Pith. sign in

REVIEW 2 major objections 6 minor 44 references

H$^2$GFM: Towards unifying Homogeneity and Heterogeneity on Text-Attributed Graphs

T0 review · 2 major / 6 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 H2GFM claims a single pretrained model can handle both homogeneous and heterogeneous text-attributed graphs, transferring zero-shot to unseen graph types.

desk verdict A promising unified GFM for homogeneous and heterogeneous TAGs, but the link prediction evaluations need a clearer edge-masking protocol before the headline numbers can be trusted. read the letter →

arxiv 2506.08298 v2 pith:CODYUFW2 submitted 2025-06-10 cs.LG cs.SI

Trung-Kien Nguyen , Heng Ping , Shixuan Li , Peiyu Zhang , Nikos Kanakaris , Nicholas Kotov , Paul Bogdan This is my paper
classification cs.LGcs.SI
keywords graphfoundationmodeltext-attributedgraphsheterogeneousnodeclassificationlinkpredictionmixtureofexpertscontextrandom-walkmeta-path
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 introduces H2GFM, a graph foundation model trained once on a mix of homogeneous and heterogeneous text-attributed graphs and then applied to new graphs with little or no adaptation. It claims this works by converting every node and relation type into a shared language embedding space, so graphs with different schemas become comparable. Around each target node it samples random-walk meta-paths as a context graph, encodes multi-hop relation semantics with a weighted sum of meta-relation embeddings, and processes the result with a mixture of context-adaptive graph transformers. The central claim is that this design lets one model outperform per-graph and other foundation baselines on both node classification and link prediction, in co-training, zero-shot transfer, and fine-tuning settings.

What carries the argument

The central mechanism is the context graph plus context encoding: for each node, random walks of length up to Lmax=4 are sampled and interpreted as meta-paths; each edge's textual meta-relation (endpoint types plus relation type, rendered as a sentence) is embedded with Sentence-BERT, and a path embedding is computed as hp = sum_{m=1}^L (1/m) r_{e_m}, a hop-decayed sum of unit meta-relation embeddings. This encoding gives different representations to paths of different lengths and relation compositions. The transformed embeddings are then processed by a context-adaptive graph transformer (CGT) that uses the path embedding to modulate attention and feature-wise linear modulation (FiLM) of neighbor messages; a sparse top-k gating network selects among a pool of CGT experts per node, letting different structural patterns be handled by different experts.

What would settle it

Take a heterogeneous graph where the defining relation, such as a collaboration pattern or a multi-hop citation chain, extends beyond 4 steps. Train H2GFM with Lmax fixed at 4 and again with Lmax increased to 8 on a held-out subset; if the longer-horizon model does not outperform the fixed model, or if the context encoding with hop weights adds noise versus a simple 1-hop aggregation, the central claim that random-walk context of length up to 4 captures informative higher-order semantics is contradicted.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single model, H2GFM, generalizes across both homogeneous TAGs (single node/edge type) and heterogeneous TAGs (multiple node/edge types) by unifying meta-relations in text space and learning from context graphs. The authors show that encoding a target node's neighborhood as multi-hop, multi-relation context, rather than as a flat 1-hop neighborhood, plus routing through a mixture of expert transformers, yields representations that transfer to unseen graph datasets. In their experiments, H2GFM outperforms existing graph foundation models and self-supervised baselines on node classification and link prediction across six co-training datasets and six unseen cross-training datasets, with the largest zero-shot gains on heterogeneous graphs.

Load-bearing premise

Random walks of maximum length 4, with no dataset-specific meta-path design, are assumed to sample context that captures the relations that matter for every graph type.

Editorial extensions

If this is right

  • A single pretrained H2GFM can be applied to a new text-attributed graph with no per-dataset meta-path design, because random walks replace hand-crafted meta-paths.
  • Co-training on homogeneous and heterogeneous graphs from related domains improves performance on both, suggesting that cross-genre training is beneficial rather than harmful.
  • Zero-shot node classification and link prediction on unseen graphs are substantially improved over prior foundation baselines, especially for heterogeneous graphs.
  • Because context graph sampling is inductive and mini-batch friendly, the approach can scale to large graphs and to graphs whose structure evolves over time.

