Pith. sign in

REVIEW 6 major objections 6 minor 38 references

Multi-Granular Attention based Heterogeneous Hypergraph Neural Network

T0 review · 6 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read This paper claims that converting symmetric meta-paths in heterogeneous graphs into multi-view hypergraphs, then applying node-level and hyperedge-level attention, yields node representations that consistently outperform seven baselines…

desk verdict Interesting hyperedge construction idea, but the equations don't specify a workable model and the empirical claims overstate the results. read the letter →

arxiv 2505.04340 v1 pith:CBQRVK7J submitted 2025-05-07 cs.LG cs.AI

Hong Jin , Kaicheng Zhou , Jie Yin , Lan You , Zhifeng Zhou This is my paper · ORCID
classification cs.LGcs.AI
keywords heterogeneousgraphneuralnetworkshypergraphmeta-pathattentionmechanismnodeclassificationclusteringrepresentationlearningover-squashing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that the expressive ceiling of meta-path based heterogeneous graph neural networks is set by their pairwise message-passing, and that replacing pairwise meta-path graphs with hypergraphs built from symmetric meta-paths lifts that ceiling. The proposed model, MGA-HHN, constructs one hyperedge type per symmetric meta-path by grouping all target nodes that share the meta-path's central 'identifier' node, then applies attention at two granularities: node-level transformer attention within each hyperedge-type view, and hyperedge-level attention across views. The paper reports consistent gains over seven baselines on node classification across DBLP, IMDB, and ACM, plus better clustering and cleaner visualization. A sympathetic reader would take the claim as: high-order relations in heterogeneous graphs are worth modeling directly, and a multi-view hypergraph with two-level attention is a workable way to do it.

What carries the argument

The load-bearing mechanism is the identifier-node grouping rule for hyperedge construction. For each symmetric meta-path, the central node type determines a hyperedge type; every instance of that central node groups its two end nodes into one hyperedge, thereby encoding a multi-way relation (e.g., several authors coauthoring one paper) as a single unit. Around this construction, the model uses per-view transformer self-attention, masked by the view's normalized hypergraph adjacency matrix, to let every node attend to all others; a hyperedge-level attention vector then learns a weighted sum of the per-view representations. The multi-head version is stabilized with residual connections and an MLP.

What would settle it

Take a heterogeneous graph and two symmetric meta-paths with the same identifier node type but different intermediate node types; if flattening both into the same hyperedge type, or replacing each hyperedge with a random set of the same target nodes, does not change classification accuracy within noise, then the claimed semantic content of the constructed hyperedges is not what drives performance.

Watch

Extended reading notes

Core claim

Meta-paths are pairwise by construction, so a conventional meta-path graph fragments a multi-author collaboration into several two-author edges and loses the fact that those authors co-occur in one event. MGA-HHN's central discovery is that a symmetric meta-path can define a hyperedge directly: take the central node type of the meta-path (the identifier node, e.g., the paper in an Author-Paper-Author path) and group all matching target nodes (the slave nodes, e.g., authors) that share it into a single hyperedge. Doing this for several meta-paths yields a multi-view heterogeneous hypergraph, and the paper shows that transformer-style node-level attention within each view plus hyperedge-level attention across views produces node representations that beat existing homogeneous, heterogeneous, and hypergraph GNNs on node classification, clustering, and visualization.

Load-bearing premise

The result stands on the assumption that grouping all target nodes of a symmetric meta-path by the central identifier node yields a hyperedge that faithfully preserves the meta-path's semantics, even though the intermediate nodes along the path are discarded.

Editorial extensions

If this is right

  • Because each hyperedge bundles all nodes that share an identifier, multi-way co-occurrences such as several authors on one paper or several papers in one venue survive as one relation, instead of being decomposed into pairwise edges.
  • Within a hyperedge-type view, attention is computed between every pair of target nodes, so information does not have to travel through many stacked layers to reach distant nodes; this is the paper's mechanism for mitigating over-squashing.
  • Hyperedge-level attention lets the model up-weight semantically informative views and down-weight noisy ones, which the ablations show is worth several F1 points.
  • On the three datasets, the full model outperforms all seven baselines at every training ratio on Macro-F1 and nearly every ratio on Micro-F1, and the gaps are largest on IMDB, the hardest dataset.
  • Replacing the hypergraph with a conventional graph, or stripping either attention level, consistently lowers F1, isolating the contribution of each component.

