Pith. sign in

REVIEW 4 major objections 7 minor 2 cited by

Hyperbolic Hypergraph Neural Networks for Multi-Relational Knowledge Hypergraph Representation

T0 review · 4 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read H2GNN: a hyperbolic hypergraph encoder that directly models multi-relational knowledge hypergraphs and beats 15 existing methods.

desk verdict Useful encoder idea with a promising ablation, but the SOTA claim rests on an underspecified link-prediction protocol; worth reviewing after the authors disclose filtered/raw ranking, curvature, splits, and baseline provenance. read the letter →

arxiv 2412.12158 v1 pith:UDMOXNLI submitted 2024-12-11 cs.LG cs.AI

Mengfan Li , Xuanhua Shi , Chenqi Qiao , Teng Zhang , Hai Jin This is my paper · ORCID
classification cs.LGcs.AI
keywords knowledgehypergraphneuralnetworkhyperbolicspaceLorentzmessagepassingposition-awareembeddinglinkpredictionnodeclassification
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

Knowledge hypergraphs record facts as tuples (relation, entity1, ..., entitym), and existing methods either break such tuples into binary triples or learn each hyperedge in isolation, discarding information about how hyperedges share entities and about the order of entities inside a tuple. This paper proposes H2GNN, a hyperbolic hypergraph neural network whose hyper-star message passing expands hyperedges into position-labelled hierarchies and then aggregates in two stages, first from nodes to hyperedges and then back to nodes using hyper-relation and position-aware embeddings. The paper's claim is that this direct, structure-based encoding preserves both adjacency and entity roles, and that hyperbolic (Lorentz) space fits the tree-like structure of knowledge hypergraphs better than Euclidean space. It reports that H2GNN outperforms 15 baselines on node classification across four citation and co-authorship datasets and on link prediction across two knowledge bases, with ablations attributing the gain to the combination of hyperbolic operations and position-aware composition. The result matters because it suggests GNN-style structure learning can replace pairwise decomposition for n-ary relational data.

What carries the argument

The central mechanism is the hyper-star message-passing scheme, a two-stage aggregation designed for multi-relational knowledge hypergraphs. In the first stage, nodes of a hyperedge are aggregated by a Lorentz centroid into a hyperedge embedding; in the second, each node is updated by a centroid composition of its own embedding, the hyperedge embedding, the hyper-relation embedding, and a position-aware embedding indexed by relation type and entity position (relations like Roster-1, Roster-2, Roster-3). The position expansion is what turns a hyperedge into a tree-like hierarchy, and the Lorentz centroid is the operation that keeps the whole procedure fully hyperbolic; the linear transformation used to map embeddings between hyperboloid layers follows a fully hyperbolic construction that guarantees outputs remain on the hyperboloid.

What would settle it

Reproduce H2GNN and all 15 baselines in one codebase with identical filtered/raw ranking, identical negative samples, and identical train/validation/test splits, and check whether the Hits@1, Hits@10, MRR, and accuracy differences survive; if they shrink to within one standard deviation, the central claim is refuted.

Watch

Extended reading notes

Core claim

The central discovery claimed by the paper is that a hypergraph neural network can learn multi-relational knowledge hypergraph representations without decomposing hyperedges into binary relations or treating them as isolated. H2GNN expands each hyperedge (r, x1, ..., xm) into a position-labelled tree by creating relations r-1, ..., r-m, then performs two-stage message passing in a Lorentz space with negative curvature: centroid aggregation over the nodes of each hyperedge yields a hyperedge embedding, and a second centroid composition of the node, hyperedge, hyper-relation, and position-aware embeddings updates the node. Because the centroids and linear transformations are defined entirely inside the hyperboloid, the model avoids the common hybrid that maps vectors back and forth to the tangent space. The reported experiments show H2GNN ahead of every compared method on node classification and link prediction, and the ablation shows the largest drop when position-aware composition is removed, which the paper reads as evidence that entity position semantics matter in multi-relational knowledge hypergraphs.