Reading between the lines

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

  • The hop decay 1/m in the context encoding is a strong inductive bias: it assumes closer relations matter more, which may not hold in graphs where distant same-type relations determine the label; a learned decay could adapt to such cases.
  • The text template for meta-relations ('A paper is written by an author') puts heavy weight on the quality of the LLM's understanding of relation semantics; embedding models that conflate similar relation phrases could blur distinct meta-relations.
  • If random-walk context sampling is biased by node degree, high-degree nodes would dominate context graphs, making the sampled context less representative of the actual local structure; a degree-normalized sampling variant is a straightforward testable extension.
  • The method's reliance on Sentence-BERT for both node and relation embeddings suggests that improvements to the text encoder, or fine-tuning it on graph-aware objectives, could yield further gains without architectural changes.
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

2 major / 6 minor

Summary. The manuscript introduces H2GFM, a text-space graph foundation model that aims to generalize across both homogeneous and heterogeneous text-attributed graphs. It unifies node and edge type information as textual meta-relations embedded with Sentence-BERT, samples multi-hop context graphs by random-walk meta-paths, encodes paths with a hop-weighted sum (Eq. 3), and processes them with a sparse mixture of context-adaptive graph transformers that use FiLM-conditioned attention. The model is evaluated on node classification and link prediction under co-training, zero-shot cross-training, and fine-tuning settings across six co-training and six cross-training datasets, with ablations, parameter sensitivity, and scalability studies. The paper's central claim is in Section 5.2: H2GFM significantly outperforms comparative baselines on both node classification and link prediction.

Significance. If the empirical claims hold, this is a meaningful step toward graph foundation models that cover heterogeneous text-attributed graphs, a gap relative to prior text-space GFMs focused on homogeneous graphs. The paper's strengths include the breadth of the evaluation (12 datasets, three learning scenarios), the transparent use of prior random-walk sampling [21], the explicit context encoding design, and the reported parameter efficiency and linear-time scaling. I do not view the absence of first-principles derivation as a flaw, since the contribution is empirical. The main uncertainty is whether the link-prediction numbers are inflated by leakage in context-graph construction; this must be resolved before the central LP claim can be accepted.

major comments (2)
  1. [Section 4.1 and Section 5.1] The link-prediction results in Tables 2-4, including the large zero-shot improvement on Pubmed (AUC 36.95 to 63.31), are only valid if validation/test edges are masked when H2GFM constructs context graphs. The paper states that 'models are only trained on the subgraph constructed from train edges,' but it does not state that random-walk context sampling for validation/test nodes also uses only train edges. Since H2GFM aggregates up to Lmax=4 hops, a positive test edge (u,v) can enter u's context graph and be directly observed in h_u, and similarly for h_v. Please specify the exact graph used for context sampling at evaluation time, and rerun or confirm the LP experiments with validation/test edges excluded from context subgraphs. Also specify whether the 2-hop negative edges used for LP are drawn from the training subgraph only; otherwise the 1:1 positive/negative sets are not comparable across models.
  2. [Section 5.2, Table 2] The blanket claim that H2GFM 'significantly outperforms other comparative baselines on both node classification and link prediction tasks' is not supported on He_Croval LP, where H2GFM (97.93 ± 0.96) is slightly below OFA (98.28 ± 0.95), with the paper's own 'Improve' column showing -0.36%. Please qualify this claim and provide pairwise significance tests or confidence intervals for the headline comparisons, since the current text uses 'significantly' without statistical support.
minor comments (6)
  1. [Eq. (4)] The text defines learnable parameters W_q, W_k, and W_r, but Eq. (4) uses W_p for the path embedding projection; please align the notation.
  2. [Algorithm 1, line 14] Line 14 says 'Calculate final h_u by Eq. (12)', but Eq. (12) defines the TopK gating operator; the final node representation is computed by Eq. (9).
  3. [Table 3 caption] The caption says 'link prediction (PD)' but should say 'link prediction (LP)'.
  4. [Appendix A] The text says 'Since number of experts k is small (8 in our setting)', but k denotes the number of active experts, which is 4 in Section 5.1; the total number of experts is n=8.
  5. [Section 5.3, Fig. 2] The 'no context graph' ablation changes both the sampling radius and the use of meta-path/context encoding, so the drop cannot be attributed solely to removing the context graph; a 1-hop variant that retains the relation encoding would provide a cleaner control.
  6. [Eq. (13)] The notation h_y is used for node classification, but for link prediction the predictor takes (h_u,h_v) and y is a scalar label; please clarify that h_y applies only to the NC term.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper makes empirical claims supported by external benchmarks; its hand-designed components are not derived from the quantities they predict.

