Pith. sign in

REVIEW 3 major objections 5 minor 49 references

Bridging RDF Knowledge Graphs with Graph Neural Networks for Semantically-Rich Recommender Systems

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read RDF knowledge graphs, converted into heterogeneous graphs with both literal-content and topology features, push GNN-based recommenders to F1 0.940 and AUC 0.987, outperforming bipartite and homogeneous graph settings.

desk verdict Useful datasets and a plausible claim, but the evaluation protocol is under-specified enough that the headline result—heterogeneous graphs beat bipartite ones—needs confirmation before it should be cited. read the letter →

arxiv 2506.08743 v1 pith:JK354YXG submitted 2025-06-10 cs.IR cs.AIcs.DBcs.LG

Michael Färber , David Lamprecht , Yuni Susanti This is my paper · ORCID
classification cs.IRcs.AIcs.DBcs.LG
keywords RDFknowledgegraphsgraphneuralnetworksrecommendersystemsheterogeneouslinkpredictionsemanticnodefeaturesAutoRDF2GML
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 tries to establish that the full semantic content of RDF knowledge graphs—text literals and graph topology alike—can be fed to graph neural network recommenders as node features and that doing so improves link prediction beyond simpler graph configurations. The authors evaluate GraphSAGE, GAT, and HGT on two large RDF-derived datasets, SOA-SW and LPWC, comparing ten feature initializations that combine BERT-based content embeddings with TransE-based topology embeddings. Their best reported configuration, GraphSAGE with weighted addition of content and topology features on the full heterogeneous graph, reaches an F1 of 0.940 and an AUC of 0.987 for paper recommendation. These results matter because they suggest that the existing Linked Open Data cloud can serve as a ready-made feature source for recommenders, removing the need to hand-write SPARQL queries or curate interaction data.

What carries the argument

The load-bearing mechanism is a four-stage pipeline. AutoRDF2GML converts an RDF knowledge graph into a heterogeneous graph whose nodes carry two kinds of features: topology-based vectors from knowledge graph embeddings such as TransE over object properties, and content-based vectors from BERT embeddings over datatype-property literals. A learned linear layer projects the differing node-type feature spaces into a common dimension, then a GNN encoder—GraphSAGE, GAT, or HGT—propagates messages along the heterogeneous edge types. Finally a dot-product decoder scores candidate links, and the experiments vary the feature combination and the graph structure to isolate what each contributes.

What would settle it

Train the best configuration (GraphSAGE with weighted addition on the full heterogeneous SOA-SW graph) but compute the TransE embeddings on a copy of the graph with all author-work edges deleted, then evaluate on held-out author-work links; if the F1 gap between TransE-initialized and one-hot runs collapses, the semantic gains are leakage rather than real content. Alternatively, inspect the released AutoRDF2GML configuration and data splits for a masking statement.

Watch

Extended reading notes

Core claim

The paper's central claim, stated on its own terms, is that semantic depth—not architecture complexity—drives GNN-based recommendation performance: using RDF datatype properties as content-based features and RDF object properties as topology-based features, and combining them by weighted addition or averaging, reliably beats one-hot initialization, and the full heterogeneous graph outperforms bipartite and homogeneous subgraphs in every scenario tested. On the SOA-SW and LPWC datasets, GraphSAGE with combined semantic features achieves the best F1 and AUC across paper, collaboration, and task recommendation, while simpler configurations fall behind. The paper further claims that existing heterogeneous-graph benchmarks lack such all-node-type content and topology features, so this combination is unique to RDF-derived datasets and is the reason the gains are visible here.

Load-bearing premise

The paper's reported gains assume that the graph embeddings used to initialize the topology features were trained without seeing the exact author-work, author-author, and dataset-task links that the recommender later predicts; the paper never states that those edges were excluded.

Editorial extensions

If this is right

  • GNN recommenders can be built directly on RDF knowledge graphs from the Linked Open Data cloud, with no SPARQL query writing, by feeding the converted graph into a standard GNN.
  • Combining content and topology embeddings with weighted addition or averaging becomes a cheap and effective default initialization for heterogeneous graph recommenders.
  • Benchmark results on older datasets that lack all-node-type semantic features may understate GNN performance, since part of the observed gain in this paper is feature richness rather than model design.
  • Heterogeneous GNN architectures are not strictly necessary: GraphSAGE, a homogeneous architecture, wins once it is given semantically rich heterogeneous features.