Load-bearing premise

The performance advantage assumes that H2GNN's training and evaluation setup — ranking filter, negative samples, data splits, and hyperparameter choices — is directly comparable to the setup that produced the baseline numbers it is measured against.

Editorial extensions

If this is right

  • If the reported results are correct, knowledge hypergraph representation learning no longer needs to flatten n-ary facts into binary triples, so high-order structure and entity-order semantics can be used directly in downstream reasoning.
  • The encoder is plug-and-play: pairing H2GNN with decoders like HSimplE, mTransH, or m-DistMult improves over using those decoders alone, suggesting that structural encoding and task-specific scoring can be separated.
  • Position-aware expansion gives entity order a concrete role in the embedding, which should help in domains where tuple positions encode roles such as origin, intermediate, and destination.
  • Fully hyperbolic operations suggest a lower-dimensional representation can capture the scale-free, hierarchical structure of knowledge hypergraphs without the cost of tangent-space projections.
  • Inductive results on evolving hypergraphs indicate the model can assign embeddings to previously unseen nodes from the historical hypergraph structure, not just classify or rank seen nodes.

Reading between the lines

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

  • Going beyond the paper, the same position-aware expansion could be applied to any n-ary relational data where argument order matters, such as biomedical event tuples or temporal fact logs, even if the data is not organized as a knowledge base.
  • The paper leaves curvature as a fixed negative scalar; a natural next test is to learn or search the curvature per dataset, since flatter or noisier graphs might favor a different geometry.
  • If the evaluation protocol (filtered versus raw ranking, negative sample generation, validation splits) turns out to differ from the baselines' protocols, the numerical lead would need to be re-estimated; the architectural benefits could still stand but the headline numbers might not.
  • One could stress-test the claim by swapping the simple centroid aggregator for attention-based or relation-specific aggregators in the same fully hyperbolic framework to see whether position-aware composition remains the key factor.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes H2GNN, a hyperbolic hypergraph neural network for multi-relational knowledge hypergraphs. The core idea is a hyper-star message-passing scheme that expands each hyperedge into position-specific relation slots and aggregates information in the Lorentz space via centroid operations, thereby incorporating entity positions, hyper-relation embeddings, and adjacency structure. The authors claim that H2GNN outperforms 15 baselines on node classification (co-authorship/co-citation datasets) and link prediction (JF17K and FB-AUTO knowledge hypergraphs). The node classification experiments use homogeneous hypergraphs, while the link prediction experiments are the only multi-relational evaluations.

Significance. If the empirical claims hold, H2GNN is a meaningful contribution: it is a fully hyperbolic, instance-dependent encoder for n-ary knowledge facts that avoids the information loss of clique/star decompositions and does not require per-relation hypergraph splitting. The reported gains over HypE and G-MPNN on link prediction are potentially valuable. The method is conceptually clean, with a simple centroid-based composition that is fully in hyperbolic space. However, the significance is conditional on the reproducibility of the link prediction numbers, because the evaluation protocol is not fully disclosed and the multi-relational claim rests entirely on that experiment.

