. The paper's acceleration comes from removing that MLP and feeding the latent representation directly to the classifier.

What would settle it

Decode the latent-space negatives back to node sets and count how many coincide with observed training hyperedges or with validation/test positives; if a meaningful share of 'negatives' are actually positive hyperedges, the boundary-aware signal is not producing true negatives. Alternatively, train the discriminator using only heuristic negatives and only SEHP-generated negatives, and compare AUROC on the MIX test set: if the heuristic-only model matches or exceeds SEHP's, the generator is not the source of the reported gains.

Watch

Extended reading notes

Core claim

The central claim is that a 'good' negative hyperedge can be defined operationally as a candidate the discriminator currently scores as positive, and that such candidates can be produced by denoising a positive hyperedge representation under the supervision signal $score_t < score_{t+1}$ (equivalently, minimizing $\mathcal{L}_{\mathrm{diff}} = \log(s^{t-1}_i / s^t_i)$). The generator therefore refines a positive embedding into a boundary-hugging negative rather than sampling from noise or corrupting observed hyperedges by hand. The diffusion is conditioned on the average-pooled node embeddings of the current sub-hypergraph, so the produced negative is anchored to the batch's global structure. The efficiency discovery is that the continuous latent representation produced by the denoiser can be fed directly to the classifier, bypassing the node-ID extraction step; the paper reports a 20 to 71-fold training speedup with under one percent average loss in AUROC and Precision.

Load-bearing premise

The method assumes that raising the discriminator's score moves a candidate toward the decision boundary, so the diffusion process trained on that signal yields samples that are true negatives rather than positives in disguise; it also assumes those latent-space representations are valid negative hyperedges without being checked against the observed or future hyperedge sets.

Editorial extensions

If this is right

  • Hyperedge prediction systems can be trained on batches of sub-hypergraphs rather than the whole graph, which removes a central scalability bottleneck for large interaction networks.
  • The boundary-aware loss provides a template for generating hard negatives with diffusion models in other imbalanced classification tasks where the set of true negatives is not well defined.
  • Because the latent-space variant skips discrete decoding, generative negative sampling becomes cheap enough for per-epoch training times of roughly one to two seconds on datasets with over one hundred thousand nodes.
  • The reported robustness from SNS to MNS/CNS test sets suggests that diffusion-generated negatives teach the classifier to separate positives from structurally similar distractors, a property that should transfer to settings where future hyperedges resemble observed ones.

Reading between the lines

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

  • The score-increasing signal is close to an adversarial objective in disguise; a natural extension not explored in the paper is whether boundary-hugging negatives also improve calibration or out-of-distribution rejection, which would make them useful beyond AUROC and Precision.
  • The latent-space shortcut assumes the aggregated embedding space carries the same boundary meaning for any aggregator; the speedup and accuracy claims are therefore likely tied to the max-min pooling aggregator used in the experiments, and may not transfer unchanged to attention-based aggregators.
  • The neighbor-expansion sampling biases generated negatives toward the vicinity of observed hyperedges. On datasets with strong community structure this is helpful, but on datasets where future hyperedges form across communities, the conditioning could omit the true decision region; comparing SEHP on datasets with differing homophily would test this.
  • A direct testable extension is to replace the diffusion denoiser with a non-diffusive MLP that is given the same boundary-aware loss; the paper's ablations use an MLP without that loss, so isolating the diffusion component's contribution to the reported gains remains open.
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

5 major / 5 minor

Summary. SEHP is a proposed framework for hyperedge prediction that generates negative samples by iteratively refining positive hyperedge embeddings with an MLP 'denoising' process, using a 'boundary-aware' loss and sub-hypergraph batching for scalability. The paper also introduces SEHP-epre, a latent-space variant that bypasses discrete node-ID extraction and reports large training speedups. The method is evaluated on six datasets against eight baselines, with AUROC and Precision metrics, ablations, and training-time comparisons.