Reading between the lines

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

  • An implication the authors leave implicit is that the same pipeline applied to any RDF knowledge graph in the Linked Open Data cloud would yield a recommender with essentially no user-item interaction history, because the graph itself supplies the features.
  • A strictly stronger test than the paper's would be to mask the predicted edge types during TransE training; until that is done, some fraction of the reported F1/AUC gains from topology features might reflect information leakage rather than semantic content.
  • A natural extension is a time-based data split on bibliographic graphs, where future citations are predicted from past literature, to check whether heterogeneous context helps when the predictive signal is genuinely temporal.
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

3 major / 5 minor

Summary. The paper presents a GNN-based recommender pipeline that converts RDF knowledge graphs into heterogeneous graphs via the authors' AutoRDF2GML tool, using content-based features (e.g., BERT embeddings of RDF literals) and topology-based features (e.g., TransE embeddings of RDF object properties). The approach is evaluated on two large RDF-derived datasets, SOA-SW and LPWC, for paper, collaboration, and task recommendation. The manuscript compares three GNN encoders (GraphSAGE, GAT, HGT), ten feature initialization strategies, and full heterogeneous versus bipartite or homogeneous graph settings. The central claim is that semantically rich RDF features and heterogeneous graph structures significantly improve GNN-based recommender performance, with best reported results of F1 0.940 and AUC 0.987 for paper recommendation.

Significance. If the evaluation protocol is sound, the paper would be a useful empirical contribution: it provides a large-scale comparison of feature initialization strategies and graph heterogeneity for GNN recommenders, and it ships code and data links, which is a concrete strength for reproducibility. The paper also identifies a gap in existing benchmarks, which typically lack both content-based and topology-based features for all node types. However, the central claims rest on an under-specified evaluation protocol. The manuscript does not state whether the knowledge-graph embeddings used as topology features are trained on the same edges that are later predicted, nor does it describe the train/validation/test split, edge masking, or negative sampling. Several head-to-head comparisons also change two variables at once (feature initialization and graph structure). These issues must be resolved before the empirical claims can be accepted.

major comments (3)
  1. [§2.1 and §3] The manuscript does not state whether the topology-based features x_t (e.g., TransE embeddings produced by AutoRDF2GML) are trained on the full RDF graph including the exact edges used as link-prediction targets (author-work, author-author, and dataset-task edges). If the KGE model is fit on the same graph before the split, the dot-product decoder in §2.2 can recover target edges almost directly from the input features, making the reported gains of the 'tb' and 'comb*' initializations circular. The authors must specify the exact split procedure, state whether target edges are excluded from embedding training, and describe edge masking and negative sampling for the GNN evaluation.
  2. [§4, Tables 1 and 3] The comparison of full heterogeneous versus bipartite/homogeneous graph settings is confounded by feature initialization. For paper recommendation, the best full heterogeneous result uses combWAddition (F1 0.940) while the bipartite result uses topology-based features only (F1 0.936); for task recommendation, the full heterogeneous result uses combAverage (F1 0.923) while the bipartite result uses tb (F1 0.915). Since both feature initialization and graph structure differ, the observed differences cannot be attributed to graph heterogeneity. The authors should compare settings with identical feature initialization, or explicitly separate the effect of each factor.
  3. [Tables 1–3 and §4] All evaluation metrics are reported as single point estimates without standard deviations, number of independent runs, or statistical significance tests. Several differences that support the paper's main claims are very small: for example, the full-heterogeneous versus bipartite F1 gap is 0.4 percentage points in Table 1 and 0.8 points in Table 3, and in Table 1 the GAT cbnld F1 (0.877) is nearly identical to one-hot (0.875). The claim that semantic features 'consistently' and 'significantly' improve performance is not supported by the reported evidence. The authors should report variance over multiple seeds and perform significance testing, especially for the small-margin comparisons.
minor comments (5)
  1. [§4 heading] The heading 'Analyis of Semantic Feature Initialization Methods' contains a typo; it should be 'Analysis'.
  2. [§4] The text refers to 'Tables 5-7' when discussing the analysis, but the paper contains only Tables 1-3; the references should be corrected.
  3. [Table 2] In the homogeneous graph section, HGT entries are marked with '-' but no explanation is given; the authors should state why HGT is not applied to the homogeneous setting.
  4. [§2.2, item 5] The initialization method 'combnld|tb' is described as a conditional fallback ('NLD if available, otherwise topology-based'), not as a combination; the naming and description should be clarified to avoid confusion with the truly combined methods such as combConcat and combWAddition.
  5. [§3] Hyperparameters are described only as 'default parameters from GNN literature' with further details deferred to GitHub; for a self-contained evaluation, key hyperparameters such as hidden dimensions, number of layers, learning rate, number of epochs, negative sampling ratio, and BERT model variant should be stated in the paper.

