Pith. sign in

REVIEW 4 major objections 5 minor 48 references

DualHNIE: Dual-Channel Hypergraph Learning for Node Importance Estimation in Heterogeneous Knowledge Graphs

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read This paper claims that relation-wide typed hyperedges, with structure and semantics encoded in separate channels and aligned contrastively, improve node-importance ranking in heterogeneous knowledge graphs.

desk verdict Plausible incremental hypergraph NIE model, but the headline numbers don't agree across tables; the FB15K SOTA claim flips depending on which table you read. read the letter →

arxiv 2512.12477 v3 pith:QGMLWWON submitted 2025-12-13 cs.AI cs.LG

classification cs.AIcs.LG
keywords heterogeneousknowledgegraphnodeimportanceestimationhypergraphlearningcontrastivemeta-pathattentionsparsetransformerranking
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 is trying to establish a representation and architecture for estimating node importance in heterogeneous knowledge graphs: instead of treating links pairwise, it groups all entities involved in the same relation into a single typed hyperedge, forming a higher-order graph. It then runs two separate encoders—a hypergraph attention network for local structure and a sparse-chunked hypergraph transformer for global semantics—and aligns the two views with a contrastive loss. The claim is that this disentanglement plus higher-order grouping yields better Spearman correlation and NDCG@100 than existing methods on four benchmark datasets. If correct, it would show that relation-grouped hypergraphs and cross-modal contrastive alignment are useful inductive biases for importance estimation in heterogeneous knowledge graphs.

What carries the argument

The load-bearing object is the relation-typed hyperedge e_r = {u_1,...,u_n, i_1,...,i_m} defined by Eq. (1). It replaces pairwise triples with one set per relation, and hyperedge type features are derived from the relation types appearing across all pairs inside that set. The argument runs through two mechanisms that consume this object: a multi-head hypergraph attention (HGAT) that passes scores from nodes to hyperedges and back, and a Sparse-Chunk-Aggregate transformer (SAHGT) that computes attention only on nonzero incidence entries and processes them in chunks, reducing O(N E d) to O(nnz(H) d). The two channels are fused by an adaptive gate and aligned with a symmetric contrastive loss.

What would settle it

Run the same dual-channel model on an otherwise identical hypergraph in which each relation hyperedge is replaced by (a) random subsets of its members, or (b) pairwise edges only, while keeping every other component fixed. If Spearman and NDCG@100 do not fall, the grouping into relation-wide hyperedges is not the source of the gains; also compare against a model that simply reads node degree and hyperedge size as features.

Watch

Extended reading notes

Core claim

The central claim, on the paper's own terms, is that a heterogeneous higher-order knowledge graph whose typed hyperedges are defined by Eq. (1)—e_r collects all users and items appearing in any triple of relation r—supports a dual-channel encoder that outperforms pairwise and single-channel baselines. The structural channel aggregates messages through hypergraph attention with node-to-hyperedge and hyperedge-to-node steps; the semantic channel applies a sparse-chunked hypergraph transformer over incidence-matrix nonzeros; and a contrastive loss aligns the two embedding spaces while unimodal regression losses retain each channel's predictive power. The reported result is best or near-best Spe

Load-bearing premise

The load-bearing premise is that a hyperedge formed by lumping all users and items of a relation type captures a meaningful higher-order interaction; if that group is mostly noise or so dense that it carries little discriminative signal, the higher-order machinery adds little.

Editorial extensions

If this is right

  • If the central claim holds, relation-grouped hyperedges are a viable replacement for pairwise message passing in importance estimation, capturing associations that pairwise paths miss.
  • Disentangling structural and semantic channels, rather than concatenating features, is directly responsible for part of the gain—the ablation shows the full dual-channel model beats structure-only, semantic-only, and concat variants.
  • Sparse-chunked attention over incidence nonzeros cuts runtime by at least 30% and GPU memory by 28–56% on the reported benchmarks, making large hypergraph transformers feasible.
  • The contrastive alignment loss is load-bearing: removing it causes a substantial drop on MUSIC10K, so cross-modal consistency contributes beyond simple feature fusion.
  • The recipe transfers across relation-rich and relation-sparse graphs, suggesting a general approach to HKG importance estimation.