Significance. If the empirical results were supported by the stated mechanism, SEHP would be a practical contribution: it is one of the first attempts to make negative-sample generation for hyperedge prediction scalable to large hypergraphs, and the latent-space acceleration is a plausible engineering contribution. The paper includes extensive comparisons, ablations, and a code link, which are strengths. However, the central claims about 'diffusion' and 'boundary-aware' refinement are not substantiated by the equations, negative-sample validity is never verified, and several reported numbers are inconsistent. The conceptual contribution therefore rests on unverified or contradicted premises.

major comments (5)
  1. [§4.2.2–§4.3, Abstract] The method described by Eq. (4) and Eq. (5) is not a diffusion model. There is no forward noising process, no noise schedule, no corruption-reconstruction training objective, and no stochastic reverse process. Eq. (4) is simply iterative MLP refinement, and L_diff in Eq. (9) is not a diffusion loss. The abstract and contribution list claim 'uses diffusion models' and 'first diffusion model in hyperedge prediction'; these claims are unsupported. Either implement a proper conditional diffusion process (with a defined noise schedule and a denoising objective) or rename the generator 'iterative adversarial refinement' and adjust the novelty claims accordingly.
  2. [§4.3, Eq. (8)–(10)] The 'boundary-aware' loss does not move samples toward the decision boundary. Minimizing L_gen = L_neg + L_diff increases s^t over t because both L_neg = -1/N Σ s_i and L_diff = log(s^{t-1}/s^t) encourage the discriminator score to increase. Since s_i is a sigmoid probability, increasing it moves representations into the positive-class region, away from the boundary s = 0.5. The initialization h^0_e from a positive hyperedge representation makes this more severe: the refinement starts from a positive example and pushes it toward even higher positive scores. The claim that this 'moves closer to the decision boundary' is contradicted by the equations. A true boundary objective (e.g., |s - 0.5|) or a clear redefinition of 'boundary' is needed.
  3. [§4.2.4, §4.4, Eq. (7)] Negative-sample validity is never verified. The node sets from Eq. (6) are not filtered against the observed hyperedge set E or the future set E', and the SEHP-epre variant uses arbitrary latent vectors as 'negative hyperedge embeddings' without checking that they lie on the manifold of aggregator outputs. The paper itself states in §4.2 that 'defining an effective negative sample' is an open challenge, but it never validates that the generated samples satisfy any such definition. Consequently, L_dis in Eq. (7) may be trained with false negatives or out-of-distribution features, and the gains in Tables 1–4 cannot be unambiguously attributed to better negative sampling. The authors should report overlap rates with E and E', and compare generated latent vectors to actual aggregator outputs.
  4. [Abstract, §4.4, §5.2.4, Table 3] The claim of 'under 1% average metric loss' for SEHP-epre relative to SEHP is contradicted by Table 3. On Cora, SEHP-epre has average AUROC 0.882 vs. SEHP's 0.908 (a relative drop of ~2.9%) and average Precision 0.672 vs. 0.703 (a drop of ~4.4%). On NDC_class the AUROC drop is ~3.4%. These are not 'negligible' or 'under 1%' losses. Please recompute the average loss or qualify the claim with the actual measured ranges.
  5. [Table 2 vs. Table 4] The reported results for the same model and dataset are inconsistent. Table 2 gives SEHP on Recipe200k as AUROC 0.9179 and Precision 0.6734, while Table 4 gives SEHP on Recipe200k as AUROC 0.9749 and Precision 0.7786. Recipe100k values match between the two tables, but the Recipe200k discrepancy is large and undermines the reproducibility of the main results. Please reconcile these numbers and report the exact evaluation protocol for each table.
minor comments (5)
  1. [§5.2] There is a typo 'seperately' that should read 'separately'.
  2. [§5.1, Tables 1–5] The hyperparameters used in the experiments are not reported: number of denoising refinement steps T, sub-hypergraph batch size, neighbor-expansion depth, learning rate, and the negative-to-positive sample ratio are all absent from the main text. Without these, the reported speed and accuracy results cannot be reproduced.
  3. [§5.2.3, Table 3] The text states that SEHP scores 0.908 on Cora, but does not clarify that this is the average AUROC across the four test sets. The table includes both per-set and average columns; the text should be explicit.
  4. [Figure 2] The caption refers to a 'purple arrow' but the figure is likely printed in black-and-white; please use a visible marker or describe the arrow position in text.
  5. [Tables 1–4] No standard deviations or significance tests are reported. Given the magnitude of some claimed improvements, the authors should report variance across multiple runs or justify why deterministic training makes this unnecessary.

Circularity Check

1 steps flagged · score 4.0 of 10

The training-time definition of a good negative sample is the discriminator's own score-increase signal, making the 'boundary-aware' mechanism self-definitional; final external test negatives keep the headline AUROC/Precision claims partially independent.

