Pith. sign in

REVIEW 3 major objections 6 minor 52 references

Disentangling Homophily and Heterophily in Multimodal Graph Clustering

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A new framework, DMGC, disentangles multimodal graphs into homophilic and heterophilic views and reports the best unsupervised clustering results on six benchmarks.

desk verdict A useful framework for an under-studied task, but the headline numbers are not trustworthy until the hyperparameter selection and run-to-run variance are reported honestly. read the letter →

arxiv 2507.15253 v1 pith:WQA6OXFC submitted 2025-07-21 cs.AI cs.LGcs.SI

classification cs.AIcs.LGcs.SI
keywords multimodalgraphclusteringunsupervisedlearninghomophilyandheterophilydisentanglementdual-frequencyfusionmultiplexself-supervisedalignment
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 sets out to solve clustering on multimodal graphs—networks whose nodes carry several feature types such as text and images—without labels, when the edges mix same-class and different-class connections. It argues that this hybrid neighborhood pattern breaks methods that assume either pure homophily or pure heterophily. To fix that, the proposed DMGC framework first disentangles the observed graph into a cross-modality homophilic graph and per-modality heterophilic graphs, then filters the two views through low-pass and high-pass operations and fuses them adaptively. Self-supervised reconstruction, contrastive alignment, and clustering objectives train the model without labels. On six real-world datasets, including one with nearly 98,000 nodes, DMGC reports the best clustering accuracy, normalized mutual information, and adjusted Rand index against prior graph and multimodal baselines.

What carries the argument

The load-bearing machinery is disentangled graph construction paired with dual-frequency fusion. Starting from per-modality aggregated features, the framework forms a cross-modality consensus feature $\tilde{X}=\sum_i w_i\tilde{X}_i$, builds a homophilic kNN graph $A_l=\mathrm{Top}(\tilde{S},k_l)$ from its normalized similarity, and builds per-modality heterophilic graphs $A_h^i=\mathrm{Top}(1-\tilde{S}_i,k_h)$ from the complements of each modality's similarity. Low-pass filtering is applied with the normalized homophilic adjacency $\hat{A}_l$, while high-pass filtering uses the heterophilic Laplacian $\hat{L}_h^i=I-\hat{D}^{-1/2}\hat{A}_h^i\hat{D}^{-1/2}$. A trainable coefficient $\gamma_i$ balances the two frequency views per modality, and attention weights $\beta_i$ fuse modalities into a single representation $Z$. InfoNCE losses align the low-pass and high-pass views with the fused representation and align modality pairs, while a KL-divergence clustering loss sharpens the final assignments.

What would settle it

On a labeled multimodal graph, compute the precision of the homophilic neighbor set $\mathrm{Top}(\tilde{S},k_l)$, defined as the fraction of chosen neighbors that share the node's class. If that fraction is at or below random chance while the reported clustering gains persist, the homophilic branch is not the carrier of performance; conversely, permuting the heterophilic neighbor sets $\mathrm{Top}(1-\tilde{S}_i,k_h)$ and observing no change in ACC or NMI would show that the heterophily branch is decorative.

Watch

Extended reading notes

Core claim

The paper's central claim is that hybrid neighborhoods in multimodal graphs can be disentangled into two complementary structures: a consensus homophilic graph built from weighted fused modality features, and modality-specific heterophilic graphs built from the complements of each modality's similarity matrix. The framework then treats the problem as one of dual-frequency filtering: low-pass filtering on the homophilic graph extracts shared, category-consistent semantics, while high-pass filtering through the heterophilic graph Laplacian preserves inter-class distinctions. Adaptive fusion with a trainable per-modality coefficient, attention-based multimodal fusion, and contrastive and reconstruction losses produce representations for clustering. The paper reports state-of-the-art results across multimodal and multi-relational benchmarks and positions the work as the first systematic unsupervised pipeline for raw multimodal graph data.