Circularity Check

1 steps flagged · score 6.0 of 10

Link prediction results may reduce to TransE embeddings fit on the same graph whose edges are predicted; the evaluation protocol is underspecified.

  1. fitted input called prediction [Section 2.1 (Topology-based features x_t), Section 2.2 step 4 (dot-product decoder), Section 3 (Paper Recommendation scenario); Tables 1-3]
    "Techniques for obtaining these representations include Knowledge Graph Embedding (KGE) techniques such as TransE [3], DistMult [41], ComplEx [32], and RotatE [30], which effectively encode RDF entities into feature vectors. AutoRDF2GML automates this process using these techniques. ... Paper Recommendation: In this scenario, we perform link prediction on author-work edge using the SOA-SW dataset"

    The paper describes producing topology features by fitting KGE models (e.g., TransE) to the RDF KG, then evaluating link prediction on the resulting dataset. It never states that the target edges (author-work, author-author, dataset-task) are excluded from the KG before KGE training, nor that the GNN masks validation/test edges during message passing. As written, the TransE embeddings are fit on the same graph whose edges are subsequently predicted. Because the decoder is a dot product of node representations, a GNN can essentially pass through these fitted embeddings; the dot product then reproduces the TransE scoring signal for the target relation.

full rationale

The central empirical claim—that semantic RDF features and heterogeneous graph structure significantly improve GNN recommenders—rests entirely on the evaluation in Section 3. The paper's own pipeline (§2.1) generates topology-based features x_t by fitting KGE models to the RDF KG, and §3 performs link prediction on edges of that same KG (author-work, author-author, dataset-task). No train/validation/test split, edge masking, or exclusion of target edges from the KGE training set is described anywhere in the manuscript. If the standard protocol of removing target edges before embedding training was followed, the results would be independent; but as written, the link prediction scores could reduce to the TransE features computed on the very edges being predicted, making the 'semantic feature' gains and the heterogeneous-vs-bipartite comparison (Table 1: 0.940 vs 0.936, 0.4% gap) partially circular. The self-citations to AutoRDF2GML and the SOA-SW/LPWC datasets are extensive but are not by themselves circular; the load-bearing issue is the unstated leakage boundary in the evaluation. Because the protocol is underspecified rather than an explicit equation-level reduction, the score is 6 rather than higher; if the authors confirm a leak-free split, the circularity concern would be resolved.

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

The central claim rests on the evaluation protocol being free of leakage, on the transfer of default hyperparameters, and on the fidelity of the authors' own conversion pipeline. There are no new theoretical entities, but the model parameters and feature combinations are fit to the data.

free parameters (3)
  • Feature projection matrix W and bias b = Learned during training
    Used to map node-type-specific features into a common space in Section 2.2, step 2.
  • Neural combinator MLP weights = Learned during training
    combnc uses z = f(W·[a,b] + c), with the network trained end-to-end on the recommendation task.
  • Weighted addition combination weights = Learned during training
    combWAddition combines content and topology features using weights fit to the target task.
assumptions (3)
  • domain assumption The train/test split is independent of feature extraction, so topology embeddings do not encode target links.
    The evaluation in Section 3 does not describe splitting before AutoRDF2GML feature generation. If KGE embeddings are fit on the full graph, the central comparison is compromised.
  • domain assumption Default hyperparameters from prior GNN literature transfer to these RDF-derived datasets.
    Section 3 states that default parameters from GNN literature are used, but no evidence shows these defaults are appropriate for the large heterogeneous RDF graphs.
  • domain assumption The AutoRDF2GML conversion faithfully represents RDF semantics for GNN input.
    The pipeline relies on the authors' own conversion tool, and the paper provides no independent verification that the transformation preserves the semantics needed for recommendation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging RDF Knowledge Graphs with Graph Neural Networks for Semantically-Rich Recommender Systems." pith.science (2026). https://pith.science/paper/JK354YXG