major comments (4)
  1. [Section 4.1, Section 4.3, Table 4] The link prediction evaluation protocol is underspecified in a way that is load-bearing for the central claim. The paper does not state whether Hits@K and MRR are computed under filtered or raw ranking; it does not report the curvature value k used in the Lorentz space; it does not describe how JF17K and FB-AUTO are split into train/validation/test; and it does not state how many random seeds or runs produce the reported numbers. Because the baseline entries appear to be taken from previously published results (e.g., HypE), any mismatch in protocol—particularly filtered versus raw ranking or a different validation split—could inflate H2GNN's apparent gains. The authors must disclose the exact protocol and, ideally, rerun all baselines under the identical protocol to support the 'outperforms' claim.
  2. [Section 4.3, Table 4] The comparison is not actually against 15 baselines on the multi-relational link prediction task. Table 4 lists 10 baselines on FB-AUTO and only 9 on JF17K, because G-MPNN is omitted on JF17K after a two-day timeout. The abstract claims comparison with 15 baselines, but on the dataset that matters most for the multi-relational claim, the comparison is incomplete. The authors should either run G-MPNN on JF17K with a bounded time budget and report its results, or explicitly revise the claim to state that G-MPNN was excluded from JF17K and that the comparison is against 14 full baselines plus a partial comparison.
  3. [Section 6 vs. Abstract] There is a direct inconsistency between the abstract and the conclusion. The abstract states H2GNN 'outperforms state-of-the-art approaches' on both tasks, while Section 6 says the encoder 'yields results comparable to the baselines on knowledge hypergraph link prediction and node classification tasks.' These are different claims. The authors should align the language: if the improvements are numerically consistent but not statistically significant, 'comparable' is appropriate; if they are significant improvements, the conclusion should say so. As written, the reader cannot tell which claim is intended.
  4. [Section 4.2, Tables 2 and 3] The node classification experiments use homogeneous co-authorship and co-citation hypergraphs, which contain no relation types and no entity positions. Therefore, these experiments do not exercise the position-aware, multi-relational message passing that distinguishes H2GNN from UniGNN-family methods. The only multi-relational evidence is the link prediction table, which makes the protocol disclosure in the first major comment essential: if the link prediction evaluation is not reproducible, the paper's central novelty claim is unsupported.
minor comments (7)
  1. [Section 3.2, Eq. (5)] Equation (5) contains an extra closing parenthesis: 'xi = centroid(xi, he, re, hp)e∈Ei )' should be corrected, and the notation should make clear that the centroid is taken over the set of neighboring hyperedges with their associated relation and position embeddings.
  2. [Section 3.1] The phrase 'we employ the aggregation operation operation' has a duplicated word; change to 'we employ the aggregation operation'.
  3. [Section 4.4] The text says 'massage-passing process' where 'message-passing' is meant; fix this typo in the ablation subsection.
  4. [Title] The arXiv title reads 'MULTI-R ELATIONAL' with an erroneous space; it should be 'MULTI-RELATIONAL'.
  5. [References] References [12] and [28] are the same work (Fatemi et al., 'Knowledge hypergraphs: Prediction beyond binary relations'); deduplicate them and use consistent citation numbering.
  6. [Section 3.3, Eq. (7)] The text says negative samples are generated by replacing each correct entity with N other entities, creating N*r negative samples per positive sample, but Section 4.1 reports only a 'negative ratio of 10.' Clarify whether N is the negative ratio or the number of negatives per position, and specify the value used in the experiments.
  7. [Figure 6] The caption does not define 'HO', 'CO', or the label 'H²GNN E'; add definitions in the caption or in the main text so the ablation results are self-contained.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: H2GNN's derivation is self-contained and its components are drawn from external prior work; the empirical claims are evaluated on held-out data, with protocol concerns that are correctness issues, not circularity.

full rationale

I found no circular derivation in this paper. The core components are stated with explicit formulas: the Lorentz linear transformation and centroid aggregation are directly cited to external prior work ([21] Fully Hyperbolic Neural Networks and [26] Lorentzian Distance Learning), none of which is authored by the present authors, so there is no self-citation chain. The position-aware composition operation comp(he, r, hp) = centroid(he, r, hp) is a non-parametric aggregation over learned entity, relation, and position embeddings; no parameter is fitted to a target metric and then reported as a prediction. Node classification and link prediction are evaluated on test data against external baselines, so the claimed advantages are empirical rather than consequences of definitions. The two most notable concerns are not circularity: the link-prediction protocol is under-specified (no filtered/raw ranking statement, no curvature value, no exact data splits), and the paper itself notes that G-MPNN timed out on JF17K while the conclusion says results are 'comparable' to baselines. These are reproducibility and consistency issues, not cases where a derived quantity equals an input by construction. Therefore the appropriate score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. The main free parameters are the learned position-aware embeddings and the unstated curvature k. The axioms are standard hyperbolic geometry results plus the domain assumptions that entity order is meaningful and that hyperbolic space helps. The 'lossless expansion' claim is an unproven assumption specific to this paper.