Reading between the lines

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

  • The construction in Eq. (1) fabricates a complete bipartite link between every user and every item within a relation group. On dense benchmarks the incidence matrix is near-complete, so the 'higher-order' hypergraph carries little discriminative structure; the reported gains may be mostly a feature-aggregation effect, not evidence for genuine higher-order semantics.
  • A testable extension would compare the full relation group against random subsets or against hyperedges built from longer meta-paths; if performance does not drop, the collective group itself is not doing the work.
  • Because every entity inherits the popularity statistics of its relation group, the model could be learning a smoothed degree or popularity signal. A controlled baseline that feeds node degree and group size as plain features would isolate this.
  • The abstract describes hyperedges as coming from meta-path sequences, but Algorithm A.1 builds one hyperedge per single relation type. Extending to multi-hop path patterns is a natural next step the paper does not evaluate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes MetaHGNIE (listed as DualHNIE in the submission metadata and abstract) for node importance estimation in heterogeneous knowledge graphs. The method constructs a hypergraph by grouping all users and items that participate in each relation type into a hyperedge, then encodes structural information with a hypergraph attention network and semantic information with a sparse-chunked hypergraph transformer. The two channels are fused via a contrastive loss and unimodal regression losses. Experiments are reported on FB15K, TMDB5K, IMDB, and MUSIC10K, with Spearman correlation and NDCG@100 as metrics, and the paper claims consistent state-of-the-art performance.

Significance. If the empirical claims are correct, the paper would provide a modest but useful contribution to node importance estimation by combining typed hyperedges, dual-channel encoding, and contrastive alignment. The framework is extensively compared with many baselines and includes ablations and efficiency analyses. However, the current manuscript contains unresolved internal inconsistencies in the headline numbers, and no significance testing is provided, so the core claim of consistent superiority over state-of-the-art methods cannot currently be assessed. The code and dataset links are given, which is helpful for reproducibility.

major comments (4)
  1. [Tables 1, 3, 5 and §5.3.1–5.3.2] The headline result for FB15K is internally inconsistent. Table 1 reports MetaHGNIE Spearman = 0.787 ± 0.004, while Table 3, Table 5, and the text in §5.3.2 report 0.778 ± 0.005. This is not a rounding artifact: the strongest baseline, DVHGNN, has Spearman = 0.779 ± 0.006 in Table 1. If the reproducible value is 0.778, MetaHGNIE is below DVHGNN on FB15K, and the claim of consistent state-of-the-art performance fails on this benchmark. Additionally, §5.3.2 text states the TMDB5K best Spearman is 0.737, while Tables 1 and 3 print 0.762, and Table B.3 reports 0.760/0.787 for TMDB5K/FB15K. The authors must identify which set of numbers is canonical and reconcile all tables and text before the empirical contribution can be evaluated.
  2. [§5.3 and Table 1] No statistical significance tests are reported for any comparison. Given the small differences (e.g., FB15K 0.787±0.004 vs. DVHGNN 0.779±0.006; IMDB 0.793±0.004 vs. SKES 0.779±0.006), overlapping standard deviations mean the claimed improvements may not be significant. The authors should provide paired significance tests or at least confidence intervals for the main comparisons. This is particularly important because the FB15K result flips sign depending on which inconsistent number is used.
  3. [§3.2, Eq. (1), and Abstract] The abstract and introduction claim that typed hyperedges are formed 'from meta-path sequences,' but Eq. (1) and Algorithm A.1 construct each hyperedge by grouping all users and items that appear in triples of a single relation type. This is a per-relation complete bipartite grouping, not a meta-path-induced hyperedge. The construction also creates unobserved user–item links (as acknowledged in Figure 2c), and on three benchmarks the incidence density is high (0.39–0.71), so the hypergraph is nearly complete. The paper should either align the claims with the actual construction or justify why this grouping encodes genuine higher-order interactions rather than spurious associations. This also bears on the 'sparse' scalability claims in §4.2 and Appendix C.
  4. [Appendix C and §4.2] The complexity analysis claims a reduction from O(N E d) to O(nnz(H) d). However, with the dense incidence matrices produced by Eq. (1), the savings factor ρ = nnz(H)/(N E) is close to 1 on TMDB5K (0.39), IMDB (0.64), and MUSIC10K (0.71). The reported runtime reductions of 30–85% therefore depend on implementation details of chunked sparse attention, not on the asymptotic sparsity of the hypergraph. The manuscript should report actual nnz(H)/(N E) values for each dataset and clarify that the efficiency gains are not due to inherent hypergraph sparsity.
minor comments (5)
  1. [Title/Abstract/Code] The arXiv title is 'DualHNIE: Dual-Channel Hypergraph Learning for Node Importance Estimation in Heterogeneous Knowledge Graphs,' while the paper's own title is 'MetaHGNIE: Meta-Path Induced Hypergraph Contrastive Learning in Heterogeneous Knowledge Graphs.' The abstract code link is github.com/jiawenchen10/DualHNIE, and the first-page footer lists github.com/SEU-WENJIA/DualHNIE. These should be harmonized.
  2. [§5.3.2] The text 'MetaHGNIE achieves the best performance (Spearman:0.737, NDCG@100:0.896)' is inconsistent with Table 3, which lists Spearman 0.762 for the same row.
  3. [Table B.2] The table header contains a typo, 'Sperrman' instead of 'Spearman.'
  4. [Figure 2] The labels in Figure 2 are very small and hard to read; enlarging the font would improve clarity.
  5. [Algorithm A.1] The algorithm uses the same symbol E for the hyperedge set and the hyperedge feature tensor. Using different notation (e.g., E and F) would avoid confusion.