@misc{pith2026250608743,
  author       = {Pith},
  title        = {Pith review of: Bridging RDF Knowledge Graphs with Graph Neural Networks for Semantically-Rich Recommender Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JK354YXG}},
  note         = {Machine review of arXiv:2506.08743}
}
read the original abstract

Graph Neural Networks (GNNs) have substantially advanced the field of recommender systems. However, despite the creation of more than a thousand knowledge graphs (KGs) under the W3C standard RDF, their rich semantic information has not yet been fully leveraged in GNN-based recommender systems. To address this gap, we propose a comprehensive integration of RDF KGs with GNNs that utilizes both the topological information from RDF object properties and the content information from RDF datatype properties. Our main focus is an in-depth evaluation of various GNNs, analyzing how different semantic feature initializations and types of graph structure heterogeneity influence their performance in recommendation tasks. Through experiments across multiple recommendation scenarios involving multi-million-node RDF graphs, we demonstrate that harnessing the semantic richness of RDF KGs significantly improves recommender systems and lays the groundwork for GNN-based recommender systems for the Linked Open Data cloud. The code and data are available on our GitHub repository: https://github.com/davidlamprecht/rdf-gnn-recommendation

Figures

Figures reproduced from arXiv: 2506.08743 by the authors.

Figure 1
Figure 1. Illustration of homogeneous (left) vs. heterogeneous (right) graphs (RDF). [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of automatic features extraction with [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. GNN-based recommendation system with RDF-based features. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 29 canonical work pages

  1. [1]

    Pattern Recognit

    Ansarizadeh, F., Tay, D.B.H., Thiruvady, D.R., Robles-Kelly, A.: Deterministic sampling in heterogeneous graph neural networks. Pattern Recognit. Lett.172, 74–81 (2023)

    work page 2023

  2. [2]

    van den Berg, R., Kipf, T.N., Welling, M.: Graph convolutional matrix completion (2017), https://arxiv.org/abs/1706.02263

    arXiv 2017

  3. [3]

    Advances in neural information processing systems26(2013)

    Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J., Yakhnenko, O.: Translating embeddings for modeling multi-relational data. Advances in neural information processing systems26(2013)

    work page 2013

  4. [4]

    In: Journal of Physics

    Chi, H., Wang, Y., Hao, Q., Xia, H.: Residual network and embedding usage: New tricks of node classification with graph convolutional networks. In: Journal of Physics. vol. 2171, p. 012011. IOP Publishing (2022)

    work page 2022

  5. [5]

    In: Proceedings of the 31st ACM International Conference on Information & Knowledge Management

    Cui, H., Lu, Z., Li, P., Yang, C.: On positional and structural node features for graph neural networks on non-attributed graphs. In: Proceedings of the 31st ACM International Conference on Information & Knowledge Management. pp. 3898–

    work page

  6. [6]

    In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Com- putational Linguistics: Human Language Technologies

    Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Com- putational Linguistics: Human Language Technologies. pp. 4171–4186 (2019). https://doi.org/10.18653/v1/N19-1423

    doi:10.18653/v1/n19-1423 2019

  7. [7]

    Dong, Y., Chawla, N.V., Swami, A.: metapath2vec: Scalable representation learn- ingforheterogeneousnetworks.In:SIGKDDinternationalconferenceonknowledge discovery and data mining. pp. 135–144 (2017)

    work page 2017

  8. [8]

    Semantic Web9(1), 77–129 (2018), https://doi.org/10.3233/SW-170275

    Färber, M., Bartscherer, F., Menne, C., Rettinger, A.: Linked data quality of dbpe- dia, freebase, opencyc, wikidata, and YAGO. Semantic Web9(1), 77–129 (2018), https://doi.org/10.3233/SW-170275

    doi:10.3233/sw-170275 2018

Show all 49 references
  1. [9]