Load-bearing premise

The load-bearing premise is that the kNN graphs built from the weighted consensus features and from each modality's complement similarity actually separate class-consistent from class-divergent neighbors; if those neighbor lists are contaminated, the noise propagates through the dual-frequency fusion and every loss built on it.

Editorial extensions

If this is right

  • Graph-based multimodal clustering can be performed without labels directly on raw topology, rather than only on feature matrices.
  • Explicitly modeling heterophily as a complementary signal, instead of discarding it as noise, becomes a viable design principle for unsupervised graph learning.
  • The same architecture transfers to multi-relational graphs with a single modality, so the benefits are not restricted to genuinely multimodal data.
  • The scalable constructions, locality-sensitive hashing for the homophilic graph and anchor-based sampling for the heterophilic graphs, allow the disentanglement to operate on graphs with tens of thousands of nodes.

Reading between the lines

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

  • A natural extension the paper leaves implicit is learning the modality weights $w_i$ and the neighbor counts $k_l,k_h$ from data rather than fixing them by grid search, since the sensitivity analysis shows that performance depends on these choices.
  • The dual-frequency framing connects to spectral graph theory: low-pass and high-pass filtering correspond to smooth and difference components of node features, so the design could inform broader architectures for heterophilic message passing.
  • If the disentanglement claim is right, a testable consequence is that DMGC's advantage should peak at intermediate homophily ratios and shrink on extremely homophilic or extremely heterophilic graphs; this could be checked by rewiring benchmark graphs to sweep the homophily ratio.
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 / 6 minor

Summary. The paper proposes DMGC, an unsupervised framework for clustering multimodal and multi-relational graphs. The method constructs a cross-modality homophilic graph and per-modality heterophilic graphs from aggregated node features, applies dual low-pass/high-pass graph filtering, fuses the modality-specific representations via attention, and trains the network with reconstruction, contrastive, cross-modality alignment, and clustering losses. Experiments on six datasets (IMDB, Amazon, Ele-Fashion, ACM, DBLP, Yelp) report ACC/NMI/ARI values claimed to be state-of-the-art, with additional ablations and sensitivity analyses.

Significance. If the reported results are reliable, the paper makes a useful contribution by addressing multimodal graph clustering in an unsupervised setting and by explicitly modeling hybrid homophilic/heterophilic neighborhoods. The framework is technically coherent, code is released, and the ablation study shows that each loss component and each disentangled graph type contributes positively. The qualitative class-wise connection matrices in Appendix A.2 provide some evidence that the constructed graphs separate homophilic and heterophilic structure. However, the central state-of-the-art claim is not yet established because the experimental protocol has a model-selection leakage problem and no variance reporting, as detailed in the major comments.