full rationale

H2GFM does not present a first-principles derivation of a prediction from an input that already contains that prediction. The central claims in Section 5.2 are empirical comparisons on public benchmarks (datasets from [3] and [17], baselines such as OFA, GFT, AnyGraph, LlaGA, GraphMAE). The model components are explicitly constructed rather than fitted to the target: Eq. (3) defines a context encoding as a fixed 1/m-weight sum of meta-relation embeddings, Eq. (5) applies softmax attention, Eq. (8) aggregates with attention weights, and Eqs. (9)-(12) implement a standard MoE gating mechanism. None of these equations defines a prediction in terms of a fitted parameter that is then renamed as a prediction. The random-walk sampling is adopted from the first author's prior work [21], but it is an openly cited methodological choice rather than an unverified uniqueness theorem, and the paper's own ablation study (Fig. 2) separately tests the contribution of the context graph. The concern about possible validation/test-edge leakage in link prediction is a correctness or evaluation-protocol issue, not a circularity by construction. No equation in the paper reduces to its own input, so the circularity score is 0.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The paper introduces no physical or domain entities. Its contributions are architectural, and the main load-bearing choices are hand-chosen hyperparameters and the domain assumptions listed above.

free parameters (7)
  • Number of random walks per node N = 50
    Chosen by hand for all experiments (Section 5.1); the paper does not report sensitivity to N, so the central results depend on this choice.
  • Maximum meta-path length Lmax = 4
    Chosen by hand (Section 5.1); the paper tests Lmax 1 to 4 and reports larger values tend to be more consistent, but the exact value affects performance in Fig. 4.
  • Total number of CGT experts n = 8
    Chosen by hand (Section 5.1); Fig. 5(b) shows performance varies with n.
  • Number of active experts k = 4
    Chosen by hand (Section 5.1); Fig. 5(a) shows active expert count affects accuracy and cost.
  • CGT hidden dimension = 768
    Chosen by hand (Section 5.1).
  • Context encoding weighting schedule 1/m = 1/m for hop m
    Hand-designed in Eq. (3), aligns with homophily assumption; not learned or fitted.
  • CGT layer count = 1
    Chosen by hand (Section 5.1); the paper argues one layer suffices because variable-length paths carry long-range information.
assumptions (4)
  • domain assumption Sentence-BERT text embeddings provide a unified semantic space across different graphs and relation types.
    Used in Eqs. (1) and (2); if textual descriptions are inconsistent across datasets, the unified space breaks down.
  • domain assumption Random-walk sampled meta-paths of maximum length 4 capture meaningful high-order semantics without manual meta-path design.
    Central to context graph construction in Section 4.1.
  • domain assumption Closer relations are more important than farther relations, motivating the 1/m weighting.
    Stated in Section 4.1: 'aligns with the homophily principle commonly observed in graphs.' This may fail on heterophilic heterogeneous graphs, which the paper itself notes in the introduction.
  • domain assumption The top-k gating mechanism with added Gaussian noise maintains load balance across experts.
    Standard MoE assumption in Eq. (11); if load balancing fails, some experts may be undertrained.

how reviews work

0 comments
Cite this review

Pith. "Pith review of H$^2$GFM: Towards unifying Homogeneity and Heterogeneity on Text-Attributed Graphs." pith.science (2026). https://pith.science/paper/CODYUFW2

@misc{pith2026250608298,
  author       = {Pith},
  title        = {Pith review of: H$^2$GFM: Towards unifying Homogeneity and Heterogeneity on Text-Attributed Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CODYUFW2}},
  note         = {Machine review of arXiv:2506.08298}
}
abstract