    In: Proceedings of the 22nd International Semantic Web Conference

    Färber, M., Lamprecht, D.: Linked Papers With Code: The Latest in Machine Learning as an RDF Knowledge Graph. In: Proceedings of the 22nd International Semantic Web Conference. ISWC’23 (2023)

    2023

  2. [10]

    In: Proceedings of the 22nd Inter- national Semantic Web Conference

    Färber, M., Lamprecht, D., Krause, J., Aung, L., Haase, P.: SemOpenAlex: The Scientific Landscape in 26 Billion RDF Triples. In: Proceedings of the 22nd Inter- national Semantic Web Conference. pp. 94–112. ISWC’23, Springer (2023)

    2023

  3. [11]

    In: Proceedings of the 2024 International Semantic Web Conference

    Färber, M., Lamprecht, D., Susanti, Y.: Autordf2gml: Facilitating rdf integration in graph machine learning. In: Proceedings of the 2024 International Semantic Web Conference. ISWC 2024 (2024)

    2024

  4. [12]

    arXiv preprint arXiv:1903.02428 (2019)

    Fey, M., Lenssen, J.E.: Fast graph representation learning with pytorch geometric. arXiv preprint arXiv:1903.02428 (2019)

    2019 arXiv

  5. [13]

    In: Proceedings of the 2020 Web Conference

    Fu, X., Zhang, J., Meng, Z., King, I.: MAGNN: Metapath Aggregated Graph Neu- ral Network for Heterogeneous Graph Embedding. In: Proceedings of the 2020 Web Conference. pp. 2331–2341. Web’20 (2020)

    2020

  6. [14]

    ACM Transactions on Recommender Systems 1(1), 1–51 (2023)

    Gao, C., Zheng, Y., Li, N., Li, Y., Qin, Y., Piao, J., Quan, Y., Chang, J., Jin, D., He, X., et al.: A survey of graph neural networks for recommender systems: Challenges, methods, and directions. ACM Transactions on Recommender Systems 1(1), 1–51 (2023)

    2023

  7. [15]

    In: Proceedings of the 16th ACM Conference on Recommender Systems

    Geng, S., Liu, S., Fu, Z., Ge, Y., Zhang, Y.: Recommendation as language pro- cessing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). In: Proceedings of the 16th ACM Conference on Recommender Systems. p. 299–315. RecSys ’22 (2022), https://doi.org/10.11...

    2022

  8. [16]

    GroupLens Research: MovieLens (2023), https://movielens.org, accessed on: 29.09.2023

    2023

  9. [17]

    IEEE Transactions on Knowledge and Data Engineering34(8), 3549–3568 (2020)

    Guo, Q., Zhuang, F., Qin, C., Zhu, H., Xie, X., Xiong, H., He, Q.: A survey on knowledge graph-based recommender systems. IEEE Transactions on Knowledge and Data Engineering34(8), 3549–3568 (2020)

    2020

  10. [18]

    Advances in neural information processing systems30(2017)

    Hamilton, W., Ying, Z., Leskovec, J.: Inductive representation learning on large graphs. Advances in neural information processing systems30(2017)

    2017

  11. [19]

    arXiv preprint arXiv:2305.19872 (2023)

    He, M., Wei, Z., Feng, S., Huang, Z., Li, W., Sun, Y., Yu, D.: Spectral heteroge- neous graph convolutions via positive noncommutative polynomials. arXiv preprint arXiv:2305.19872 (2023)

    2023 arXiv

  12. [20]

    Advances in neural information processing systems33, 22118–22133 (2020)

    Hu, W., Fey, M., Zitnik, M., Dong, Y., Ren, H., Liu, B., Catasta, M., Leskovec, J.: Open graph benchmark: Datasets for machine learning on graphs. Advances in neural information processing systems33, 22118–22133 (2020)

    2020

  13. [21]

    In: Pro- ceedings of the 2020 Web Conference

    Hu, Z., Dong, Y., Wang, K., Sun, Y.: Heterogeneous Graph Transformer. In: Pro- ceedings of the 2020 Web Conference. pp. 2704–2710. Web’20 (2020)

    2020

  14. [22]

    arXiv preprint arXiv:1609.02907 (2016)