free parameters (3)
  • Position-aware embedding hp (per relation type and position index) = learned; dimension equals hidden size (8 for node classification, 200 for link prediction)
    Introduced in Section 3.2 to encode entity position within a hyperedge; shared across instances of the same relation-position pair; central to the composition operation and to the claimed improvement.
  • Lorentz space curvature k = not reported
    The paper defines operations in H^n_k but never states the value of k used in experiments or whether it is fixed or learned. This is a load-bearing hyperparameter for all hyperbolic operations.
  • Linear maps Wh, Wr, Wp and biases = learned during training
    Used in Figure 4 and Section 3.2 to transform hyperedge, relation, and position embeddings before composition; standard trainable parameters but unstated initialization and whether bias is included.
assumptions (5)
  • standard math The exponential and logarithmic maps plus the Lorentz linear transformation Fx(M) from [21] map correctly between tangent space and hyperboloid, preserving hyperbolicity.
    Invoked in Section 3.1 to justify that all operations stay in hyperbolic space; taken as proven in [21].
  • standard math The centroid operation defined in Section 3.1 is a valid aggregation in Lorentz space that minimizes squared Lorentzian distance.
    Used as the composition and aggregation operator in Eq. 4-5; follows [26].
  • domain assumption Entity order in a knowledge hypergraph tuple is semantically meaningful and can be encoded by position-specific relation variants (r-1, ..., r-m).
    Core motivation in Section 2; without this, the position-aware expansion and composition have no benefit.
  • domain assumption Knowledge hypergraphs or their expanded hierarchies are tree-like enough that hyperbolic space is beneficial.
    Motivation in Section 2 and Figure 2; the paper provides visualization but no quantitative evidence of hyperbolicity.
  • ad hoc to paper The star expansion of hyperedges into position-specific relations is lossless.
    Claimed in the abstract and Section 2 without formal definition or proof; the subsequent centroid aggregation is lossy, so the 'lossless' premise is not established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hyperbolic Hypergraph Neural Networks for Multi-Relational Knowledge Hypergraph Representation." pith.science (2026). https://pith.science/paper/UDMOXNLI

@misc{pith2026241212158,
  author       = {Pith},
  title        = {Pith review of: Hyperbolic Hypergraph Neural Networks for Multi-Relational Knowledge Hypergraph Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UDMOXNLI}},
  note         = {Machine review of arXiv:2412.12158}
}
read the original abstract

Knowledge hypergraphs generalize knowledge graphs using hyperedges to connect multiple entities and depict complicated relations. Existing methods either transform hyperedges into an easier-to-handle set of binary relations or view hyperedges as isolated and ignore their adjacencies. Both approaches have information loss and may potentially lead to the creation of sub-optimal models. To fix these issues, we propose the Hyperbolic Hypergraph Neural Network (H2GNN), whose essential component is the hyper-star message passing, a novel scheme motivated by a lossless expansion of hyperedges into hierarchies. It implements a direct embedding that consciously incorporates adjacent entities, hyper-relations, and entity position-aware information. As the name suggests, H2GNN operates in the hyperbolic space, which is more adept at capturing the tree-like hierarchy. We compare H2GNN with 15 baselines on knowledge hypergraphs, and it outperforms state-of-the-art approaches in both node classification and link prediction tasks.

Figures

Figures reproduced from arXiv: 2412.12158 by the authors.