major comments (3)
  1. [Sec. 4.1, 'Parameter setting'; Figs. 5–7] The paper lists hyperparameter ranges (learning rates, L2 weights, k_l, k_h, layers, lambda, mu) but never states how the final hyperparameters are selected. For clustering tasks there is no held-out validation split, so the reported ACC/NMI/ARI numbers appear to be the best over the grid evaluated on the same test labels used to compute the metrics. This constitutes label leakage into model selection and makes the comparison with baselines unfair, directly undermining the state-of-the-art claim in Tables 1 and 2. The sensitivity analysis in Figures 5–7, which plots ACC against lambda, mu, k_l, and k_h, confirms that test labels were used for model selection. Please specify a selection protocol that does not use test labels (e.g., a validation-based criterion such as cluster stability or silhouette score on a held-out portion), or report results for a fixed default configuration and for the full grid as a sensitivity range.
  2. [Tables 1–3] All experimental results are reported without standard deviations, number of seeds, or statistical significance tests. Given the large hyperparameter grid and the stochastic nature of deep clustering training, the point estimates may be unstable; the margins over the best baselines are often small (e.g., 0.94 ACC points on ACM and 1.04 ACC points on Yelp in Table 2), potentially within run-to-run noise. Please report mean and standard deviation over at least five independent runs (with different random seeds) for the proposed method and for the main baselines, and perform a significance test (e.g., paired t-test or Wilcoxon) for the key comparisons. This is essential to support the central claim that DMGC achieves state-of-the-art performance.
  3. [Sec. 3.1.2, Eqs. (2)–(3); Appendix A.2] The disentangled graph construction rests on the heuristic that the kNN graph of the weighted consensus feature X_tilde reliably identifies class-consistent neighbors, and that the complement graphs Top(1 - S_tilde_i, k_h) identify class-divergent neighbors. This assumption is load-bearing because all downstream filtering and alignment losses depend on these graphs. The qualitative class-wise connection matrices in Appendix A.2 (Yelp only) provide some support, but the paper does not quantify how accurately the constructed graphs recover true homophilic and heterophilic edges across the datasets. Please add a quantitative evaluation, such as label-based precision/recall or the homophily ratio of the constructed graphs versus the original graphs for all datasets, to substantiate the disentanglement claim.
minor comments (6)
  1. [Fig. 4 and Sec. 4.3 ('Effectiveness of homophily ratio improvement')] The description of Figure 4 does not clearly state whether the reported homophily ratio is computed on the constructed homophilic graph, on a fused unified graph, or on the bipartite graph used for alignment; please clarify the exact object being measured and how it is generated.
  2. [Sec. 3.1.3 and Sec. 4.1] The scalable constructions using locality-sensitive hashing and anchor-based sampling are introduced, but the experimental section does not state whether the exact or the scalable variant is used for the large Ele-Fashion dataset; please specify which implementation is evaluated and report the actual runtime or memory usage.
  3. [Table 3, rows for w/o L_Cm] The entries for w/o L_Cm on ACM, DBLP, and Yelp are shown as '—' without explanation; since these datasets have only a single modality, the cross-modality alignment loss is not defined, and this should be stated explicitly in the table caption or text.
  4. [Eq. (15)] The function ell in the definition of L_contrast is used before being defined; please define it explicitly (e.g., as the InfoNCE loss for a single pair) and clarify the summation range in the denominator.
  5. [Appendix B] The Texas and Chameleon results are reported as raw numbers without comparison to any baseline, and the robustness claim on Amazon with 20% noise is presented without a baseline or statistical context; please either add comparative experiments or clearly label these as exploratory analyses.
  6. [Sec. 4.4, Fig. 5] The heatmap in Figure 5 is described in detail (e.g., 'reaching its peak at lambda = 10') but the axis labels and colorbar are not visible in the provided text; please ensure the figure is legible and the color scale is explicitly labeled, and align the text with the actual axis values.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; DMGC's components are defined from features and standard objectives, and benchmark comparisons include independent baselines.

full rationale

DMGC's derivation chain is self-contained. The disentangled graphs (Eqs. 2-3) are constructed from modality features and kNN selection, not from clustering labels. The dual-frequency filtering (Eqs. 5-8) applies standard low-pass and high-pass graph filters to these graphs. The alignment objectives (Eqs. 12-15) compare representations within the model, and the clustering loss (Eqs. 16-18) is the standard DEC self-training objective, which is an optimization target rather than a claimed external prediction. The empirical claim of state-of-the-art performance is measured against both independent baselines (HDMI, DMG, O2MAC, VGAE, DGI, GWMAC, GCFAgg) and the authors' own earlier methods (InfoMGF-RA, BTGF, MCGC, MvAGC); the inclusion of independent baselines means the comparison does not reduce to self-citation. No step in the derivation imports a uniqueness theorem or ansatz solely from the authors' prior work. The potential issue of test-label-based hyperparameter selection is a statistical-validity concern, not a circular derivation, and therefore does not affect the circularity score.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The framework relies on a standard set of assumptions about feature similarity, neighborhood semantics, and self-training, plus a large set of hyperparameters that are tuned per dataset. No new physical or mathematical entities are postulated. The main risk is that the disentanglement heuristics (kNN on consensus features, complement graphs) are unproven and dataset-specific, which shifts much of the explanatory burden to the hyperparameter search.