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

    2016 arXiv

  15. [23]

    https://doi

    Lamprecht, D., Färber, M.: Graph Machine Learning Dataset LPWC. https://doi. org/10.5281/zenodo.10299366 (2023)

    2023 doi

  16. [24]

    https: //doi.org/10.5281/zenodo.10299428 (2023)

    Lamprecht, D., Färber, M.: Graph Machine Learning Dataset SOA-SW. https: //doi.org/10.5281/zenodo.10299428 (2023)

    2023 doi

  17. [25]

    arXiv preprint arXiv:2307.08430 (2023)

    Li, C., Guo, Z., He, Q., Xu, H., He, K.: Long-range dependency based multi-layer perceptron for heterogeneous information networks. arXiv preprint arXiv:2307.08430 (2023)

    2023 arXiv

  18. [26]

    In: ACM SIGKDD Conference on Knowledge Discovery & Data Mining

    Lv, Q., Ding, M., Liu, Q., Chen, Y., Feng, W., He, S., Zhou, C., Jiang, J., Dong, Y., Tang, J.: Are we really making much progress? Revisiting, benchmarking and refining heterogeneous graph neural networks. In: ACM SIGKDD Conference on Knowledge Discovery & Data Mining. pp. 11...

    2021

  19. [27]

    https://www.w3.org/TR/ rdf-primer/ (2004), [Accessed 04-09-2024]

    Manola, F., Miller, E.: RDF Primer — w3.org. https://www.w3.org/TR/ rdf-primer/ (2004), [Accessed 04-09-2024]

    2004

  20. [28]

    Current Research in Structural Biology7, 100122 (2024)

    Momanyi, B.M., Zhou, Y.W., Grace-Mercure, B.K., Temesgen, S.A., Basharat, A., Ning, L., Tang, L., Gao, H., Lin, H., Tang, H.: SAGESDA: Multi-GraphSAGE net- works for predicting SnoRNA-disease associations. Current Research in Structural Biology7, 100122 (2024). https://doi.org...

    2024

  21. [29]

    Semantic Web11(1), 101–113 (2020)

    Polleres, A., Kamdar, M.R., Fernández, J.D., Tudorache, T., Musen, M.A.: A more decentralized vision for linked data. Semantic Web11(1), 101–113 (2020). https://doi.org/10.3233/SW-190380, https://doi.org/10.3233/SW-190380

    2020 doi

  22. [30]

    Sun, Z., Deng, Z., Nie, J., Tang, J.: Rotate: Knowledge graph embedding by rela- tional rotation in complex space (2019)

    2019

  23. [31]

    Nucleic acids research51(D1), D638–D646 (2023)

    Szklarczyk, D., Kirsch, R., Koutrouli, M., Nastou, K., Mehryary, F., Hachilif, R., Gable, A.L., Fang, T., Doncheva, N.T., Pyysalo, S., et al.: The string database in 2023: protein–protein association networks and functional enrichment analyses for any sequenced genome of inter...

    2023

  24. [32]

    In: International conference on machine learning

    Trouillon, T., Welbl, J., Riedel, S., Gaussier, É., Bouchard, G.: Complex embed- dings for simple link prediction. In: International conference on machine learning. pp. 2071–2080. PMLR (2016)

    2016

  25. [33]

    arXiv preprint arXiv:1710.10903 (2017)

    Veličković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)

    2017 arXiv

  26. [34]

    In: Proceedings of the 27th ACM International Conference on Information and Knowledge Management

    Wang, H., Zhang, F., Wang, J., Zhao, M., Li, W., Xie, X., Guo, M.: RippleNet: Propagating User Preferences on the Knowledge Graph for Recommender Systems. In: Proceedings of the 27th ACM International Conference on Information and Knowledge Management. CIKM ’18 (2018)

    2018

  27. [35]

    In: Proceedings of the World Wide Web Con- ference

    Wang, H., Zhao, M., Xie, X., Li, W., Guo, M.: Knowledge graph convolutional networks for recommender systems. In: Proceedings of the World Wide Web Con- ference. pp. 3307–3313. WWW’19 (2019)

    2019

  28. [36]

    In: Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval

    Wang, X., He, X., Wang, M., Feng, F., Chua, T.S.: Neural graph collaborative filtering. In: Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval. pp. 165–174 (2019)