  1. self definitional [Section 4.2.2 (Eq. 4), Section 4.3 (Eqs. 8-10)]
    "However, in hyperedge prediction, there is no clear notion of what constitutes a good negative sample, so there is no original sample to start with. Instead, the score predicted by the classifier can be used as a signal to supervise the denoising process. ... The signal for denoising is score_t_hyperedge < score_{t+1}_hyperedge. ... The second component, L_diff, which is the boundary-aware loss function, ensures that the generated negative samples move closer to the decision boundary through the iterative denoising process: L_diff = log(s^{t-1}_i / s^t_i)."

    The paper explicitly says there is no clear notion of a good negative sample, then defines that notion as 'denoising increases the classifier score' (Eq. 4 and the score_t < score_{t+1} signal). The generator losses L_neg = -1/N sum s_i and L_diff = log(s^{t-1}/s^t) both push the discriminator's positive-class probability upward; no term in Eq. 10 measures distance to a decision boundary. Thus the claimed property 'moves closer to the decision boundary' is not derived from an independent boundary geometry; it is, by construction, a renaming of the score-increasing objective. The discriminator is simultaneously trained to label these same generated representations as y=0 in Eq. 7, so the training-time notion of 'negative' is endogenous.

full rationale

The paper's headline empirical claims are not circular: AUROC and Precision are computed on fixed held-out negative sets (SNS, MNS, CNS, MIX) that are generated by external heuristics, and the 20-71x speedup is a direct timing comparison of SEHP versus SEHP-epre, not a fitted parameter relabeled as a prediction. The self-citations in the reference list (e.g., prior hypergraph work by the Yin group) are motivational and not load-bearing. The one genuine definitional loop is in the generator's training criterion: Section 4.2.2 states that no clear notion of a good negative sample exists, then adopts the discriminator's score as the denoising signal, and Section 4.3's L_gen minimizes only functions of that same score (Eqs. 8-9). Calling this 'boundary-aware' is therefore an assumption encoded in the loss rather than a consequence of the equations. Because the final evaluation is anchored to external test negatives, the central accuracy and efficiency results retain independent content; the circularity burden is moderate rather than total.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new ontological entities are introduced; the structural condition and generated representations are standard learned vectors. The central claim rests on a small set of unreported hyperparameters and on the assumption that classifier-guided refinement produces valid negatives.

free parameters (5)
  • Embedding dimension and MLP hidden sizes = not reported
    Architecture choices for the encoder, denoiser, node-ID MLP, and classifier are not fully specified; performance likely depends on them.
  • Number of denoising refinement steps T = not reported
    Eq. 4 and Eq. 5 describe iterative refinement but no step count or convergence criterion is given.
  • Sub-hypergraph batch size and neighbor-expansion depth = not reported
    These control the structural condition h_sub_structure and scalability; no values or sensitivity analysis are provided.
  • Loss mixing weight between L_neg and L_diff = not reported
    L_gen is stated as a sum with implicit equal weighting; no ablation of alternative weights is reported.
  • Negative-to-positive sample ratio = not reported
    The paper does not state how many generated negatives are used per positive during training, although this ratio is known to affect hyperedge prediction quality.
assumptions (4)
  • domain assumption Test and training negatives produced by SNS, MNS, and CNS are true negatives not in E or the future set E'.
    If random negatives overlap future positives, AUROC and Precision are biased; the paper does not filter or discuss this in Section 5.1.
  • ad hoc to paper The classifier score s_e is a valid proxy for sample hardness or distance to a decision boundary.
    The generator's entire objective in Eq. 8 and Eq. 9 is defined through the discriminator's own scores; no calibration to an external boundary is provided.
  • domain assumption Neighbor-based sub-hypergraph sampling preserves the global structure needed for conditional denoising.
    The expansion by shared nodes in Section 4.2.1 is assumed to capture global patterns; no formal guarantee or sensitivity analysis is given.
  • ad hoc to paper Continuous latent-space denoising yields embeddings that behave like aggregator outputs for real node sets.
    The accelerated variant feeds generated vectors directly to the classifier in Section 4.4, yet no analysis shows these vectors lie on the same manifold as pooled node embeddings.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Scalable and Effective Negative Sample Generation for Hyperedge Prediction." pith.science (2026). https://pith.science/paper/GZH4VLGM