Circularity Check

1 steps flagged · score 6.0 of 10

TMDB5K semantic prompts contain the ground-truth popularity label, so the semantic branch (and hence the fused SOTA claim) is partially circular; otherwise the empirical pipeline is self-contained.

  1. self definitional [Section 3.3 (Eq. 2), Eqs. (21)-(22), Appendix A.1]
    "we also include item popularity p_i∈R. A text prompt is constructed as “the description of item n_i is text_i, and its popularity rating is p_i.” we compute the semantic embeddings using a pretrained language model [7]M ... node importance is annotated using the official movie popularity rating."

    TMDB5K's regression target is the official movie popularity rating (Appendix A.1). Section 3.3 puts that same value p_i into the semantic prompt before encoding, and the semantic branch then outputs s_semantic = FFN(semantic features). The 'prediction' is therefore partly a read-off of an input token rather than an independent estimate. Since the fused model (Eq. 22: s_fusion = η1·s_struct + η2·s_semantic) inherits this shortcut, part of the reported TMDB5K gain is forced by construction, not by the higher-order hypergraph machinery.

full rationale

The paper's central empirical claim is evaluated on fixed external labels with 7:1:2 splits, and the hypergraph construction, dual encoders, contrastive loss, and fusion weights are trained or validated rather than set to the test labels, so most of the derivation is not circular. The two self-citations ([1] and [38]) are background and are not used to justify the framework, so no self-citation chain is load-bearing. The FB15K Spearman inconsistency (0.787 in Table 1 vs 0.778 in Tables 3/5) is a correctness/reproducibility concern, not a circularity. The one concrete circular step is in the semantic channel: the prompt in Sec. 3.3 includes 'popularity rating is p_i', and Appendix A.1 states that TMDB5K labels are the official movie popularity rating. Thus, for TMDB5K, the semantic branch's input contains the regression target, and the fused output inherits this shortcut, making part of the reported SOTA improvement over baselines on TMDB5K forced by input/label overlap rather than by higher-order hypergraph modeling. Because only one branch on one dataset is directly identified, the circularity is partial rather than total.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