Reading between the lines

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

  • The construction suggests a general recipe: any symmetric meta-path can define a view, so the framework could be extended with automatically discovered meta-paths, something the authors list as future work rather than a demonstrated result.
  • The over-squashing mitigation is argued heuristically; one could test it directly by measuring how much a change in a distant node's label or feature alters a target node's representation, and comparing MGA-HHN with a pairwise meta-path baseline under controlled path lengths.
  • Because node-level attention scales as $O(KN^2 d d')$, the method as presented is suited to graphs with a few thousand target nodes; extending to larger graphs would require sparse or sampled attention, which the paper does not address.
  • A testable implication of the identifier-node rule is that two meta-paths with the same identifier type but different intermediate semantics (e.g., co-authorship vs. same-venue publication) should contribute different views; if model performance does not change when those views are merged or shuffled, the claimed semantic separation is not doing the work.
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

6 major / 6 minor

Summary. This paper proposes MGA-HHN, a heterogeneous hypergraph neural network for node representation learning. The model first constructs a multi-view heterogeneous hypergraph by grouping target nodes according to identifier nodes of symmetric meta-paths (e.g., APA and APVPA). It then applies transformer-style node-level attention within each hyperedge-type view, combining the attention matrix with a normalized hypergraph adjacency matrix, and finally fuses the per-view representations with hyperedge-level attention. The authors evaluate node classification, clustering, and visualization on DBLP, IMDB, and ACM, report improvements over seven baselines, and provide ablation, sensitivity, and case studies.

Significance. The high-level idea is relevant and plausible: converting symmetric meta-paths into typed hyperedges is a natural way to expose higher-order relations, and combining node-level and hyperedge-level attention is a reasonable architectural choice. The paper also assembles a broad empirical comparison on three standard benchmarks, including clustering and a case study. However, the core equations in Section 4 contain multiple dimensional inconsistencies, and Eq. (9) as written does not implement the claimed sparse structural attention. In addition, several quantitative claims in Sections 5.2 and 5.3 are not supported by the reported tables. Because the model as specified cannot be executed or traced to the empirical results, the contribution is currently not verifiable; the underlying approach remains potentially viable after correction.