free parameters (8)
  • k_l = grid searched over {10, 15, 20} per dataset
    Number of nearest neighbors in the homophily-enhanced graph (Eq. 3); the paper reports sensitivity but not the final value per dataset.
  • k_h = grid searched over {2, 3, 4, 5, 6} per dataset
    Number of nearest neighbors in the modality-specific heterophily graph (Eq. 3).
  • alpha = not reported
    Residual weight in the propagation step (Eq. 1), which controls the balance between aggregated and raw features; no value or sensitivity analysis is given.
  • number of filtering layers = grid searched over {1, 2, 3, 5}
    Depth of the low-pass/high-pass filtering stack in Eq. (7).
  • lambda = grid searched over {0, 1e-3, 1e-1, 1}; best on Amazon reported at 0.001 in Sec. 4.4
    Weight of the dual-frequency contrastive loss in Eq. (19).
  • mu = grid searched over {0, 1e-3, 1e-1, 1}; best on Amazon reported at 1
    Weight of the cross-modality alignment loss in Eq. (19).
  • w_i = not reported
    Per-modality weights in Eq. (2) that define the consensus feature; no default or tuning procedure is described.
  • tau = not reported
    Temperature in the InfoNCE contrastive loss (Eq. 15); no value or sensitivity analysis is given.
assumptions (5)
  • domain assumption Real-world multimodal graphs exhibit hybrid homophilic and heterophilic neighborhoods.
    Motivates the entire disentanglement approach; evidenced by Fig. 1 on four datasets but assumed as general.
  • domain assumption A kNN graph on the weighted sum of modality-specific aggregated features identifies class-consistent neighbors.
    Eq. (2)-(3) build the homophilic graph from S = X_tilde X_tilde^T; this assumes the consensus feature is a reliable indicator of same-class similarity.
  • domain assumption The complementary graph of intra-modality similarity captures useful inter-class distinctions.
    Heterophily graphs are built from Top(1 - S_tilde_i, k_h); this assumes the most dissimilar nodes are informative rather than just noisy.
  • domain assumption Low-pass filtering on homophilic graphs and high-pass filtering on heterophilic graphs yield complementary clustering signals.
    Core of the dual-frequency fusion; supported by t-SNE visualizations in Appendix A but not analytically justified.
  • standard math The self-training target (Eq. 17) improves clustering without labels.
    Adopted from DEC-style clustering; the sharpened target distribution is a standard heuristic, not a proven convergence guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Disentangling Homophily and Heterophily in Multimodal Graph Clustering." pith.science (2026). https://pith.science/paper/WQA6OXFC

@misc{pith2026250715253,
  author       = {Pith},
  title        = {Pith review of: Disentangling Homophily and Heterophily in Multimodal Graph Clustering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WQA6OXFC}},
  note         = {Machine review of arXiv:2507.15253}
}
read the original abstract

Multimodal graphs, which integrate unstructured heterogeneous data with structured interconnections, offer substantial real-world utility but remain insufficiently explored in unsupervised learning. In this work, we initiate the study of multimodal graph clustering, aiming to bridge this critical gap. Through empirical analysis, we observe that real-world multimodal graphs often exhibit hybrid neighborhood patterns, combining both homophilic and heterophilic relationships. To address this challenge, we propose a novel framework -- \textsc{Disentangled Multimodal Graph Clustering (DMGC)} -- which decomposes the original hybrid graph into two complementary views: (1) a homophily-enhanced graph that captures cross-modal class consistency, and (2) heterophily-aware graphs that preserve modality-specific inter-class distinctions. We introduce a \emph{Multimodal Dual-frequency Fusion} mechanism that jointly filters these disentangled graphs through a dual-pass strategy, enabling effective multimodal integration while mitigating category confusion. Our self-supervised alignment objectives further guide the learning process without requiring labels. Extensive experiments on both multimodal and multi-relational graph datasets demonstrate that DMGC achieves state-of-the-art performance, highlighting its effectiveness and generalizability across diverse settings. Our code is available at https://github.com/Uncnbb/DMGC.