Everything the central empirical claim rests on beyond standard supervised learning. The paper-specific postulates: (i) per-relation hyperedge grouping (Eq. 1) is a faithful higher-order encoding of multi-entity interactions and its fabricated indirect links are benign; (ii) frozen sentence embeddings (Eq. 2) provide a semantic modality complementary to structure; (iii) scatter-softmax attention over non-zero incidence pairs preserves the expressivity of dense attention (Appendix C). Free parameters the reported numbers depend on: fusion weights η1/η2 (Eq. 22), loss weights α/β (Eq. 25), contrastive temperature τ (Eq. 23), chunk size C, and architecture counts (layers/heads), several selected per dataset via validation sweeps (Figs 6–7). No new entities are postulated; HHKGs and the typed-hyperedge tensor are re-labelings of standard constructions.

free parameters (6)
  • Fusion weights η1, η2 = initialized η1=0.3, η2=0.7; MUSIC10K peak at η1=0.4 (Fig. 6)
    Learnable convex weights in Eq. 22; per-dataset best values used in reported results.
  • Contrastive loss weight α = 0.1–0.3 best per dataset (Fig. 7)
    Trade-off weight in Eq. 25; tuned per dataset.
  • Unimodal loss weight β = 0.2–0.4 best per dataset (Fig. 7)
    Trade-off weight in Eq. 25; tuned per dataset.
  • Temperature τ = not reported
    Contrastive temperature in the similarity matrix of Eq. 23; value never stated.
  • Chunk size C = not reported (swept 10^1–10^5 in Fig. C.1)
    Sparse-chunked attention chunk size; runtime-dependent, memory-independent per Appendix C.
  • Architecture counts (layers L, heads, hidden dim) = default 1 layer, 4–16 heads, hidden 20; deeper layers degrade (Fig. 5)
    Fig. 5 shows NDCG@100 dropping from 0.90 to 0.79 with 3 layers on FB15K; selected by validation.
assumptions (5)
  • domain assumption Ground-truth importance labels (Wikipedia page views, TMDB popularity, IMDB votes, artist familiarity) faithfully represent node importance
    All four datasets labeled as in [11,26]; no robustness check on label noise (§5.1, Table A.1).
  • domain assumption Frozen pretrained sentence embeddings (BERT/all-mpnet-base) of name+description+popularity capture task-relevant semantics
    Semantic channel uses Eq. 2 with §3.3 prompts; no analysis of which semantics matter for ranking.
  • ad hoc to paper A per-relation hyperedge containing all participating users and items (Eq. 1) encodes genuine higher-order interactions rather than introducing spurious associations
    Core modeling postulate of §3.2; creates dense incidence (ρ ≈ 0.39–0.71 on 3 of 4 datasets) and indirect unobserved links whose false-association rate is never validated.
  • ad hoc to paper Sparse scatter-softmax attention over non-zero incidence pairs preserves the expressivity of full attention
    Appendix C claims 'linear savings in ρ without sacrificing attention expressivity'; scatter-softmax normalizes within hyperedges and changes the attention distribution vs dense attention.
  • standard math Standard supervised regression setup (MSE losses, early stopping, 7:1:2 splits) is valid for performance comparison
    §5.1/A.2; the description of 3-fold CV alongside fixed splits is ambiguous.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DualHNIE: Dual-Channel Hypergraph Learning for Node Importance Estimation in Heterogeneous Knowledge Graphs." pith.science (2026). https://pith.science/paper/QGMLWWON

@misc{pith2026251212477,
  author       = {Pith},
  title        = {Pith review of: DualHNIE: Dual-Channel Hypergraph Learning for Node Importance Estimation in Heterogeneous Knowledge Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QGMLWWON}},
  note         = {Machine review of arXiv:2512.12477}
}
read the original abstract

Estimating node importance in heterogeneous knowledge graphs is a fundamental problem underlying recommendation, search, and knowledge decision systems. However, most existing methods rely on pairwise message passing mechanisms that fail to capture higher-order interactions induced by meta-relational structures. Furthermore, structural topology and semantic attributes are typically entangled within a unified embedding space, which obscures their distinct inductive biases and limits the discriminative capacity of learned importance representations. To address these limitations, we propose DualHNIE, a principled dual-channel hypergraph learning framework for node importance estimation. DualHNIE first constructs a higher-order knowledge graph by forming typed hyperedges from meta-path sequences, enabling explicit modeling of higher-order relational patterns. It then introduces two complementary encoders: a structure-aware hypergraph attention network that performs locally normalized aggregation over meta-path--induced hyperedges to capture localized structural dependencies, and a sparse--chunked hypergraph transformer that captures global semantic interactions while maintaining scalable computation. We further design a contrastive alignment mechanism with auxiliary supervision, ensuring cross-view consistency while preserving modality-specific representation. Extensive experiments on multiple benchmark datasets demonstrate that DualHNIE outperforms state-of-the-art methods, validating the effectiveness of explicit high-order modeling and disentangled dual-channel representation learning for heterogeneous knowledge graphs. Code and datasets are available\footnote[1]{https://github.com/jiawenchen10/DualHNIE}.