major comments (6)
  1. [Section 4.2, Eq. (4)] The expression \bar A^r = (D_v^r)^{-1/2} (H^r)^T W_e^r (D_e^r)^{-1} H^r (D_v^r)^{1/2} is not dimensionally valid for the stated sizes. With H^r in R^{|V|*|E|}, the product (H^r)^T W_e^r (D_e^r)^{-1} H^r lies in R^{|E|*|E|}, so left multiplication by (D_v^r)^{-1/2} in R^{|V|*|V|} is undefined when |V| is not equal to |E|, as in the DBLP APA view with 4,057 authors and 14,328 paper hyperedges. The standard node-level normalized incidence product is D_v^{-1/2} H^r W_e^r (D_e^r)^{-1} (H^r)^T D_v^{-1/2} in R^{|V|*|V|}, with D_v^{-1/2} on both sides; this correction is needed before Eq. (9) can be evaluated.
  2. [Section 4.3, Eqs. (7)-(8)] The stated dimensions are inconsistent. With X' in R^{|V|*d'} and W_Q, W_K, W_V in R^{d'*d'}, the products Q = W_Q X', K = W_K X', V = W_V X' are not well-defined unless |V| = d', and the claim that Q, K, V lie in R^{d'*d'} is incompatible with there being |V| nodes. Moreover, if Q and K are |V|*d' matrices, then K^T Q is d'*d', not the |V|*|V| similarity matrix required by Eqs. (9) and (10). The standard construction would be Q = X' W_Q, K = X' W_K, and \alpha^r = Q K^T / sqrt(d'), or an equivalent orientation with Q, K in R^{d'*|V|}. As written, node-level attention cannot produce the matrix used in the subsequent equations.
  3. [Section 4.3, Eq. (9)] Even after correcting Eq. (4), the operation Softmax(\alpha^r \odot \bar A^r) does not implement the claimed sparse attention. Since \bar A^r_{ij} = 0 for every pair (i,j) that does not share a hyperedge, those entries contribute e^0 = 1 to the softmax denominator; in the DBLP view, where a hyperedge contains only a small fraction of the 4,057 authors, the denominator is dominated by non-neighbors and the attention is nearly uniform. The described mechanism requires a masked softmax in which zero entries are replaced by -infinity before normalization, or an equivalent sparse softmax over hyperedge neighbors. Eq. (9) should be amended accordingly.
  4. [Section 4.3, Eq. (11)] The residual connection cat(Z^{r,(1)}, ..., Z^{r,(h)}) \oplus X is dimensionally undefined unless d = d'. X is defined in Section 3.1 as R^{|V|*d}, while X' is R^{|V|*d'} after the MLP in Eq. (6). If the residual is meant to use the projected features X', this should be stated; if it is meant to use X, an output projection matching the dimensions is required.
  5. [Sections 5.2-5.3, Tables III-IV] Several quantitative claims are not supported by the reported tables. The statement that MGA-HHN "consistently outperforms all baselines across three datasets" is contradicted by the IMDB 20% Micro-F1 row, where HGTN scores 62.74 and MGA-HHN scores 61.75. The ablation narrative in Section 5.3 attributes drops of 10.7% on DBLP, 6.6% on IMDB, and 2.2% on ACM to removing hyperedge-level attention, but Table IV implies drops of only 0.80, 1.81, and 0.39 points, respectively, from the full model's Micro-F1. In addition, the DBLP row "MGA-HHN hyperedge-concate" (84.90/84.05) is far below "w/o hyperedge-level attention" (93.85/93.54), which is internally implausible and suggests swapped labels or numbers. The ablation narrative should be recomputed and corrected.
  6. [Abstract and Section 6] The claim that MGA-HHN "effectively mitigates over-squashing" is not tested. No experiment measures sensitivity to network depth, information distortion, or long-range dependencies, and the conclusion explicitly defers theoretical treatment of over-squashing to future work. If over-squashing mitigation is a central claimed contribution, it needs direct empirical support, such as depth-wise accuracy or a controlled long-range benchmark, or the claim should be softened.
minor comments (6)
  1. [Throughout] Please correct typos such as "Hetergeneous" (Section 5.1), "matirx" (Section 4.2), "mechnisam" (Sections 1 and 4.1), and "MGA-HNN" (Section 5.4).
  2. [Section 5.1, Table III] The experiments are averaged over five runs, but no standard deviations are reported; several margins over baselines are below 0.5 points, so variance information is needed to assess whether the improvements are significant.
  3. [Section 5.1] The exact hyperparameters used for the main results in Table III, such as hidden dimension, number of attention heads, dropout, and weight decay, are not stated; the sensitivity figures in Section 5.4 do not specify which values were selected for each dataset.
  4. [Section 5.2] The reported gains over baselines (8.5%, 1.9%, 0.5%; 15.7%; 2.6%) are not accompanied by the training-ratio setting or the metric to which they refer, and they are not directly derivable from Table III as presented.
  5. [References] Reference [10] prints "PAKDD 2028"; the correct year is 2018.
  6. [Section 4.5] The complexity analysis uses d and ed while the model text uses d'; please align the notation and clarify that the O(K N^2 d ed) term already assumes the dense attention of Eq. (9).