Figure 1
Figure 1. Considering the tuple (Roster, Bucks, Guard, Jrue Holiday) and (Award, Bucks, champion, Season 20-21). We integrate position-aware information into the knowledge hypergraph by introducing relations like Roster-1, Roster-2, Roster-3, Award-1, Award-2, and Award-3. These relations correspond to their order within the tuple and hyperedges and the expansion results in a hierarchical structure. Based on these observation… view at source ↗
Figure 2
Figure 2. Demonstration of instantiated hierarchies in Euclidean and hyperbolic space. The entity [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization 2-dimensional representations on the DBLP dataset in both Euclidean and Lorentz spaces. Each point denotes an individual author, with color differentiation illustrating their respective labels. In practice, we employ the DBLP dataset as a case to visualize the embeddings acquired through UniGNN [16], a unified framework for graph and hypergraph neural networks. We apply the uniform manifold approximati… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Illustration showcasing the H2GNN architecture with operations executed in Lorentz space through the Exponential Map. The two-stage hyper-star message passing, denoted as ϕ1 and ϕ2, is intricately designed for knowledge hypergraphs. In the function ϕ1, the aggregation …
Figure 5
Figure 5. Figure 5: Comparison experiments: encoding the hypergraph structure information with different methods for the same [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Sensitivity analysis of H2GNN modules: hyperbolic operation (HO) and position-aware composition operation (CO) in link prediction task. respectively remove the two modules and report these model variants’ performance in [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Two-dimensional Taxonomy for N-ary Knowledge Representation Learning Methods

    cs.LG 2025-06 conditional novelty 6.0 of 10

    A survey of n-ary knowledge representation learning methods proposes a two-dimensional taxonomy based on modeling technique and entity role/position awareness.

  2. Knowledge Graphs Meet Graph Neural Networks: A Comprehensive Survey

    cs.LG 2026-05 conditional novelty 4.0 of 10

    A two-level taxonomy (KG pipeline stages × GNN architectures) systematically reviews GNN methods for knowledge-graph construction, embedding, reasoning, and applications.

Reference graph

Works this paper leans on

48 extracted references · 45 canonical work pages · cited by 2 Pith papers

  1. [21]

    Fully hyperbolic neural networks

    Weize Chen, Xu Han, Yankai Lin, Hexu Zhao, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. Fully hyperbolic neural networks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022

    work page 2022

  2. [1]

    Poskhg: A position-aware knowledge hypergraph model for link prediction

    Zirui Chen, Xin Wang, Chenxu Wang, and Zhao Li. Poskhg: A position-aware knowledge hypergraph model for link prediction. Data Science and Engineering, 8(2):135–145, 2023

    work page 2023

  3. [2]

    Enhance: Enhanced entity and relation embedding for knowledge hypergraph link prediction

    Chenxu Wang, Zhao Li, Xin Wang, and Zirui Chen. Enhance: Enhanced entity and relation embedding for knowledge hypergraph link prediction. In Companion Proceedings of the ACM Web Conference 2023, pages 115–118, 2023

    work page 2023

  4. [3]

    Medical knowledge graph: Data sources, construction, reasoning, and applications

    Xuehong Wu, Junwen Duan, Yi Pan, and Min Li. Medical knowledge graph: Data sources, construction, reasoning, and applications. Big Data Mining and Analytics, 6(2):201–217, 2023

    work page 2023

  5. [4]

    A universal quaternion hypergraph network for multimodal video question answering

    Zhicheng Guo, Jiaxuan Zhao, Licheng Jiao, Xu Liu, and Fang Liu. A universal quaternion hypergraph network for multimodal video question answering. IEEE Transactions on Multimedia, 2021

    work page 2021

  6. [5]

    Reinforcement learning with actor-critic for knowledge graph reasoning

    Linli Zhang, Dewei Li, Yugeng Xi, and Shuai Jia. Reinforcement learning with actor-critic for knowledge graph reasoning. Science China. Information Sciences, 63(6):169101, 2020

    work page 2020

  7. [6]

    Self- supervised multi-channel hypergraph convolutional network for social recommendation

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

    work page 2021

  8. [7]

    Using rich social media information for music recommendation via hypergraph model

    Shulong Tan, Jiajun Bu, Chun Chen, Bin Xu, Can Wang, and Xiaofei He. Using rich social media information for music recommendation via hypergraph model. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 7, 2011

    work page 2011

Show all 48 references
  1. [8]