Figures

Figures reproduced from arXiv: 2512.12477 by the authors.

Figure 1
Figure 1. Illustration of HKGs. (a) Knowledge graph elements. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of direct and indirect relations in HKGs. (a) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed MetaHGNIE framework. (a) Input heterogeneous knowledge graph. (b) Construction of the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison of MetaHGNIE in different levels. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Ablation study of layer depth and heads’ number. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Hyperparameter 𝜂 analysis of multimodal fusion. Parameter Analysis of Fusion and Loss Function [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Hyperparameter analysis of loss function. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 4 linked inside Pith

  1. [1]

    Critical nodes identification in complex networks: a survey.arXiv preprint arXiv:2507.06164(2025)

    CHEN, D., CHEN, J., ZHANG, X., JIA, Q., LIU, X., SUN, Y., LV, L.,ANDYU, W. Critical nodes identification in complex networks: a survey.arXiv preprint arXiv:2507.06164(2025)

  2. [2]

    Deep structural knowledge exploitation and synergy for estimating node importance value on heterogeneous information networks

    CHEN, Y., FANG, Y., WANG, Q., CAO, X.,ANDKING, I. Deep structural knowledge exploitation and synergy for estimating node importance value on heterogeneous information networks. InAAAI(2024), vol. 38, pp. 8302–8310

  3. [3]

    Semi-supervised node impor- tance estimation with informative distribution modeling for uncertainty regulariza- tion

    CHEN, Y., WANG, T., FANG, Y.,ANDXIAO, Y. Semi-supervised node impor- tance estimation with informative distribution modeling for uncertainty regulariza- tion. InProceedings of the ACM on Web Conference 2025(2025), pp. 3108–3118

  4. [4]

    Hyperedge importance estimation via identity-aware hypergraph attention network

    CHEN, Y., WANG, X.,ANDCHEN, C. Hyperedge importance estimation via identity-aware hypergraph attention network. InCIKM(2024), pp. 334–343

  5. [5]

    Llm-guided multi-view hypergraph learning for human-centric explainable recom- mendation.arXiv preprint arXiv:2401.08217(2024)

    CHU, Z., WANG, Y., CUI, Q., LI, L., CHEN, W., QIN, Z.,ANDREN, K. Llm-guided multi-view hypergraph learning for human-centric explainable recom- mendation.arXiv preprint arXiv:2401.08217(2024)

  6. [6]

    CHUNG, C., LEE, J.,ANDWHANG, J. J. Representation learning on hyper- relational and numeric knowledge graphs with transformers. InSIGKDD(2023), pp. 310–322

  7. [7]

    BERT: pre- training of deep bidirectional transformers for language understanding.CoRR abs/1810.04805(2018)

    DEVLIN, J., CHANG, M., LEE, K.,ANDTOUTANOVA, K. BERT: pre- training of deep bidirectional transformers for language understanding.CoRR abs/1810.04805(2018)

  8. [8]

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

    DING, K., WANG, J., LI, J., LI, D.,ANDLIU, H. Be more with less: Hypergraph attention networks for inductive text classification. InEMNLP(2020), pp. 4927– 4936

Show all 48 references
  1. [9]

    Knowledge hypergraphs: prediction beyond binary relations

    FATEMI, B., TASLAKIAN, P., VAZQUEZ, D.,ANDPOOLE, D. Knowledge hypergraphs: prediction beyond binary relations. InIJCAI(2021), IJCAI’20