@misc{pith2026241112354,
  author       = {Pith},
  title        = {Pith review of: Scalable and Effective Negative Sample Generation for Hyperedge Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GZH4VLGM}},
  note         = {Machine review of arXiv:2411.12354}
}
read the original abstract

Hyperedge prediction is crucial in hypergraph analysis for understanding complex multi-entity interactions in various web-based applications, including social networks and e-commerce systems. Traditional methods often face difficulties in generating high-quality negative samples due to the imbalance between positive and negative instances. To address this, we present the Scalable and Effective Negative Sample Generation for Hyperedge Prediction (SEHP) framework, which utilizes diffusion models to tackle these challenges. SEHP employs a boundary-aware loss function that iteratively refines negative samples, moving them closer to decision boundaries to improve classification performance. SEHP samples positive instances to form sub-hypergraphs for scalable batch processing. By using structural information from sub-hypergraphs as conditions within the diffusion process, SEHP effectively captures global patterns. To enhance efficiency, our approach operates directly in latent space, avoiding the need for discrete ID generation and resulting in significant speed improvements while preserving accuracy. Extensive experiments show that SEHP outperforms existing methods in accuracy, efficiency, and scalability, representing a substantial advancement in hyperedge prediction techniques. Our code is available here.

Figures

Figures reproduced from arXiv: 2411.12354 by the authors.

Figure 1
Figure 1. Overview of the Conditional Diffusion-Based Framework for Negative Hyperedge Generation in Hyperedge Prediction. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Framework of the conditional diffusion-based negative hyperedge generation for hyperedge prediction using generated [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 48 canonical work pages

  1. [1]

    Devanshu Arya, Deepak K Gupta, Stevan Rudinac, and Marcel Worring. 2020. Hypersage: Generalizing inductive representation learning on hypergraphs.arXiv preprint arXiv:2010.04558 (2020)

    arXiv 2020

  2. [2]

    Devanshu Arya and Marcel Worring. 2018. Exploiting Relational Information in Social Networks using Geometric Deep Learning on Hypergraphs. In Proceedings of the 2018 ACM on International Conference on Multimedia Retrieval (Yokohama, Japan) (ICMR ’18). ACM, 117–125

    work page 2018

  3. [3]

    Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. 2021. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems 34 (2021), 17981–17993

    2021

  4. [4]

    Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu. 2023. One Transformer Fits All Distributions in Multi-Modal Diffusion at Scale. In Proceedings of the 40th International Conference on Machine Learning, Andreas Krause (Ed.), Vol. 202. PMLR, 1692–1717

    work page 2023

  5. [5]

    Benson, Rediet Abebe, Michael T

    Austin R. Benson, Rediet Abebe, Michael T. Schaub, Ali Jadbabaie, and Jon Klein- berg. 2018. Simplicial closure and higher-order link prediction. Proceedings of the National Academy of Sciences (2018)

    work page 2018

  6. [6]

    Bhuiyan, Mahmudur Rahman, Mahmuda Rahman, and Mohammad Al Hasan

    Mansurul A. Bhuiyan, Mahmudur Rahman, Mahmuda Rahman, and Mohammad Al Hasan. 2012. GUISE: Uniform Sampling of Graphlets for Large Graph Analysis. In Proceedings of the 2012 IEEE 12th International Conference on Data Mining (ICDM ’12). 91–100

    work page 2012

  7. [7]

    Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. 2023. DiffusionDet: Diffu- sion Model for Object Detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) . 19830–19843

    work page 2023

  8. [8]

    Dian Cheng, Jiawei Chen, Wenjun Peng, Wenqin Ye, Fuyu Lv, Tao Zhuang, Xiaoyi Zeng, and Xiangnan He. 2022. Ihgnn: Interactive hypergraph neural network for personalized product search. In Proceedings of the ACM Web Conference 2022 . 256–265

    work page 2022

Show all 61 references
  1. [9]

    Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion Models Beat GANs on Image Synthesis. In Advances in Neural Information Processing Systems , M. Ran- zato (Ed.), Vol. 34. Curran Associates, Inc., 8780–8794

    2021

  2. [10]

    Abhishek Dixit and Ashish Mani. 2023. Sampling technique for noisy and bor- derline examples problem in imbalanced classification. Applied Soft Computing 142 (2023), 110361

    2023

  3. [11]