    Kbhn: A knowledge-aware bi-hypergraph network based on visual-knowledge features fusion for teaching image annotation.Information Processing & Management, 60(1):103106, 2023

    Hao Li, Jing Wang, Xu Du, Zhuang Hu, and Shuoqiu Yang. Kbhn: A knowledge-aware bi-hypergraph network based on visual-knowledge features fusion for teaching image annotation.Information Processing & Management, 60(1):103106, 2023

    2023

  2. [9]

    Multi-modal knowledge hypergraph for diverse image retrieval

    Yawen Zeng, Qin Jin, Tengfei Bao, and Wenfeng Li. Multi-modal knowledge hypergraph for diverse image retrieval. In The 37th AAAI Conference on Artificial Intelligence, 2023

    2023

  3. [10]

    Megacare: Knowledge-guided multi-view hypergraph predictive framework for healthcare

    Jialun Wu, Kai He, Rui Mao, Chen Li, and Erik Cambria. Megacare: Knowledge-guided multi-view hypergraph predictive framework for healthcare. Information Fusion, 100:101939, 2023. 11

    2023

  4. [11]

    On the representation and embedding of knowledge bases beyond binary relations

    Jianfeng Wen, Jianxin Li, Yongyi Mao, Shini Chen, and Richong Zhang. On the representation and embedding of knowledge bases beyond binary relations. In Proceedings of the 25th International Joint Conference on Artificial Intelligence, 2016

    2016

  5. [14]

    Hypergraph neural networks

    Yifan Feng, Haoxuan You, Zizhao Zhang, Rongrong Ji, and Yue Gao. Hypergraph neural networks. InProceedings of the 33rd AAAI conference on artificial intelligence, 2019

    2019

  6. [15]

    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 Proceedings of the 33rd Annual Conference on Neural Information Processing Systems, 2019

    2019

  7. [16]

    Unignn: a unified framework for graph and hypergraph neural networks

    Jing Huang and Jie Yang. Unignn: a unified framework for graph and hypergraph neural networks. InProceedings of the 13th International Joint Conference on Artificial Intelligence, 2021

    2021

  8. [17]

    Heterogeneous hypergraph variational autoencoder for link prediction

    Haoyi Fan, Fengbin Zhang, Yuxuan Wei, Zuoyong Li, Changqing Zou, Yue Gao, and Qionghai Dai. Heterogeneous hypergraph variational autoencoder for link prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8), 2021

    2021

  9. [18]

    Ffhr: Fully and flexible hyperbolic representation for knowledge graph completion

    Wentao Shi, Junkang Wu, Xuezhi Cao, Jiawei Chen, Wenqiang Lei, Wei Wu, and Xiangnan He. Ffhr: Fully and flexible hyperbolic representation for knowledge graph completion. arXiv preprint arXiv:2302.04088, 2023

    2023 arXiv

  10. [19]

    Hyperbolic graph convolutional neural networks

    Ines Chami, Zhitao Ying, Christopher Ré, and Jure Leskovec. Hyperbolic graph convolutional neural networks. 2019

    2019

  11. [20]

    Machine learning meets complex networks via coalescent embedding in the hyperbolic space

    Alessandro Muscoloni, Josephine Maria Thomas, Sara Ciucci, Ginestra Bianconi, and Carlo Vittorio Cannis- traci. Machine learning meets complex networks via coalescent embedding in the hyperbolic space. Nature communications, 8(1), 2017

    2017

  12. [22]

    On the hyperbolicity of large-scale networks

    W Sean Kennedy, Onuttom Narayan, and Iraj Saniee. On the hyperbolicity of large-scale networks. arXiv preprint arXiv:1307.0031, 2013

    2013 arXiv

  13. [23]

    Tree-like structure in large social and information networks

    Aaron B Adcock, Blair D Sullivan, and Michael W Mahoney. Tree-like structure in large social and information networks. In Proceedings of the 13th international conference on data mining, 2013