  2. [10]

    Hypergraph neural networks

    FENG, Y., YOU, H., ZHANG, Z., JI, R.,ANDGAO, Y. Hypergraph neural networks. InAAAI(2019), vol. 33, pp. 3558–3565

  3. [11]

    GE, K.,ANDHAN, Q.-B. Node importance estimation for knowledge graphs based on multi-perspectives attention fusion mechanism.International Journal of Pattern Recognition and Artificial Intelligence 38, 14 (2024), 2459017

  4. [12]

    Inductive representation learning on large graphs.Advances in neural information processing systems 30(2017)

    HAMILTON, W., YING, Z.,ANDLESKOVEC, J. Inductive representation learning on large graphs.Advances in neural information processing systems 30(2017)

  5. [13]

    HAVELIWALA, T. H. Topic-sensitive pagerank. InWWW(2002), pp. 517–526

  6. [14]

    D., GUTIERREZ, C., KIRRANE, S., GAYO, J

    HOGAN, A., BLOMQVIST, E., COCHEZ, M.,D’AMATO, C., MELO, G. D., GUTIERREZ, C., KIRRANE, S., GAYO, J. E. L., NAVIGLI, R., NEUMAIER, S., ET AL. Knowledge graphs.ACM Computing Surveys 54, 4 (2021), 1–37

  7. [15]

    Estimating node importance values in heterogeneous information networks

    HUANG, C., FANG, Y., LIN, X., CAO, X., ZHANG, W.,ANDORLOWSKA, M. Estimating node importance values in heterogeneous information networks. In ICDE(2022), IEEE, pp. 846–858

  8. [16]

    Representation learning on knowledge graphs for node importance estimation

    HUANG, H., SUN, L., DU, B., LIU, C., LV, W.,ANDXIONG, H. Representation learning on knowledge graphs for node importance estimation. InSIGKDD(2021), pp. 646–655

  9. [17]

    Retrieval and reasoning on KGs: Integrate knowledge graphs into large language models for complex question answering

    JI, Y., WU, K., LI, J., CHEN, W., ZHONG, M., JIA, X.,ANDZHANG, M. Retrieval and reasoning on KGs: Integrate knowledge graphs into large language models for complex question answering. InFindings of the Association for Computational Linguistics: EMNLP 2024(Nov. 2024), Y . Al-On...

  10. [18]

    Hypergraph-enhanced dual semi-supervised graph classification

    JU, W., MAO, Z., YI, S., QIN, Y., GU, Y., XIAO, Z., WANG, Y., LUO, X.,AND ZHANG, M. Hypergraph-enhanced dual semi-supervised graph classification. In ICML(2024), JMLR.org

  11. [19]

    Heterogeneous hypergraph neural network for social recommendation using attention network.ACM Transactions on Recommender Systems 3, 3 (Mar

    KHAN, B., WU, J., YANG, J.,ANDMA, X. Heterogeneous hypergraph neural network for social recommendation using attention network.ACM Transactions on Recommender Systems 3, 3 (Mar. 2025)

  12. [20]

    N.,ANDWELLING, M

    KIPF, T. N.,ANDWELLING, M. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907(2016)

  13. [21]

    LEE, J.,ANDWHANG, J. J. Structure is all you need: Structural representation learning on hyper-relational knowledge graphs. InForty-second International Conference on Machine Learning(2025)

  14. [22]

    Dvhgnn: Multi-scale dilated vision hgnn for efficient vision recognition

    LI, C., LI, T., HU, X., LUO, D.,ANDJIN, T. Dvhgnn: Multi-scale dilated vision hgnn for efficient vision recognition. InProceedings of the Computer Vision and Pattern Recognition Conference(2025), pp. 20158–20168

  15. [23]

    Hyper-relational knowledge repre- sentation learning with multi-hypergraph disentanglement

    LI, J., LUO, X., LU, G.,ANDZHANG, S. Hyper-relational knowledge repre- sentation learning with multi-hypergraph disentanglement. InProceedings of the ACM on Web Conference 2025(2025), pp. 3288–3299

  16. [24]