The growing interests and applications of graph learning in diverse domains have propelled the development of a unified model generalizing well across different graphs and tasks, known as the Graph Foundation Model (GFM). Existing research has leveraged text-attributed graphs (TAGs) to tackle the heterogeneity in node features among graphs. However, they primarily focus on homogeneous TAGs (HoTAGs), leaving heterogeneous TAGs (HeTAGs), where multiple types of nodes/edges reside, underexplored. To enhance the capabilities and applications of GFM, we introduce H$^2$GFM, a novel framework designed to generalize across both HoTAGs and HeTAGs. Our model projects diverse meta-relations among graphs under a unified textual space, and employs a context encoding to capture spatial and higher-order semantic relationships. To achieve robust node representations, we propose a novel context-adaptive graph transformer (CGT), effectively capturing information from both context neighbors and their relationships. Furthermore, we employ a mixture of CGT experts to capture the heterogeneity in structural patterns among graph types. Comprehensive experiments on a wide range of HoTAGs and HeTAGs as well as learning scenarios demonstrate the effectiveness of our model.

Figures

Figures reproduced from arXiv: 2506.08298 by the authors.

Figure 1
Figure 1. Overview framework of H2GFM Mixture of Experts (MoE). MoE [25, 4] approach comprises M sub-networks (experts) such as MLP, where each network aims to specialize in a specific sub-domain. Based on the input, a gating (routing) mechanism is utilized to dynamically assign proper experts for the process. In graph learning, MoE has been explored to enhance the capabilities of GNNs in supervised settings [32, 19] or tackl… view at source ↗
Figure 1
Figure 1. Fig.1. Given HoTAG/HeTAG datasets, a context graph is sampled for each target node, consisting [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Ablation studies on co-training datasets. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: Visualizaion of node features of all datasets [PITH_FULL_IMAGE:figures/full_fig_p014_3.png]
Figure 5
Figure 5. Figure 5: Parameter studies. • Pre-training: Loss weights β1 = 10, β2 = 100, β3 = 1, β4 = 0.01; orthogonal regularizer λ = 1; AdamW optimizer with learning rate 1e-4 and weight decay 1e-5; 25 epochs; batch size 1024; edge and node feature drop rates of 0.2 • Fine-tuning: Learnin…
Figure 4
Figure 4. Figure 4: Impact of maximum path length. We study the impact of maximum length Lmax for the random-walk meta-path sampling strategy. From Fig.4, while performance varies across datasets, larger values of Lmax tend to achieve more consistent results on a wide range of data, sugge…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 24 canonical work pages

  1. [21]

    Link prediction on latent heterogeneous graphs

    Trung-Kien Nguyen, Zemin Liu, and Yuan Fang. Link prediction on latent heterogeneous graphs. InProceedings of the ACM Web Conference 2023, pages 263–273, 2023

    work page 2023

  2. [1]

    Unifying structural descriptors for biological and bioinspired nanoscale complexes.Nature Computational Science, 2(4):243–252, 2022

    Minjeong Cha, Emine Sumeyra Turali Emre, Xiongye Xiao, Ji-Young Kim, Paul Bogdan, J Scott VanEpps, Angela Violi, and Nicholas A Kotov. Unifying structural descriptors for biological and bioinspired nanoscale complexes.Nature Computational Science, 2(4):243–252, 2022

    work page 2022

  3. [2]

    Llaga: Large language and graph assistant.arXiv preprint arXiv:2402.08170, 2024

    Runjin Chen, Tong Zhao, Ajay Jaiswal, Neil Shah, and Zhangyang Wang. Llaga: Large language and graph assistant.arXiv preprint arXiv:2402.08170, 2024

    arXiv 2024

  4. [3]

    Text-space graph foundation models: Comprehensive benchmarks and new insights.arXiv preprint arXiv:2406.10727, 2024

    Zhikai Chen, Haitao Mao, Jingzhe Liu, Yu Song, Bingheng Li, Wei Jin, Bahare Fatemi, Anton Tsitsulin, Bryan Perozzi, Hui Liu, et al. Text-space graph foundation models: Comprehensive benchmarks and new insights.arXiv preprint arXiv:2406.10727, 2024

    arXiv 2024

  5. [4]

    Towards understanding the mixture-of-experts layer in deep learning.Advances in neural information processing systems, 35:23049–23062, 2022

    Zixiang Chen, Yihe Deng, Yue Wu, Quanquan Gu, and Yuanzhi Li. Towards understanding the mixture-of-experts layer in deep learning.Advances in neural information processing systems, 35:23049–23062, 2022

    2022

  6. [5]

    Adam: A method for stochastic optimization.(No Title), 2014

    Kingma Diederik. Adam: A method for stochastic optimization.(No Title), 2014

    2014

  7. [6]