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: MGA-HHN is evaluated on external benchmarks with standard supervised training; no derivation reduces to its inputs or to self-citations.

full rationale

The paper's central claim is an empirical one---MGA-HHN outperforms baselines on external node-classification and clustering benchmarks (Section 5.2). The model is specified by an explicit hypergraph construction (Section 4.2) and by attention equations (Sections 4.3--4.4); none of these equations is fitted to the labels it later predicts, and no prediction target is defined in terms of a fitted constant. The only author self-citation I could identify is Ref. [10] (Metagraph2vec, co-authored by J. Yin), which appears in a general survey citation for heterogeneous representation learning and is not load-bearing. Ablation results (Section 5.3) compare model variants rather than renaming fitted quantities as predictions. The over-squashing-mitigation narrative is not derived from a self-referential theorem; the conclusion explicitly defers theoretical analysis to future work, so the paper does not import a uniqueness result or ansatz via self-citation. Any concerns about the dimensional consistency of Eq. (4) or the masking behavior of Eq. (9) are correctness or executability issues, not circularity. I therefore find no circular step under the required evidence standard.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

MGA-HHN introduces no new physical or conceptual entities. The hyperedges are derived from existing meta-paths and node types. The ledger lists the hand-tuned hyperparameters and the domain assumptions that the hypergraph construction and attention design rely on.

free parameters (3)
  • hidden_dimension = 256 (ACM), 512 (DBLP and IMDB)
    Tuned via sensitivity analysis in Section 5.4 and Figure 8; the reported best values vary by dataset.
  • number_of_attention_heads = 8 (ACM and DBLP), 16 (IMDB)
    Tuned via sensitivity analysis in Section 5.4 and Figure 9.
  • learning_rate = 0.001
    Fixed Adam learning rate, stated in Section 5.1; a standard hyperparameter choice.
assumptions (4)
  • domain assumption The selected symmetric meta-paths (APA, APCPA, APTPA for DBLP; MAM, MDM for IMDB; PAP, PSP for ACM) capture the task-relevant semantics.
    Used to construct hyperedges in Section 4.2 and the experiments; if these paths are uninformative, the hypergraph views add no signal.
  • domain assumption A symmetric meta-path's central node type uniquely defines a meaningful hyperedge type, and grouping all end nodes by the identifier node preserves the semantics of the original path.
    Definition 4 and 5, Section 4.2; the construction discards intermediate nodes and assumes the central node grouping is a faithful representation.
  • domain assumption The Hadamard product of self-attention with the normalized hypergraph adjacency yields an effective and trainable attention mechanism.
    Equations (8) and (9); a design choice not compared with alternatives such as additive structural bias in the experiments.
  • domain assumption The benchmark datasets and label splits are representative and comparable to those used for the baselines.
    Experimental setup in Section 5.1 gives no split or preprocessing details, so comparability with the cited baseline results is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Granular Attention based Heterogeneous Hypergraph Neural Network." pith.science (2026). https://pith.science/paper/CBQRVK7J

@misc{pith2026250504340,
  author       = {Pith},
  title        = {Pith review of: Multi-Granular Attention based Heterogeneous Hypergraph Neural Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CBQRVK7J}},
  note         = {Machine review of arXiv:2505.04340}
}
read the original abstract

Heterogeneous graph neural networks (HeteGNNs) have demonstrated strong abilities to learn node representations by effectively extracting complex structural and semantic information in heterogeneous graphs. Most of the prevailing HeteGNNs follow the neighborhood aggregation paradigm, leveraging meta-path based message passing to learn latent node representations. However, due to the pairwise nature of meta-paths, these models fail to capture high-order relations among nodes, resulting in suboptimal performance. Additionally, the challenge of ``over-squashing'', where long-range message passing in HeteGNNs leads to severe information distortion, further limits the efficacy of these models. To address these limitations, this paper proposes MGA-HHN, a Multi-Granular Attention based Heterogeneous Hypergraph Neural Network for heterogeneous graph representation learning. MGA-HHN introduces two key innovations: (1) a novel approach for constructing meta-path based heterogeneous hypergraphs that explicitly models higher-order semantic information in heterogeneous graphs through multiple views, and (2) a multi-granular attention mechanism that operates at both the node and hyperedge levels. This mechanism enables the model to capture fine-grained interactions among nodes sharing the same semantic context within a hyperedge type, while preserving the diversity of semantics across different hyperedge types. As such, MGA-HHN effectively mitigates long-range message distortion and generates more expressive node representations. Extensive experiments on real-world benchmark datasets demonstrate that MGA-HHN outperforms state-of-the-art models, showcasing its effectiveness in node classification, node clustering and visualization tasks.