Figures

Figures reproduced from arXiv: 2507.15253 by the authors.

Figure 1
Figure 1. Node-level homophily ratio distributions across different datasets reveal a prevalent [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of our proposed DMGC. First, it disentangles a cross-modality consensus homophilic graph [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of the learned node representation on [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Homophily ratio comparison of different datasets. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 7
Figure 7. Figure 7: Sensitivity of 𝑘ℎ in Heterophily Graph Construction Hyperparameter Analysis in 𝐾-Nearest Neighbor Graph Construction. Figures 6 and 7 illustrate the effect of hyperpa￾rameters 𝑘𝑙 and 𝑘ℎ for ACC in 𝑘-nearest neighbor (𝑘NN) graph construction, where 𝑘𝑙 and 𝑘ℎ control the…
Figure 5
Figure 5. Figure 5: Accuracy heatmap of Amazon between loss balanc [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 9
Figure 9. Figure 9: Class-wise connection matrices visualized from the [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 47 canonical work pages

  1. [1]

    Graph heterogeneous multi-relational recommendation

    Chong Chen, Weizhi Ma, Min Zhang, Zhaowei Wang, Xiuqiang He, Chenyang Wang, Yiqun Liu, and Shaoping Ma. Graph heterogeneous multi-relational recommendation. In Proceedings of the AAAI conference on artificial intelligence , volume 35, pages 3958–3966, 2021

  2. [2]

    Polygcl: Graph contrastive learning via learnable spectral polynomial filters

    Jingyu Chen, Runlin Lei, and Zhewei Wei. Polygcl: Graph contrastive learning via learnable spectral polynomial filters. In The Twelfth International Conference on Learning Representations, 2023

  3. [3]

    Hybrid transformer with multi-level fusion for multimodal knowledge graph completion

    Xiang Chen, Ningyu Zhang, Lei Li, Shumin Deng, Chuanqi Tan, Changliang Xu, Fei Huang, Luo Si, and Huajun Chen. Hybrid transformer with multi-level fusion for multimodal knowledge graph completion. In Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval, pages 904–915, 2022

  4. [4]

    Adaptive universal generalized pagerank graph neural network

    Eli Chien, Jianhao Peng, Pan Li, and Olgica Milenkovic. Adaptive universal generalized pagerank graph neural network. arXiv preprint arXiv:2006.07988 , 2020

  5. [5]

    Multimodal learning with graphs

    Yasha Ektefaie, George Dasoulas, Ayush Noori, Maha Farhat, and Marinka Zitnik. Multimodal learning with graphs. Nature Machine Intelligence , 5(4):340–350, 2023

  6. [6]

    One2multi graph autoencoder for multi-view graph clustering

    Shaohua Fan, Xiao Wang, Chuan Shi, Emiao Lu, Ken Lin, and Bai Wang. One2multi graph autoencoder for multi-view graph clustering. In proceedings of the web conference 2020 , pages 3070–3076, 2020

  7. [7]

    Slaps: Self-supervision improves structure learning for graph neural networks, 2021

    Bahare Fatemi, Layla El Asri, and Seyed Mehran Kazemi. Slaps: Self-supervision improves structure learning for graph neural networks, 2021

  8. [8]

    Magnn: Metapath aggre- gated graph neural network for heterogeneous graph embedding

    Xinyu Fu, Jiani Zhang, Ziqiao Meng, and Irwin King. Magnn: Metapath aggre- gated graph neural network for heterogeneous graph embedding. In Proceedings of The Web Conference 2020, pages 2331–2341, 2020

Show all 52 references
  1. [9]

    Predict then propagate: Graph neural networks meet personalized pagerank

    Johannes Gasteiger, Aleksandar Bojchevski, and Stephan Günnemann. Predict then propagate: Graph neural networks meet personalized pagerank. In Interna- tional Conference on Learning Representations , 2018

  2. [10]

    Imagebind: One embedding space to bind them all

    Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15180–15190, 2023

  3. [11]