    Unigraph: Learning a unified cross-domain foundation model for text-attributed graphs.arXiv preprint arXiv:2402.13630, 2024

    Yufei He, Yuan Sui, Xiaoxin He, and Bryan Hooi. Unigraph: Learning a unified cross-domain foundation model for text-attributed graphs.arXiv preprint arXiv:2402.13630, 2024

    arXiv 2024

  8. [7]

    Graphmae: Self-supervised masked graph autoencoders

    Zhenyu Hou, Xiao Liu, Yukuo Cen, Yuxiao Dong, Hongxia Yang, Chunjie Wang, and Jie Tang. Graphmae: Self-supervised masked graph autoencoders. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 594–604, 2022

    2022

Show all 44 references
  1. [8]

    Heterogeneous graph transformer

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun. Heterogeneous graph transformer. In Proceedings of the web conference 2020, pages 2704–2710, 2020

    2020

  2. [9]

    Prodigy: Enabling in-context learning over graphs.Advances in Neural Information Processing Systems, 36:16302–16317, 2023

    Qian Huang, Hongyu Ren, Peng Chen, Gregor Kržmanc, Daniel Zeng, Percy S Liang, and Jure Leskovec. Prodigy: Enabling in-context learning over graphs.Advances in Neural Information Processing Systems, 36:16302–16317, 2023

    2023

  3. [10]

    In vitro neurons learn and exhibit sentience when embodied in a simulated game-world.Neuron, 110(23):3952–3969, 2022

    Brett J Kagan, Andy C Kitchen, Nhi T Tran, Forough Habibollahi, Moein Khajehnejad, Bradyn J Parker, Anjali Bhat, Ben Rollo, Adeel Razi, and Karl J Friston. In vitro neurons learn and exhibit sentience when embodied in a simulated game-world.Neuron, 110(23):3952–3969, 2022

    2022

  4. [11]

    Semi-supervised classification with graph convolutional networks

    Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. InICLR, 2017

    2017

  5. [12]

    When heterophily meets heterogeneity: New graph benchmarks and effective methods.arXiv preprint arXiv:2407.10916, 2024

    Junhong Lin, Xiaojie Guo, Shuaicheng Zhang, Dawei Zhou, Yada Zhu, and Julian Shun. When heterophily meets heterogeneity: New graph benchmarks and effective methods.arXiv preprint arXiv:2407.10916, 2024

    2024 arXiv

  6. [13]

    One for all: Towards training one graph model for all classification tasks.arXiv preprint arXiv:2310.00149, 2023

    Hao Liu, Jiarui Feng, Lecheng Kong, Ningyue Liang, Dacheng Tao, Yixin Chen, and Muhan Zhang. One for all: Towards training one graph model for all classification tasks.arXiv preprint arXiv:2310.00149, 2023

    2023 arXiv

  7. [15]

    Graph foundation models: Concepts, opportunities and challenges.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Jiawei Liu, Cheng Yang, Zhiyuan Lu, Junze Chen, Yibo Li, Mengmei Zhang, Ting Bai, Yuan Fang, Lichao Sun, Philip S Yu, et al. Graph foundation models: Concepts, opportunities and challenges.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  8. [16]

    One model for one graph: A new perspective for pretraining with cross-domain graphs.arXiv preprint arXiv:2412.00315, 2024

    Jingzhe Liu, Haitao Mao, Zhikai Chen, Wenqi Fan, Mingxuan Ju, Tong Zhao, Neil Shah, and Jiliang Tang. One model for one graph: A new perspective for pretraining with cross-domain graphs.arXiv preprint arXiv:2412.00315, 2024

    2024 arXiv

  9. [17]

    Multi-scale heterogeneous text-attributed graph datasets from diverse domains.arXiv preprint arXiv:2412.08937, 2024

    Yunhui Liu, Qizhuo Xie, Jinwei Shi, Jiaxu Shen, and Tieke He. Multi-scale heterogeneous text-attributed graph datasets from diverse domains.arXiv preprint arXiv:2412.08937, 2024. 11

    2024 arXiv

  10. [18]

    Are we really making much progress? revisiting, benchmarking and refining heterogeneous graph neural networks

    Qingsong Lv, Ming Ding, Qiang Liu, Yuxiang Chen, Wenzheng Feng, Siming He, Chang Zhou, Jianguo Jiang, Yuxiao Dong, and Jie Tang. Are we really making much progress? revisiting, benchmarking and refining heterogeneous graph neural networks. InProceedings of the 27th ACM SIGKDD ...