    Yifan Feng, Haoxuan You, Zizhao Zhang, Rongrong Ji, and Yue Gao. 2019. Hy- pergraph neural networks. In Proceedings of the AAAI conference on artificial intelligence, Vol. 33. 3558–3565

    2019

  4. [12]

    Johannes Gasteiger, Stefan Weißenberger, and Stephan Günnemann. 2019. Diffu- sion improves graph learning. Advances in neural information processing systems 32 (2019)

    2019

  5. [13]

    Zhangxuan Gu, Haoxing Chen, and Zhuoer Xu. 2024. Diffusioninst: Diffusion Model for Instance Segmentation. In ICASSP 2024 - 2024 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) , Vol. 1. 2730–2734

    2024

  6. [14]

    Ishaan Gulrajani and Tatsunori B Hashimoto. 2023. Likelihood-Based Diffusion Language Models. In Advances in Neural Information Processing Systems , A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (Eds.), Vol. 36. Curran Associates, Inc., 16693–16715

    2023

  7. [15]

    Lei Guo, Hongzhi Yin, Tong Chen, Xiangliang Zhang, and Kai Zheng. 2021. Hierarchical Hyperedge Embedding-Based Representation Learning for Group Recommendation. ACM Trans. Inf. Syst. 40, 1, Article 3 (sep 2021), 27 pages

    2021

  8. [16]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. Advances in neural information processing systems 30 (2017)

    2017

  9. [17]

    Jiale Han, Bo Cheng, and Xu Wang. 2020. Open domain question answering based on text enhanced knowledge graph with hyperedge infusion. In Findings of the Association for Computational Linguistics: EMNLP 2020 . 1475–1481

    2020

  10. [18]

    Yan Han, Edward W Huang, Wenqing Zheng, Nikhil Rao, Zhangyang Wang, and Karthik Subbian. 2023. Search behavior prediction: A hypergraph perspective. In Proceedings of the sixteenth acm international conference on web search and data mining. 697–705

    2023

  11. [19]

    Zhengfu He, Tianxiang Sun, Qiong Tang, Kuanning Wang, Xuanjing Huang, and Xipeng Qiu. 2023. DiffusionBERT: Improving Generative Masked Language Models with Diffusion Models. In Proceedings of the 61st Annual Meeting of the ACL, Anna Rogers (Ed.). Toronto, Canada, 4521–4534

    2023

  12. [20]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. In Advances in Neural Information Processing Systems , H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (Eds.), Vol. 33. Curran Associates, Inc., 6840–6851

    2020

  13. [21]

    Jie Huang, Xin Liu, and Yangqiu Song. 2019. Hyper-Path-Based Representation Learning for Hyper-Networks. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (Beijing, China) (CIKM ’19). ACM, 449–458

    2019

  14. [22]

    Jing Huang and Jie Yang. 2021. UniGNN: a Unified Framework for Graph and Hypergraph Neural Networks. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, Zhi-Hua Zhou (Ed.). IJCAI-21, 2563–2569

    2021

  15. [23]

    Hyunjin Hwang, Seungwoo Lee, Chanyoung Park, and Kijung Shin. 2022. AHP: Learning to Negative Sample for Hyperedge Prediction. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (Madrid,Spain) (SIGIR ’22). 2237–2242

    2022

  16. [24]

    Zhixuan Jia, Yushun Fan, Chunyu Wei, and Ruyu Yan. 2023. A spatial–temporal hypergraph based method for service recommendation in the Mobile Internet of Things-enabled service platform. Advanced Engineering Informatics 57 (2023), 102038

    2023

  17. [25]

    Dong Li, Zhiming Xu, Sheng Li, and Xin Sun. 2013. Link prediction in social networks based on hypergraph. InProceedings of the 22nd International Conference on World Wide Web (Rio de Janeiro, Brazil) (WWW ’13 Companion). ACM, 41–42

    2013

  18. [26]

    Shuyang Li, Yufei Li, Jianmo Ni, and Julian McAuley. 2022. SHARE: a System for Hierarchical Assistive Recipe Editing. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing . 11077–11090

    2022

  19. [27]

    Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. 2022. Diffusion-LM Improves Controllable Text Generation. In Ad- vances in Neural Information Processing Systems , Vol. 35. Curran Associates, Inc., 4328–4343

    2022

  20. [28]

    Zihao Li, Aixin Sun, and Chenliang Li. 2023. DiffuRec: A Diffusion Model for Sequential Recommendation. ACM Trans. 42, 3, Article 66 (dec 2023), 28 pages