    Gromov-wasserstein multi-modal alignment and clustering

    Fengjiao Gong, Yuzhou Nie, and Hongteng Xu. Gromov-wasserstein multi-modal alignment and clustering. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management , CIKM ’22, page 603–613, New York, NY, USA, 2022. Association for Computing Machinery

  4. [12]

    Mixed-modality clustering via generative graph structure matching

    Xiaxia He, Boyue Wang, Junbin Gao, Qianqian Wang, Yongli Hu, and Baocai Yin. Mixed-modality clustering via generative graph structure matching. IEEE Trans. Knowl. Data Eng., 36(12):8773–8786, December 2024

  5. [13]

    Unigraph2: Learning a unified embedding space to bind multimodal graphs

    Yufei He, Yuan Sui, Xiaoxin He, Yue Liu, Yifei Sun, and Bryan Hooi. Unigraph2: Learning a unified embedding space to bind multimodal graphs. In THE WEB CONFERENCE, 2025

  6. [14]

    Unigraph2: Learning a unified embedding space to bind multimodal graphs

    Yufei He, Yuan Sui, Xiaoxin He, Yue Liu, Yifei Sun, and Bryan Hooi. Unigraph2: Learning a unified embedding space to bind multimodal graphs. In THE WEB CONFERENCE 2025, 2025

  7. [15]

    Multimodal heterogeneous graph attention network

    Xiangen Jia, Min Jiang, Yihong Dong, Feng Zhu, Haocai Lin, Yu Xin, and Huahui Chen. Multimodal heterogeneous graph attention network. Neural Computing and Applications, 35(4):3357–3372, 2023

  8. [16]

    Learning mul- timodal graph-to-graph translation for molecule optimization

    Wengong Jin, Kevin Yang, Regina Barzilay, and Tommi Jaakkola. Learning mul- timodal graph-to-graph translation for molecule optimization. In International Conference on Learning Representations , 2019

  9. [17]

    Hdmi: High-order deep multiplex infomax

    Baoyu Jing, Chanyoung Park, and Hanghang Tong. Hdmi: High-order deep multiplex infomax. In Proceedings of the Web Conference 2021 , pages 2414–2424, 2021

  10. [18]

    Cdc: A simple framework for complex data clustering

    Zhao Kang, Xuanting Xie, Bingheng Li, and Erlin Pan. Cdc: A simple framework for complex data clustering. IEEE Transactions on Neural Networks and Learning Systems, pages 1–12, 2024

  11. [19]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Variational graph auto-encoders, 2016

  12. [20]

    A multimodal graph neural network framework for cancer molecular subtype classification

    Bingjun Li and Sheida Nabavi. A multimodal graph neural network framework for cancer molecular subtype classification. BMC bioinformatics, 25(1):27, 2024

  13. [21]

    Imf: Interactive multimodal fusion model for link prediction

    Xinhang Li, Xiangyu Zhao, Jiaxing Xu, Yong Zhang, and Chunxiao Xing. Imf: Interactive multimodal fusion model for link prediction. In Proceedings of the ACM Web Conference 2023, pages 2572–2580, 2023

  14. [22]

    Multi-view attrib- uted graph clustering

    Zhiping Lin, Zhao Kang, Lizong Zhang, and Ling Tian. Multi-view attrib- uted graph clustering. IEEE Transactions on Knowledge & Data Engineering , 35(02):1872–1880, 2023

  15. [23]

    Self-supervised learning: Generative or contrastive

    Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Generative or contrastive. IEEE Transactions on Knowledge and Data Engineering , 35(1):857–876, 2021

  16. [24]

    Relation structure-aware heterogeneous information network embedding

    Yuanfu Lu, Chuan Shi, Linmei Hu, and Zhiyuan Liu. Relation structure-aware heterogeneous information network embedding. In Proceedings of the AAAI conference on artificial intelligence , volume 33, pages 4456–4463, 2019

  17. [25]

    The heterophilic graph learning handbook: Benchmarks, models, theoretical analysis, applications and challenges