    2013

  14. [24]

    Hyperbolic graph attention network

    Yiding Zhang, Xiao Wang, Chuan Shi, Xunqiang Jiang, and Yanfang Ye. Hyperbolic graph attention network. IEEE Transactions on Big Data, 8(6), 2022

    2022

  15. [25]

    Representation tradeoffs for hyperbolic embeddings

    Frederic Sala, Chris De Sa, Albert Gu, and Christopher Ré. Representation tradeoffs for hyperbolic embeddings. In Proceedings of the 35th International Conference on Machine Learning, 2018

    2018

  16. [26]

    Lorentzian distance learning for hyperbolic representations

    Marc Law, Renjie Liao, Jake Snell, and Richard Zemel. Lorentzian distance learning for hyperbolic representations. In Proceedings of the 36th International Conference on Machine Learning, 2019

    2019

  17. [27]

    UMAP: uniform manifold approximation and projection for dimension reduction

    Leland McInnes and John Healy. UMAP: uniform manifold approximation and projection for dimension reduction. CoRR, abs/1802.03426, 2018

    2018 arXiv

  18. [28]

    Knowledge hypergraphs: Prediction beyond binary relations

    Bahare Fatemi, Perouz Taslakian, David Vázquez, and David Poole. Knowledge hypergraphs: Prediction beyond binary relations. In Proceedings of the 29th International Joint Conference on Artificial Intelligence, 2020

    2020

  19. [29]

    Hypersage: Generalizing inductive representation learning on hypergraphs

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

    2010 arXiv

  20. [30]

    Beyond triplets: hyper-relational knowledge graph embedding for link prediction

    Paolo Rosso, Dingqi Yang, and Philippe Cudré-Mauroux. Beyond triplets: hyper-relational knowledge graph embedding for link prediction. In Proceedings of the 2020 Web Conference, 2020

    2020

  21. [31]

    Link prediction on n-ary relational data

    Saiping Guan, Xiaolong Jin, Yuanzhuo Wang, and Xueqi Cheng. Link prediction on n-ary relational data. In Proceedings of the 2019 World Wide Web Conference, 2019

    2019

  22. [32]

    Scalable instance reconstruction in knowledge bases via relatedness affiliated embedding

    Richong Zhang, Junpeng Li, Jiajie Mei, and Yongyi Mao. Scalable instance reconstruction in knowledge bases via relatedness affiliated embedding. In Proceedings of the 2018 World Wide Web Conference, 2018

    2018

  23. [33]

    Neural message passing for multi-relational ordered and recursive hypergraphs

    Naganand Yadati. Neural message passing for multi-relational ordered and recursive hypergraphs. In Proceedings of the 34th Annual Conference on Neural Information Processing Systems, 2020. 12

    2020

  24. [34]

    Bollacker, Colin Evans, Praveen K

    Kurt D. Bollacker, Colin Evans, Praveen K. Paritosh, Tim Sturge, and Jamie Taylor. Freebase: a collaboratively created graph database for structuring human knowledge. In Proceedings of the ACM SIGMOD International Conference on Management of Data, 2008

    2008

  25. [35]

    Neuinfer: Knowledge inference on n-ary facts

    Saiping Guan, Xiaolong Jin, Jiafeng Guo, Yuanzhuo Wang, and Xueqi Cheng. Neuinfer: Knowledge inference on n-ary facts. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020

    2020

  26. [36]

    Knowledge graph embedding for link prediction: A comparative analysis

    Andrea Rossi, Denilson Barbosa, Donatella Firmani, Antonio Matinata, and Paolo Merialdo. Knowledge graph embedding for link prediction: A comparative analysis. ACM Transactions on Knowledge Discovery from Data, 15(2), 2021

    2021

  27. [37]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In Proceedings of the 5th International Conference on Learning Representations, 2017

    2017

  28. [38]