    E.,ANDWU, X

    LI, M., LIU, K., LIU, H., ZHAO, Z., WARD, T. E.,ANDWU, X. Heteroge- neous meta-path graph learning for higher-order social recommendation.ACM Transactions on Knowledge Discovery from Data 18, 8 (2024), 1–25

  17. [25]

    K.,ANDYE, Y

    LI, X., NG, M. K.,ANDYE, Y. Har: hub, authority and relevance scores in multi- relational data for query search. InProceedings of the 2012 SIAM International Conference on Data Mining(2012), SIAM, pp. 141–152

  18. [26]

    Node importance estimation leveraging llms for semantic augmentation in knowledge graphs.arXiv preprint arXiv:2412.00478(2024)

    LIN, X., ZHANG, T., HOU, C., WANG, J., XUE, J.,ANDLV, H. Node importance estimation leveraging llms for semantic augmentation in knowledge graphs.arXiv preprint arXiv:2412.00478(2024)

  19. [27]

    Node importance es- timation with multiview contrastive representation learning.International Journal of Intelligent Systems 2023(2023)

    LIU, L., ZENG, W., TAN, Z., XIAO, W., ZHAO, X.,ET AL. Node importance es- timation with multiview contrastive representation learning.International Journal of Intelligent Systems 2023(2023)

  20. [28]

    Self-supervised dynamic hypergraph recommendation based on hyper-relational knowledge graph

    LIU, Y., XUAN, H., LI, B., WANG, M., CHEN, T.,ANDYIN, H. Self-supervised dynamic hypergraph recommendation based on hyper-relational knowledge graph. InCIKM(2023), pp. 1617–1626

  21. [29]

    Scalable graph neural network- based framework for identifying critical nodes and links in complex networks

    MUNIKOTI, S., DAS, L.,ANDNATARAJAN, B. Scalable graph neural network- based framework for identifying critical nodes and links in complex networks. Neurocomputing 468(2022), 211–221

  22. [30]

    The pagerank citation ranking: Bringing order to the web

    PAGE, L., BRIN, S., MOTWANI, R.,ANDWINOGRAD, T. The pagerank citation ranking: Bringing order to the web. Tech. rep., Stanford infolab, 1999

  23. [31]

    L., ZHAO, T.,ANDFALOUTSOS, C

    PARK, N., KAN, A., DONG, X. L., ZHAO, T.,ANDFALOUTSOS, C. Estimating node importance in knowledge graphs using graph neural networks. InSIGKDD (2019), pp. 596–606

  24. [32]

    L., ZHAO, T.,ANDFALOUTSOS, C

    PARK, N., KAN, A., DONG, X. L., ZHAO, T.,ANDFALOUTSOS, C. Multiimport: Inferring node importance in a knowledge graph from multiple input signals. In SIGKDD(2020), pp. 503–512

  25. [33]

    H., HUYNH, T

    SAKONG, D., VU, V. H., HUYNH, T. T., LENGUYEN, P., YIN, H., NGUYEN, Q. V. H.,ANDNGUYEN, T. T. Higher-order knowledge-enhanced recommendation with heterogeneous hypergraph multi-attention.Information Sciences 680(2024), 121165

  26. [34]

    SAXENA, S., GHATAK, S., KOLLA, R., MUKHERJEE, D.,ANDCHAKRABORTY, WWW, April 13-17, 2026, Dubai, United Arab Emirates T rovato et al. T. Dphgnn: A dual perspective hypergraph neural networks. InSIGKDD(2024), pp. 2548–2559

  27. [35]

    Dhyper: A recurrent dual hyper- graph neural network for event prediction in temporal knowledge graphs.ACM Transactions on Information Systems 42, 5 (2024), 1–23

    TANG, X., CHEN, L., SHI, H.,ANDLYU, D. Dhyper: A recurrent dual hyper- graph neural network for event prediction in temporal knowledge graphs.ACM Transactions on Information Systems 42, 5 (2024), 1–23

  28. [36]

    Dualgnn: Dual graph neural network for multimedia recommendation.IEEE Transactions on Multimedia 25(2021), 1074–1084