    2023

  21. [29]

    Lichtenwalter, Jake T

    Ryan N. Lichtenwalter, Jake T. Lussier, and Nitesh V. Chawla. 2010. New perspec- tives and methods in link prediction. In Proceedings of the 16th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (Washington, USA) (KDD ’10). ACM, 243–252

    2010

  22. [30]

    Jianghao Lin, Weiwen Liu, Xinyi Dai, Weinan Zhang, Shuai Li, Ruiming Tang, Xiuqiang He, Jianye Hao, and Yong Yu. 2021. A graph-enhanced click model for web search. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieva...

    2021

  23. [31]

    Jiaying Liu, Jing Ren, Wenqing Zheng, Lianhua Chi, Ivan Lee, and Feng Xia

  24. [32]

    Xiao Liu, Shunmei Meng, Qianmu Li, Xiaolong Xu, Lianyong Qi, Wanchun Dou, Jing Zhang, and Xuyun Zhang. 2023. Disentangled Hypergraph Collaborative Filtering for Social Recommendation. In 2023 IEEE International Conference on Web Services (ICWS). IEEE, 475–482

    2023

  25. [33]

    Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. 2013. Rectifier nonlineari- ties improve neural network acoustic models. In Proc. icml, Vol. 30. Atlanta, GA, 3

    2013

  26. [34]

    David JC MacKay. 1992. Information-based objective functions for active data selection. Neural computation 4, 4 (1992), 590–604

    1992

  27. [35]

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems 26 (2013)

    2013

  28. [36]

    Bronstein, and Xavier Bresson

    Federico Monti, Michael M. Bronstein, and Xavier Bresson. 2017. Geometric matrix completion with recurrent multi-graph neural networks. In Proceedings of the 31st International Conference on Neural Information Processing Systems (Long Beach, California, USA) (NIPS’17). Curran ...

    2017

  29. [37]

    Alexander Quinn Nichol and Prafulla Dhariwal. 2021. Improved Denoising Diffusion Probabilistic Models. In Proceedings of the 38th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 139) , Marina Meila and Tong Zhang (Eds.). PMLR, 8162–8171

    2021

  30. [38]

    Chenhao Niu, Yang Song, Jiaming Song, Shengjia Zhao, Aditya Grover, and Stefano Ermon. 2020. Permutation invariant graph generation via score-based generative modeling. In International Conference on Artificial Intelligence and Statistics. PMLR, 4474–4484

    2020

  31. [39]

    Narasimha Murty

    Prasanna Patil, Govind Sharma, and M. Narasimha Murty. 2020. Negative Sam- pling for Hyperlink Prediction in Networks. In Advances in Knowledge Discovery and Data Mining: 24th Pacific-Asia Conference, PAKDD 2020, Singapore (Singapore, Singapore). Springer-Verlag, Berlin, Heide...

    2020

  32. [40]

    Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. 2022. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings . 1–10

    2022

  33. [41]

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. 2022. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural infor...

    2022

  34. [42]

    Yuanhang Su and C.-C. Jay Kuo. 2022. Recurrent Neural Networks and Their Memory Behavior: A Survey. APSIPA Transactions on Signal and Information Processing 11, 1 (2022), –

    2022

  35. [43]

    Xiangguo Sun, Hong Cheng, Bo Liu, Jia Li, Hongyang Chen, Guandong Xu, and Hongzhi Yin. 2023. Self-supervised hypergraph representation learning for sociological analysis. IEEE Transactions on Knowledge and Data Engineering 35, 11 (2023), 11860–11871

    2023

  36. [44]

    Xiangguo Sun, Hongzhi Yin, Bo Liu, Hongxu Chen, Qing Meng, Wang Han, and Jiuxin Cao. 2021. Multi-level hyperedge distillation for social linking prediction on sparsely observed networks. In Proceedings of the Web Conference 2021 . 2934– 2945

    2021

  37. [45]

    Ke Tu, Peng Cui, Xiao Wang, Fei Wang, and Wenwu Zhu. 2018. Structural deep embedding for hyper-networks. In Proceedings of the AAAI conference on artificial intelligence, Vol. 32

    2018

  38. [46]