    2019

  29. [37]

    In: Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval

    Wang, X., He, X., Wang, M., Feng, F., Chua, T.S.: Neural graph collaborative filtering. In: Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval. p. 165–174. SIGIR’19 (2019), https://doi.org/10.1145/3331184.3331267

    2019

  30. [38]

    In: World Wide Conference

    Wang, X., Ji, H., Shi, C., Wang, B., Ye, Y., Cui, P., Yu, P.S.: Heterogeneous graph attention network. In: World Wide Conference. pp. 2022–2032 (2019)

    2019

  31. [39]

    ACM Computing Surveys55(5), 1–37 (2022)

    Wu, S., Sun, F., Zhang, W., Xie, X., Cui, B.: Graph neural networks in recom- mender systems: a survey. ACM Computing Surveys55(5), 1–37 (2022)

    2022

  32. [40]

    CoRRabs/2406.01032 (2024)

    Xu, J., Wu, Z., Lin, M., Zhang, X., Wang, S.: LLM and GNN are complemen- tary: Distilling LLM for multimodal graph learning. CoRRabs/2406.01032 (2024). https://doi.org/10.48550/ARXIV.2406.01032, https://doi.org/10.48550/ arXiv.2406.01032

    2024 doi arXiv

  33. [41]

    arXiv preprint arXiv:1412.6575 (2014)

    Yang, B., Yih, W.t., He, X., Gao, J., Deng, L.: Embedding entities and relations for learning and inference in knowledge bases. arXiv preprint arXiv:1412.6575 (2014)

    2014 arXiv

  34. [42]

    In: Proceedings of the AAAI Conference on Artificial Intel- ligence

    Yang, X., Yan, M., Pan, S., Ye, X., Fan, D.: Simple and efficient heterogeneous graph neural network. In: Proceedings of the AAAI Conference on Artificial Intel- ligence. vol. 37, pp. 10816–10824 (2023)

    2023

  35. [43]

    Knowledge-Based Systems204, 106194 (2020)

    Yang, Z., Dong, S.: HAGERec: Hierarchical Attention Graph Convolutional Network Incorporating Knowledge Graph for Explainable Recommendation. Knowledge-Based Systems204, 106194 (2020)

    2020

  36. [44]

    In: Proceedings of the 24th ACM SIGKDD International Conference on Knowl- edge Discovery & Data Mining

    Ying, R., He, R., Chen, K., Eksombatchai, P., Hamilton, W.L., Leskovec, J.: Graph convolutional neural networks for web-scale recommender systems. In: Proceedings of the 24th ACM SIGKDD International Conference on Knowl- edge Discovery & Data Mining. p. 974–983. KDD ’18, New Y...

    2018

  37. [45]

    arXiv preprint arXiv:2011.09679 (2020)

    Yu, L., Shen, J., Li, J., Lerer, A.: Scalable graph neural networks for heterogeneous graphs. arXiv preprint arXiv:2011.09679 (2020)

    2020 arXiv

  38. [46]

    Nucleic acids research52(D1) (2024)

    Zdrazil, B., Felix, E., Hunter, F., Manners, E.J., Blackshaw, J., Corbett, S., de Veij, M., Ioannidis, H., Lopez, D.M., Mosquera, J.F., et al.: The chembl database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods. Nucleic acids resea...

    2024

  39. [47]

    In: ACM SIGKDD international conference on knowledge discov- ery & data mining

    Zhang, C., Song, D., Huang, C., Swami, A., Chawla, N.V.: Heterogeneous graph neural network. In: ACM SIGKDD international conference on knowledge discov- ery & data mining. pp. 793–803 (2019)

    2019

  40. [48]

    arXiv preprint arXiv:1904.12058 (2019)

    Zhang, M., Chen, Y.: Inductive matrix completion based on graph neural networks. arXiv preprint arXiv:1904.12058 (2019)

    2019 arXiv

  41. [49]

    AI open1, 57–81 (2020)

    Zhou, J., Cui, G., Hu, S., Zhang, Z., Yang, C., Liu, Z., Wang, L., Li, C., Sun, M.: Graph neural networks: A review of methods and applications. AI open1, 57–81 (2020)

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.