Figures

Figures reproduced from arXiv: 2505.04340 by the authors.

Figure 1
Figure 1. An illustration of a heterogeneous graph, meta-paths and hypergraphs. (a) A heterogeneous graph with three types of [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Customized heterogeneous hypergraph for specific downstream tasks. (a) An example heterogeneous graph with three [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The overall framework of MGA-HHN. First, a heterogeneous hypergraph is constructed based on the task-related meta [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Node-level attention. First, we use WQ,WK,WV to project node features into query (Q), key (K) and value (V) matrices. Then we use KT √ ·Q d′ to obtain the mutual attention between nodes and update it by combining the normalized adjacency matrix A¯ r . Finally, we perfo…
Figure 5
Figure 5. Figure 5: Hyperedge-level attention. The weights of each hyper [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Performance comparison of node clustering results. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Visualization of latent node representations generated by different models on DBLP. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Effect of hidden dimension on model performance. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Effect of attention head h on model performance. explicitly preserve the high-order relations inherent in hetero￾geneous graphs. w/o Node-level Attention refers to the variant that substi￾tutes multi-head node-level attention with a hypergraph con￾volution in HGNN. On …
Figure 10
Figure 10. Figure 10: A case study based on the DBLP dataset. (a) A subgraph extracted as a heterogeneous graph. The green nodes [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Compared to HAN, the representations learned by [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 11
Figure 11. Figure 11: t-SNE visualization of the learned representations of author nodes. [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 22 canonical work pages

  1. [1]

    Mining heterogeneous information networks: a structural analysis approach,

    Y . Sun and J. Han, “Mining heterogeneous information networks: a structural analysis approach,” ACM SIGKDD explorations newsletter , vol. 14, no. 2, pp. 20–28, 2013

    work page 2013

  2. [2]

    A survey of heterogeneous information network analysis,

    C. Shi, Y . Li, J. Zhang, Y . Sun, and S. Y . Philip, “A survey of heterogeneous information network analysis,” IEEE Transactions on Knowledge and Data Engineering , vol. 29, no. 1, pp. 17–37, 2016

    work page 2016

  3. [3]

    Graph transformer networks,

    S. Yun, M. Jeong, R. Kim, J. Kang, and H. J. Kim, “Graph transformer networks,” Advances in neural information processing systems , vol. 32, 2019

    2019

  4. [4]

    Heterogeneous graph trans- former,

    Z. Hu, Y . Dong, K. Wang, and Y . Sun, “Heterogeneous graph trans- former,” in Proceedings of the web conference 2020 , 2020, pp. 2704– 2710

    work page 2020

  5. [5]

    Hypergraph transformer neural networks,

    M. Li, Y . Zhang, X. Li, Y . Zhang, and B. Yin, “Hypergraph transformer neural networks,” ACM Transactions on Knowledge Discovery from Data, vol. 17, no. 5, pp. 1–22, 2023

    2023

  6. [6]

    Heterogeneous network representation learning,

    Y . Dong, Z. Hu, K. Wang, Y . Sun, and J. Tang, “Heterogeneous network representation learning,” in Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI-20 , C. Bessiere, Ed. International Joint Conferences on Artificial Intelligence Organization, 7 2020, pp. 4861–4867, survey track. [Online]. Available: https:/...

    work page doi:10.24963/ijcai.2020/677 2020

  7. [7]

    Heterogeneous network representation learning: A unified framework with survey and benchmark,

    C. Yang, Y . Xiao, Y . Zhang, Y . Sun, and J. Han, “Heterogeneous network representation learning: A unified framework with survey and benchmark,” IEEE Transactions on Knowledge and Data Engineering , vol. 34, no. 10, pp. 4854–4873, 2022

    work page 2022

  8. [8]