    Sitao Luan, Chenqing Hua, Qincheng Lu, Liheng Ma, Lirong Wu, Xinyu Wang, Minkai Xu, Xiao-Wen Chang, Doina Precup, Rex Ying, et al. The heterophilic graph learning handbook: Benchmarks, models, theoretical analysis, applications and challenges. CoRR, 2024

  18. [26]

    Revisiting heterophily for graph neural networks

    Sitao Luan, Chenqing Hua, Qincheng Lu, Jiaqi Zhu, Mingde Zhao, Shuyuan Zhang, Xiao-Wen Chang, and Doina Precup. Revisiting heterophily for graph neural networks. Advances in neural information processing systems, 35:1362–1375, 2022

  19. [27]

    Multiplex graph representation learning via dual correlation reduction

    Yujie Mo, Yuhuan Chen, Yajie Lei, Liang Peng, Xiaoshuang Shi, Changan Yuan, and Xiaofeng Zhu. Multiplex graph representation learning via dual correlation reduction. IEEE Transactions on Knowledge and Data Engineering , 2023

  20. [28]

    Disentangled multiplex graph representation learning

    Yujie Mo, Yajie Lei, Jialie Shen, Xiaoshuang Shi, Heng Tao Shen, and Xiaofeng Zhu. Disentangled multiplex graph representation learning. In International Conference on Machine Learning , pages 24983–25005. PMLR, 2023

  21. [29]

    Multi-view contrastive graph clustering

    Erlin Pan and Zhao Kang. Multi-view contrastive graph clustering. Advances in neural information processing systems , 34:2148–2159, 2021

  22. [30]

    Beyond homophily: Reconstructing structure for graph-agnostic clustering

    Erlin Pan and Zhao Kang. Beyond homophily: Reconstructing structure for graph-agnostic clustering. In International conference on machine learning , pages 26868–26877. PMLR, 2023

  23. [31]

    Unsupervised attributed multiplex network embedding

    Chanyoung Park, Donghyun Kim, Jiawei Han, and Hwanjo Yu. Unsupervised attributed multiplex network embedding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5371–5378, 2020

  24. [32]

    Learning on multimodal graphs: A survey

    Ciyuan Peng, Jiayuan He, and Feng Xia. Learning on multimodal graphs: A survey. arXiv preprint arXiv:2402.05322, 2024

  25. [33]

    Upper bounding barlow twins: A novel filter for multi-relational clustering

    Xiaowei Qian, Bingheng Li, and Zhao Kang. Upper bounding barlow twins: A novel filter for multi-relational clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 14660–14668, 2024

  26. [34]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning , p...

  27. [35]

    Balanced multi-relational graph clustering

    Zhixiang Shen, Haolan He, and Zhao Kang. Balanced multi-relational graph clustering. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4120–4128, 2024

  28. [36]

    When heterophily meets heterogeneous graphs: Latent graphs guided unsupervised representation learning

    Zhixiang Shen and Zhao Kang. When heterophily meets heterogeneous graphs: Latent graphs guided unsupervised representation learning. IEEE Transactions on Neural Networks and Learning Systems , 36(6):10283–10296, 2025

  29. [37]

    Beyond redundancy: Information- aware unsupervised multiplex graph structure learning

    Zhixiang Shen, Shuo Wang, and Zhao Kang. Beyond redundancy: Information- aware unsupervised multiplex graph structure learning. In Advances in Neural Information Processing Systems, volume 37, pages 31629–31658, 2024

  30. [38]

    Mgat: Multimodal graph attention network for recommendation

    Zhulin Tao, Yinwei Wei, Xiang Wang, Xiangnan He, Xianglin Huang, and Tat- Seng Chua. Mgat: Multimodal graph attention network for recommendation. Information Processing & Management , 57(5):102277, 2020

  31. [39]

    Variable kernel density estimation

    George R Terrell and David W Scott. Variable kernel density estimation. The Annals of Statistics, pages 1236–1265, 1992

  32. [40]