    WANG, Q., WEI, Y., YIN, J., WU, J., SONG, X.,ANDNIE, L. Dualgnn: Dual graph neural network for multimedia recommendation.IEEE Transactions on Multimedia 25(2021), 1074–1084

  29. [37]

    WANG, X., JI, H., SHI, C., WANG, B., YE, Y., CUI, P.,ANDYU, P. S. Hetero- geneous graph attention network. InWWW(2019), pp. 2022–2032

  30. [38]

    Dual-view desynchronization hypergraph learning for dynamic hyperedge prediction.IEEE Transactions on Knowledge and Data Engineering(2024)

    WANG, Z., CHEN, J., SHAO, Z.,ANDWANG, Z. Dual-view desynchronization hypergraph learning for dynamic hyperedge prediction.IEEE Transactions on Knowledge and Data Engineering(2024)

  31. [39]

    Relation-aware entity alignment for heterogeneous knowledge graphs

    WU, Y., LIU, X., FENG, Y., WANG, Z., YAN, R.,ANDZHAO, D. Relation-aware entity alignment for heterogeneous knowledge graphs. InIJCAI(2019)

  32. [40]

    Self-supervised hypergraph transformer for recommender systems

    XIA, L., HUANG, C.,ANDZHANG, C. Self-supervised hypergraph transformer for recommender systems. InSIGKDD(2022), pp. 2100–2109

  33. [41]

    Knowledge graph self- supervised rationalization for recommendation

    YANG, Y., HUANG, C., XIA, L.,ANDHUANG, C. Knowledge graph self- supervised rationalization for recommendation. InSIGKDD(2023), pp. 3046– 3056

  34. [42]

    Knowledge graph contrastive learning for recommendation

    YANG, Y., HUANG, C., XIA, L.,ANDLI, C. Knowledge graph contrastive learning for recommendation. InSIGIR(2022), pp. 1434–1443

  35. [43]

    High-order neighbors aware representation learning for knowledge graph completion.IEEE Transactions on Neural Networks and Learning Systems 36, 3 (2024), 5273–5287

    YIN, H., ZHONG, J., LI, R., SHANG, J., WANG, C.,ANDLI, X. High-order neighbors aware representation learning for knowledge graph completion.IEEE Transactions on Neural Networks and Learning Systems 36, 3 (2024), 5273–5287

  36. [44]

    YU, J., YIN, H., LI, J., WANG, Q., HUNG, N. Q. V.,ANDZHANG, X. Self- supervised multi-channel hypergraph convolutional network for social recommen- dation. InWWW(2021), pp. 413–424

  37. [45]

    Knowgpt: Knowledge graph based prompting for large language models.Advances in Neural Information Processing Systems 37(2024), 6052–6080

    ZHANG, Q., DONG, J., CHEN, H., ZHA, D., YU, Z.,ANDHUANG, X. Knowgpt: Knowledge graph based prompting for large language models.Advances in Neural Information Processing Systems 37(2024), 6052–6080

  38. [46]

    Label informed contrastive pretraining for node importance estimation on knowledge graphs.IEEE Transactions on Neural Networks and Learning Systems (2024)

    ZHANG, T., HOU, C., JIANG, R., ZHANG, X., ZHOU, C., TANG, K.,ANDLV, H. Label informed contrastive pretraining for node importance estimation on knowledge graphs.IEEE Transactions on Neural Networks and Learning Systems (2024)

  39. [47]

    Relational graph neural network with hierarchical attention for knowledge graph completion

    ZHANG, Z., ZHUANG, F., ZHU, H., SHI, Z., XIONG, H.,ANDHE, Q. Relational graph neural network with hierarchical attention for knowledge graph completion. InAAAI(2020), vol. 34, pp. 9612–9619

  40. [48]

    Cadren: Contextual anchor- driven relational network for controllable cross-graphs node importance estimation

    ZHONG, Z., ZHANG, Y., CHANG, Z.,ANDQIN, Z. Cadren: Contextual anchor- driven relational network for controllable cross-graphs node importance estimation. InEuropean Conference on Machine Learning and 25th Principles and Practice of Knowledge Discovery in Databases(2025). A Imp...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.