    A survey on heterogeneous graph embedding: Methods, techniques, applications and sources,

    X. Wang, D. Bo, C. Shi, S. Fan, Y . Ye, and P. S. Yu, “A survey on heterogeneous graph embedding: Methods, techniques, applications and sources,” IEEE Trans. Big Data , vol. 9, no. 2, pp. 415–436, 2023. [Online]. Available: https://doi.org/10.1109/TBDATA.2022.3177455

    arXiv 2023

Show all 38 references
  1. [9]

    metapath2vec: Scalable rep- resentation learning for heterogeneous networks,

    Y . Dong, N. V . Chawla, and A. Swami, “metapath2vec: Scalable rep- resentation learning for heterogeneous networks,” in Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining , 2017, pp. 135–144

    2017

  2. [10]

    Metagraph2vec: Complex se- mantic path augmented heterogeneous network embedding,

    D. Zhang, J. Yin, X. Zhu, and C. Zhang, “Metagraph2vec: Complex se- mantic path augmented heterogeneous network embedding,” inAdvances in Knowledge Discovery and Data Mining (PAKDD 2028) . Springer International Publishing, 2018, pp. 196–208

    2018

  3. [11]

    Intentgc: a scalable graph convolution framework fusing heterogeneous information for recommendation,

    J. Zhao, Z. Zhou, Z. Guan, W. Zhao, W. Ning, G. Qiu, and X. He, “Intentgc: a scalable graph convolution framework fusing heterogeneous information for recommendation,” in Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2019, pp...

    2019

  4. [12]

    A heterogeneous network-based method with attentive meta-path extraction for predicting drug–target interactions,

    H. Wang, F. Huang, Z. Xiong, and W. Zhang, “A heterogeneous network-based method with attentive meta-path extraction for predicting drug–target interactions,” Briefings in Bioinformatics , vol. 23, no. 4, p. bbac184, 2022

    2022

  5. [13]

    The graph neural network model,

    F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini, “The graph neural network model,” IEEE Transactions on Neural Networks, vol. 20, no. 1, pp. 61–80, 2009

    2009

  6. [14]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. [Online]. Available: ht...

    2017

  7. [15]

    Graph attention networks,

    P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Li `o, and Y . Bengio, “Graph attention networks,” in 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings . OpenReview.net,

    2018

  8. [16]

    Pathsim: Meta path- based top-k similarity search in heterogeneous information networks,

    Y . Sun, J. Han, X. Yan, P. S. Yu, and T. Wu, “Pathsim: Meta path- based top-k similarity search in heterogeneous information networks,” Proceedings of the VLDB Endowment , vol. 4, no. 11, pp. 992–1003, 2011

    2011

  9. [17]

    Hypergraph learning: Methods and practices,

    Y . Gao, Z. Zhang, H. Lin, X. Zhao, S. Du, and C. Zou, “Hypergraph learning: Methods and practices,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 44, no. 5, pp. 2548–2566, 2020

    2020

  10. [18]

    Modeling relational data with graph convolutional networks,

    M. Schlichtkrull, T. N. Kipf, P. Bloem, R. Van Den Berg, I. Titov, and M. Welling, “Modeling relational data with graph convolutional networks,” in The semantic web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, proceedings 15 . Springer, ...

    2018

  11. [19]

    Hetero- geneous graph attention network,

    X. Wang, H. Ji, C. Shi, B. Wang, Y . Ye, P. Cui, and P. S. Yu, “Hetero- geneous graph attention network,” in The world wide web conference , 2019, pp. 2022–2032

    2019

  12. [20]

    Magnn: Metapath aggregated graph neural network for heterogeneous graph embedding,

    X. Fu, J. Zhang, Z. Meng, and I. King, “Magnn: Metapath aggregated graph neural network for heterogeneous graph embedding,” in Proceed- ings of the web conference 2020 , 2020, pp. 2331–2341

    2020

  13. [21]