    Hamilton, Pietro Liò, Yoshua Bengio, and R Devon Hjelm

    Petar Veličković, William Fedus, William L. Hamilton, Pietro Liò, Yoshua Bengio, and R Devon Hjelm. Deep graph infomax, 2018

  33. [41]

    An information criterion for controlled disentanglement of multimodal data, 2025

    Chenyu Wang, Sharut Gupta, Xinyi Zhang, Sana Tonekaboni, Stefanie Jegelka, Tommi Jaakkola, and Caroline Uhler. An information criterion for controlled disentanglement of multimodal data, 2025

  34. [42]

    Cooperation of experts: Fusing heterogeneous information with large margin

    Shuo Wang, Shunyang Huang, Jinghui Yuan, Zhixiang Shen, et al. Cooperation of experts: Fusing heterogeneous information with large margin. In Forty-second International Conference on Machine Learning

  35. [43]

    Tiva-kg: A multimodal knowledge graph with text, image, video and audio

    Xin Wang, Benyuan Meng, Hong Chen, Yuan Meng, Ke Lv, and Wenwu Zhu. Tiva-kg: A multimodal knowledge graph with text, image, video and audio. In Proceedings of the 31st ACM international conference on multimedia , pages 2391– 2399, 2023

  36. [44]

    Unsupervised feature learning via non-parametric instance discrimination

    Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 3733–3742, 2018

  37. [45]

    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 S Yu Philip. A comprehensive survey on graph neural networks.IEEE Transactions on Neural Networks and Learning Systems , 32(1):4–24, 2020

  38. [46]

    Graph embedding contrastive multi-modal representation learning for clustering

    Wei Xia, Tianxiu Wang, Quanxue Gao, Ming Yang, and Xinbo Gao. Graph embedding contrastive multi-modal representation learning for clustering. IEEE Transactions on Image Processing , 32:1170–1183, 2023

  39. [47]

    One node one model: Featuring the missing-half for graph clustering

    Xuanting Xie, Bingheng Li, Erlin Pan, Zhaochen Guo, Zhao Kang, and Wenyu Chen. One node one model: Featuring the missing-half for graph clustering. Proceedings of the AAAI Conference on Artificial Intelligence , 2025

  40. [48]

    When graph meets multi- modal: Benchmarking on multimodal attributed graphs learning

    Hao Yan, Chaozhuo Li, Zhigang Yu, Jun Yin, Ruochen Liu, Peiyan Zhang, Weihao Han, Mingzheng Li, Zhengxin Zeng, Hao Sun, et al. When graph meets multi- modal: Benchmarking on multimodal attributed graphs learning. arXiv preprint arXiv:2410.09132, 2024. MM ’25, October 27–31, 20...

  41. [49]

    Gcfagg: Global and cross-view feature aggregation for multi- view clustering

    Weiqing Yan, Yuanyang Zhang, Chenlei Lv, Chang Tang, Guanghui Yue, Liang Liao, and Weisi Lin. Gcfagg: Global and cross-view feature aggregation for multi- view clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19863–19872, 2023

  42. [50]

    Multimodal graph learning for generative tasks

    Minji Yoon, Jing Yu Koh, Bryan Hooi, and Russ Salakhutdinov. Multimodal graph learning for generative tasks. In NeurIPS 2023 Workshop: New Frontiers in Graph Learning, 2023

  43. [51]

    Graph transformer networks

    Seongjun Yun, Minbyul Jeong, Raehyun Kim, Jaewoo Kang, and Hyunwoo J Kim. Graph transformer networks. Advances in neural information processing systems , 32, 2019

  44. [52]

    Multimodal graph benchmark

    Jing Zhu, Yuhang Zhou, Shengyi Qian, Zhongmou He, Tong Zhao, Neil Shah, and Danai Koutra. Multimodal graph benchmark. CoRR, 2024. A ADDITIONAL VISUALIZATION DETAILS A.1 T-SNE visualization comparison between homophilic and heterophiic graph embeddings We conducted an additiona...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.