    Predict then propagate: Graph neu- ral networks meet personalized pagerank

    Johannes Klicpera, Aleksandar Bojchevski, and Stephan Günnemann. Predict then propagate: Graph neu- ral networks meet personalized pagerank. In Proceedings of the 7th International Conference on Learning Representations, 2019

    2019

  29. [39]

    Adagcn: Adaboosting graph convolutional networks into deep models

    Ke Sun, Zhanxing Zhu, and Zhouchen Lin. Adagcn: Adaboosting graph convolutional networks into deep models. In Proceedings of the 9th International Conference on Learning Representations, 2021

    2021

  30. [40]

    Heterogeneous graph transformer

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun. Heterogeneous graph transformer. InProceedings of the 2020 Web Conference, 2020

    2020

  31. [41]

    Heterogeneous information network embedding with adversarial disentangler

    Ruijia Wang, Chuan Shi, Tianyu Zhao, Xiao Wang, and Yanfang Ye. Heterogeneous information network embedding with adversarial disentangler. IEEE Transactions on Knowledge and Data Engineering, 35(2), 2023

    2023

  32. [42]

    Houye Ji, Xiao Wang, Chuan Shi, Bai Wang, and Philip S. Yu. Heterogeneous graph propagation network. IEEE Transactions on Knowledge and Data Engineering, (1), 2023

    2023

  33. [43]

    An elementary proof of a theorem of johnson and lindenstrauss

    Sanjoy Dasgupta and Anupam Gupta. An elementary proof of a theorem of johnson and lindenstrauss. Random Structures & Algorithms, 22(1), 2003

    2003

  34. [44]

    Bourgain’s discretization theorem

    Ohad Giladi, Assaf Naor, and Gideon Schechtman. Bourgain’s discretization theorem. In Proceedings of the Annales de la Faculté des sciences de Toulouse: Mathématiques, 2012

    2012

  35. [45]

    Hyperbolic graph neural networks

    Qi Liu, Maximilian Nickel, and Douwe Kiela. Hyperbolic graph neural networks. In Proceedings of the 33rd Annual Conference on Neural Information Processing Systems, 2019

    2019

  36. [46]

    Mixed-curvature multi-relational graph neural network for knowledge graph completion

    Shen Wang, Xiaokai Wei, Cicero Nogueira Nogueira dos Santos, Zhiguo Wang, Ramesh Nallapati, Andrew Arnold, Bing Xiang, Philip S Yu, and Isabel F Cruz. Mixed-curvature multi-relational graph neural network for knowledge graph completion. In Proceedings of the 2021 ACM Web Confe...

    2021

  37. [47]

    LHP: logical hypergraph link prediction

    Yang Yang, Xue Li, Yi Guan, Haotian Wang, Chaoran Kong, and Jingchi Jiang. LHP: logical hypergraph link prediction. Expert Syst. Appl., 222:119842, 2023

    2023

  38. [48]

    Hypergraph neural networks

    Yifan Feng, Haoxuan You, Zizhao Zhang, Rongrong Ji, and Yue Gao. Hypergraph neural networks. In The 33rd AAAI Conference on Artificial Intelligence, AAAI 2019, The 31st Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The 9th AAAI Symposium on Educatio...

    2019

  39. [49]

    Talukdar

    Naganand Yadati, Vikram Nitin, Madhav Nimishakavi, Prateek Yadav, Anand Louis, and Partha P. Talukdar. NHP: neural hypergraph link prediction. In Mathieu d’Aquin, Stefan Dietze, Claudia Hauff, Edward Curry, and Philippe Cudré-Mauroux, editors, CIKM ’20: The 29th ACM Internatio...

    2020

  40. [50]

    Hypergraph joint representation learning for hypervertices and hyperedges via cross expansion

    Yuguang Yan, Yuanlin Chen, Shibo Wang, Hanrui Wu, and Ruichu Cai. Hypergraph joint representation learning for hypervertices and hyperedges via cross expansion. In Michael J. Wooldridge, Jennifer G. Dy, and Sri- raam Natarajan, editors, Thirty-Eighth AAAI Conference on Artific...

    2024

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.