    Heteroge- neous graph neural network,

    C. Zhang, D. Song, C. Huang, A. Swami, and N. V . Chawla, “Heteroge- neous graph neural network,” in Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , 2019, pp. 793–803

    2019

  14. [22]

    Simple and efficient 14 heterogeneous graph neural network,

    X. Yang, M. Yan, S. Pan, X. Ye, and D. Fan, “Simple and efficient 14 heterogeneous graph neural network,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 37, no. 9, 2023, pp. 10 816– 10 824

    2023

  15. [23]

    On the bottleneck of graph neural networks and its practical implications,

    U. Alon and E. Yahav, “On the bottleneck of graph neural networks and its practical implications,” arXiv preprint arXiv:2006.05205 , 2020

    2006 arXiv

  16. [24]

    Heterogeneous graph neural networks analysis: a survey of techniques, evaluations and applications,

    R. Bing, G. Yuan, M. Zhu, F. Meng, H. Ma, and S. Qiao, “Heterogeneous graph neural networks analysis: a survey of techniques, evaluations and applications,” Artif. Intell. Rev. , vol. 56, no. 8, pp. 8003–8042, 2023. [Online]. Available: https://doi.org/10.1007/s10462-022-10375-2

    2023 doi

  17. [25]

    An attention- based graph neural network for heterogeneous structural learning,

    H. Hong, H. Guo, Y . Lin, X. Yang, Z. Li, and J. Ye, “An attention- based graph neural network for heterogeneous structural learning,” in Proceedings of the AAAI conference on artificial intelligence , vol. 34, no. 04, 2020, pp. 4132–4139

    2020

  18. [26]

    Learning with hypergraphs: Clus- tering, classification, and embedding,

    D. Zhou, J. Huang, and B. Sch ¨olkopf, “Learning with hypergraphs: Clus- tering, classification, and embedding,” Advances in neural information processing systems, vol. 19, 2006

    2006

  19. [27]

    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 Comput. Surv., vol. 56, no. 1, pp. 24:1–24:38, 2024. [Online]. Available: https://doi.org/10.1145/3605776

    2024 doi

  20. [28]

    Hypergraph neural net- works,

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

    2019

  21. [29]

    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

  22. [30]

    Hypergraph convolution and hyper- graph attention,

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

    2021

  23. [31]

    Structural deep embedding for hyper-networks,

    K. Tu, P. Cui, X. Wang, F. Wang, and W. Zhu, “Structural deep embedding for hyper-networks,” in Proceedings of the AAAI conference on artificial intelligence , vol. 32, no. 1, 2018

    2018

  24. [32]

    Be more with less: Hypergraph attention networks for inductive text classification,

    K. Ding, J. Wang, J. Li, D. Li, and H. Liu, “Be more with less: Hypergraph attention networks for inductive text classification,” in Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , B. Webber, T. Cohn, Y . He, and Y . Liu, Eds. O...

    2020

  25. [33]

    Heterogeneous hypergraph embedding for graph classification,

    X. Sun, H. Yin, B. Liu, H. Chen, J. Cao, Y . Shao, and N. Q. Viet Hung, “Heterogeneous hypergraph embedding for graph classification,” in Proceedings of the 14th ACM international conference on web search and data mining , 2021, pp. 725–733

    2021

  26. [34]

    Heterogeneous hypergraph variational autoencoder for link prediction,

    H. Fan, F. Zhang, Y . Wei, Z. Li, C. Zou, Y . Gao, and Q. Dai, “Heterogeneous hypergraph variational autoencoder for link prediction,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 44, no. 8, pp. 4125–4138,

  27. [35]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

    2017

  28. [36]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, Y . Bengio and Y . LeCun, Eds., 2015. [Online]. Available: http://arxi...

    2015 arXiv

  29. [2018]

    Available: https://openreview.net/forum?id=rJXMpikCZ

    [Online]. Available: https://openreview.net/forum?id=rJXMpikCZ

  30. [2022]

    Available: https://doi.org/10.1109/TPAMI.2021.3059313

    [Online]. Available: https://doi.org/10.1109/TPAMI.2021.3059313

    2021

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.