    2021

  11. [19]

    Mixture of link predictors on graphs.arXiv preprint arXiv:2402.08583, 2024

    Li Ma, Haoyu Han, Juanhui Li, Harry Shomer, Hui Liu, Xiaofeng Gao, and Jiliang Tang. Mixture of link predictors on graphs.arXiv preprint arXiv:2402.08583, 2024

    2024 arXiv

  12. [20]

    Birds of a feather: Homophily in social networks.Annual review of sociology, 27(1):415–444, 2001

    Miller McPherson, Lynn Smith-Lovin, and James M Cook. Birds of a feather: Homophily in social networks.Annual review of sociology, 27(1):415–444, 2001

    2001

  13. [22]

    Unsupervised attributed multiplex network embedding

    Chanyoung Park, Donghyun Kim, Jiawei Han, and Hwanjo Yu. Unsupervised attributed multiplex network embedding. InProceedings of the AAAI conference on artificial intelligence, volume 34, pages 5371–5378, 2020

    2020

  14. [23]

    Film: Visual reasoning with a general conditioning layer

    Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. InProceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  15. [24]

    Sentence-bert: Sentence embeddings using siamese bert-networks.arXiv preprint arXiv:1908.10084, 2019

    N Reimers. Sentence-bert: Sentence embeddings using siamese bert-networks.arXiv preprint arXiv:1908.10084, 2019

    1908 arXiv

  16. [25]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.arXiv preprint arXiv:1701.06538, 2017

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.arXiv preprint arXiv:1701.06538, 2017

    2017 arXiv

  17. [26]

    Pathsim: Meta path-based top-k similarity search in heterogeneous information networks.PVLDB, 4(11):992–1003, 2011

    Yizhou Sun, Jiawei Han, Xifeng Yan, Philip S Yu, and Tianyi Wu. Pathsim: Meta path-based top-k similarity search in heterogeneous information networks.PVLDB, 4(11):992–1003, 2011

    2011

  18. [27]

    Higpt: Heterogeneous graph language model

    Jiabin Tang, Yuhao Yang, Wei Wei, Lei Shi, Long Xia, Dawei Yin, and Chao Huang. Higpt: Heterogeneous graph language model. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2842–2853, 2024

    2024

  19. [28]

    Structural analysis of nanoscale network materials using graph theory.ACS nano, 15(8):12847–12859, 2021

    Drew A Vecchio, Samuel H Mahler, Mark D Hammig, and Nicholas A Kotov. Structural analysis of nanoscale network materials using graph theory.ACS nano, 15(8):12847–12859, 2021

    2021

  20. [29]

    Graph attention networks

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. InICLR, 2018

    2018

  21. [30]

    Deep graph infomax.arXiv preprint arXiv:1809.10341, 2018

    Petar Veliˇckovi´c, William Fedus, William L Hamilton, Pietro Liò, Yoshua Bengio, and R Devon Hjelm. Deep graph infomax.arXiv preprint arXiv:1809.10341, 2018

    2018 arXiv

  22. [31]

    Llms as zero-shot graph learners: Alignment of gnn representations with llm token embeddings.Advances in Neural Information Processing Systems, 37:5950–5973, 2025

    Duo Wang, Yuan Zuo, Fengzhi Li, and Junjie Wu. Llms as zero-shot graph learners: Alignment of gnn representations with llm token embeddings.Advances in Neural Information Processing Systems, 37:5950–5973, 2025

    2025

  23. [32]

    Graph mixture of experts: Learning on large-scale graphs with explicit diversity modeling.Advances in Neural Information Processing Systems, 36, 2024

    Haotao Wang, Ziyu Jiang, Yuning You, Yan Han, Gaowen Liu, Jayanth Srinivasa, Ramana Kompella, Zhangyang Wang, et al. Graph mixture of experts: Learning on large-scale graphs with explicit diversity modeling.Advances in Neural Information Processing Systems, 36, 2024

    2024

  24. [33]

    Biomorphic structural batteries for robotics.Science robotics, 5(45):eaba1912, 2020