    Changlin Wan, Muhan Zhang, Wei Hao, Sha Cao, Pan Li, and Chi Zhang. 2021. Principled hyperedge prediction with structural spectral features and neural networks. arXiv preprint arXiv:2106.04292 (2021)

    2021 arXiv

  39. [47]

    Peihao Wang, Shenghao Yang, Yunyu Liu, Zhangyang Wang, and Pan Li. 2023. Equivariant Hypergraph Diffusion Neural Operators. In International Conference on Learning Representations (ICLR)

    2023

  40. [48]

    Hanrui Wu, Yuguang Yan, and Michael Kwok-Po Ng. 2022. Hypergraph collabo- rative network on vertices and hyperedges. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 3 (2022), 3245–3258

    2022

  41. [49]

    Lei Wu, Zhen Yang, Shijun Liu, and Minggang He. 2018. An Approach to Web Service Organization Based on Hypergraph Clustering. In2018 IEEE 22nd Interna- tional Conference on Computer Supported Cooperative Work in Design ((CSCWD)) . IEEE, 87–91

    2018

  42. [50]

    Naganand Yadati, Madhav Nimishakavi, Prateek Yadav, Vikram Nitin, Anand Louis, and Partha Talukdar. 2019. HyperGCN: a new method of training graph convolutional networks on hypergraphs . Curran Associates Inc., Red Hook, USA

    2019

  43. [51]

    Naganand Yadati, Vikram Nitin, Madhav Nimishakavi, Prateek Yadav, Anand Louis, and Partha Talukdar. 2020. NHP: Neural Hypergraph Link Prediction. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (Virtual Event, Ireland) (CIKM ’20). ...

    2020

  44. [52]

    Hamdi Yahyaoui, Mohamed Almulla, and Eiman Boujarwah. 2021. Measuring semantic similarity between services using hypergraphs. InThe 23rd International Conference on Information Integration and Web Intelligence . 205–211

    2021

  45. [53]

    Jielong Yan, Yifan Feng, Shihui Ying, and Yue Gao. 2024. Hypergraph Dynamic System. In International Conference on Learning Representations (ICLR)

    2024

  46. [54]

    Yuhao Yang, Chao Huang, Lianghao Xia, Yuxuan Liang, Yanwei Yu, and Chen- liang Li. 2022. Multi-behavior hypergraph-enhanced transformer for sequential recommendation. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining . 2263–2274

    2022

  47. [55]

    Se-eun Yoon, Hyungseok Song, Kijung Shin, and Yung Yi. 2020. How Much and When Do We Need Higher-order Information in Hypergraphs? A Case Study on Hyperedge Prediction. In Proceedings of The Web Conference 2020 (Taipei, Taiwan) (WWW ’20). ACM, 2627–2633

    2020

  48. [56]

    Chia-An Yu, Ching-Lun Tai, Tak-Shing Chan, and Yi-Hsuan Yang. 2018. Modeling Multi-way Relations with Hypergraph Embedding. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management(Torino, Italy) (CIKM ’18). ACM, 1707–1710

    2018

  49. [57]

    Junliang Yu, Hongzhi Yin, Jundong Li, Qinyong Wang, Nguyen Quoc Viet Hung, and Xiangliang Zhang. 2021. Self-supervised multi-channel hypergraph convolu- tional network for social recommendation. In Proceedings of the web conference

    2021

  50. [58]

    Junwei Zhang, Min Gao, Junliang Yu, Lei Guo, Jundong Li, and Hongzhi Yin. 2021. Double-scale self-supervised hypergraph learning for group recommendation. In Proceedings of the 30th ACM international conference on information & knowledge management. 2557–2567

    2021

  51. [59]

    Muhan Zhang, Zhicheng Cui, Shali Jiang, and Yixin Chen. 2018. Beyond link prediction: predicting hyperlinks in adjacency space. In Proceedings of the Thirty- Second AAAI Conference on Artificial Intelligence (New Orleans, Louisiana, USA) (AAAI’18). Article 543, 8 pages

    2018

  52. [60]

    Ruochi Zhang, Yuesong Zou, and Jian Ma. 2020. Hyper-SAGNN: a self-attention based graph neural network for hypergraphs. In International Conference on Learning Representations (ICLR). Scalable and Effective Negative Sample Generation for Hyperedge Prediction WWW ’2025, 28 Apri...

    2020

  53. [2020]

    In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval

    Web of scholars: A scholar knowledge graph. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. 2153–2156

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.