    Mingqiang Wang, Drew Vecchio, Chunyan Wang, Ahmet Emre, Xiongye Xiao, Zaixing Jiang, Paul Bogdan, Yudong Huang, and Nicholas A Kotov. Biomorphic structural batteries for robotics.Science robotics, 5(45):eaba1912, 2020

    2020

  25. [34]

    A survey on heterogeneous graph embedding: methods, techniques, applications and sources.IEEE transactions on big data, 9(2):415–436, 2022

    Xiao Wang, Deyu Bo, Chuan Shi, Shaohua Fan, Yanfang Ye, and Philip S Yu. A survey on heterogeneous graph embedding: methods, techniques, applications and sources.IEEE transactions on big data, 9(2):415–436, 2022

    2022

  26. [35]

    Het- erogeneous graph attention network

    Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu. Het- erogeneous graph attention network. InThe world wide web conference, pages 2022–2032, 2019. 12

    2022

  27. [36]

    Self-supervised heterogeneous graph neural network with co-contrastive learning

    Xiao Wang, Nian Liu, Hui Han, and Chuan Shi. Self-supervised heterogeneous graph neural network with co-contrastive learning. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pages 1726–1736, 2021

    2021

  28. [37]

    Gft: Graph foundation model with transferable tree vocabulary.Advances in Neural Information Processing Systems, 37:107403–107443, 2025

    Zehong Wang, Zheyuan Zhang, Nitesh Chawla, Chuxu Zhang, and Yanfang Ye. Gft: Graph foundation model with transferable tree vocabulary.Advances in Neural Information Processing Systems, 37:107403–107443, 2025

    2025

  29. [38]

    Graph neural networks in recom- mender systems: a survey.ACM Computing Surveys, 55(5):1–37, 2022

    Shiwen Wu, Fei Sun, Wentao Zhang, Xu Xie, and Bin Cui. Graph neural networks in recom- mender systems: a survey.ACM Computing Surveys, 55(5):1–37, 2022

    2022

  30. [39]

    Layer-by-layer assembled nanowire networks enable graph-theoretical design of multifunctional coatings

    Wenbing Wu, Alain Kadar, Sang Hyun Lee, Hong Ju Jung, Bum Chul Park, Jeffery E Raymond, Thomas K Tsotsis, Carlos ES Cesnik, Sharon C Glotzer, Valerie Goss, et al. Layer-by-layer assembled nanowire networks enable graph-theoretical design of multifunctional coatings. Matter, 8(1), 2025

    2025

  31. [40]

    A comprehensive survey on graph neural networks.IEEE transactions on neural networks and learning systems, 32(1):4–24, 2020

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and Philip S Yu. A comprehensive survey on graph neural networks.IEEE transactions on neural networks and learning systems, 32(1):4–24, 2020

    2020

  32. [41]

    Anygraph: Graph foundation model in the wild.arXiv preprint arXiv:2408.10700, 2024

    Lianghao Xia and Chao Huang. Anygraph: Graph foundation model in the wild.arXiv preprint arXiv:2408.10700, 2024

    2024 arXiv

  33. [42]

    How powerful are graph neural networks? InICLR, 2019

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? InICLR, 2019

    2019

  34. [43]

    Network science characteristics of brain-derived neuronal cultures deciphered from quantitative phase imaging data.Scientific reports, 10(1):15078, 2020

    Chenzhong Yin, Xiongye Xiao, Valeriu Balaban, Mikhail E Kandel, Young Jae Lee, Gabriel Popescu, and Paul Bogdan. Network science characteristics of brain-derived neuronal cultures deciphered from quantitative phase imaging data.Scientific reports, 10(1):15078, 2020

    2020

  35. [44]

    Graph contrastive learning with augmentations.Advances in neural information processing systems, 33:5812–5823, 2020

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations.Advances in neural information processing systems, 33:5812–5823, 2020

    2020

  36. [45]

    Author,"

    Hanqing Zeng, Hanjia Lyu, Diyi Hu, Yinglong Xia, and Jiebo Luo. Mixture of weak & strong experts on graphs.arXiv preprint arXiv:2311.05185, 2023. Appendix A Algorithm and Complexity We outline the model training for H2GFM in Algorithm 1. In line 1, we initialize the model para...

    2023 arXiv

Pith tools

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