Pith. sign in

REVIEW 3 major objections 4 minor 57 references

RHEA: Reliability-Harmonized Reconstruction and Assignment for Robust Multimodal-Attributed Graph Clustering

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Neighborhood consensus reveals which node attributes to trust, keeping clustering accurate when text or images are noisy or missing.

desk verdict A credible, genuinely new reliability-field mechanism for multimodal graph clustering, with a real but non-fatal donor-contamination soft spot in the reconstruction step. read the letter →

arxiv 2608.00621 v1 pith:LQH7P6QO submitted 2026-08-01 cs.LG

Yinlin Zhu , Di Wu , Ziyu Han , Zekai Chenm , Wang Luo , Miao Hu , Guocong Quan This is my paper
classification cs.LG
keywords multimodalattributedgraphclusteringnode-specificmodalityreliabilityneighborhoodconsensusmissingcompletionoptimaltransportattributehomophilyunsupervisedrobustness
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

RHEA's central claim is that graph neighborhoods supply free, label-free evidence for how reliable each node's text and image attributes are: under attribute homophily, a modality that agrees with neighboring nodes is trustworthy, and one that deviates is corrupted or missing. Existing multimodal graph clustering fails, the paper argues, because it treats modality reliability as a global property when it is actually node-specific. RHEA estimates a per-node, per-modality reliability field from neighbor agreement, reconstructs unreliable or missing modalities by averaging observed-neighbor embeddings, weights modalities during fusion by that reliability, and steers optimal-transport cluster assignment with reconstruction confidence and neighbor-consensus distillation. Across four benchmarks and five attribute conditions, RHEA consistently beats the strongest baseline, with the NMI advantage widening as attribute quality worsens. The reader should care because real product, social, and citation graphs routinely contain corrupted or missing attributes, and the method offers a way to keep clustering accurate without any labels.

What carries the argument

The central object is the neighborhood-consensus reliability field rho_i,m: for each node i and modality m, the average cosine similarity between the node's modality embedding and its neighbors' embeddings, converted by a softmax across modalities into a per-node distribution over modalities. This one-dimensional signal carries the argument because it gates reconstruction (Eq. 3), weights the fused representation (Eq. 5), and, through the reconstruction confidence c_i, shapes the row marginal of an entropy-regularized optimal transport assignment (Eqs. 6–7) alongside a neighbor-consensus assignment distillation target (Eq. 9).

What would settle it

Build a graph with low attribute homophily, such as edge homophily below 0.5, inject the same Gaussian corruption used in the paper into a random subset of nodes, and measure the AUROC of the reliability-based corruption detector; if it falls to chance instead of the reported 0.95+, the neighborhood-consensus premise is broken. Alternatively, make corruption spatially correlated so that every node's neighbors are also corrupt, and check whether the reconstruction still improves clustering.

Watch

Extended reading notes

Core claim

The paper's discovery is that node-specific modality reliability can be read off the graph itself and then propagated end-to-end through representation learning, fusion, and clustering. For each modality, the average cosine agreement between a node's embedding and its graph neighbors yields a per-node reliability score; after a softmax across modalities, this score decides which modality to repair, how to weight the fused representation, and how much transport mass each node's assignment receives. The claim is validated by synthetic-corruption recoverability: a detector built from the reliability field separates injected corruption from clean nodes with AUROC above 0.95 across all datasets,

Load-bearing premise

The whole pipeline assumes that a node's neighbors are themselves mostly clean and share its semantics; if the neighborhood that is supposed to vote on reliability is full of corrupted or missing attributes, the reliability estimate and the reconstruction both get contaminated.

Editorial extensions

If this is right

  • If RHEA is right, multimodal graph clustering can be made robust to missing or corrupted attributes without labels, relying on local structure instead of a dedicated generative completion model.
  • The reliability field is not just a heuristic: it recovers injected corruption at AUROC above 0.95, so it can double as a label-free corruption detector on the same graphs.
  • The framework's gains grow precisely when attribute quality worsens, meaning performance degrades gracefully instead of collapsing as perturbation rates increase.
  • Each component carries a distinct regime of value: gated reconstruction dominates the missing-modality gains, the reliability gate prevents overwriting reliable modalities under corruption, and neighbor-consensus assignment distillation supplies the clean-setting base.
  • The reliability mechanism generalizes across asymmetric and joint corruption settings, not only single-modality damage.

Reading between the lines

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

  • Because reliability estimation depends on homophily, a natural extension is to make reconstruction donors themselves reliability-filtered, weighting each neighbor by its own reliability or confidence before averaging; this could extend RHEA to graphs with spatially correlated corruption.
  • The reliability field might transfer beyond clustering, for example as a supervision-free data-quality audit for retrieval or recommendation systems built on the same node features.
  • The confidence-weighted optimal transport marginal suggests a general principle: in clustering with unreliable observations, down-weighting low-confidence samples at the transport level is equivalent to robust assignment, a mechanism that could be tested in other OT-based clustering methods.
  • If the AUROC result is stable, the reliability field could serve as a direct estimator of per-node data quality, enabling selective re-encoding or human review of flagged nodes.
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 / 4 minor

Summary. The paper proposes RHEA, an unsupervised framework for clustering multimodal-attributed graphs under node-specific modality noise and missingness. RHEA estimates a per-node, per-modality reliability field from neighbor agreement (Eqs. 1–2), reconstructs unreliable or missing modalities by aggregating observed neighbor embeddings (Eqs. 3–4), fuses modalities according to the estimated reliability (Eq. 5), and feeds the result into a reliability-aware entropic optimal transport clustering objective with neighbor-consensus assignment distillation (Eqs. 6–9). The empirical evaluation is broad: four MAG datasets, five attribute conditions, 20 perturbation settings in the main table, plus ablations, sensitivity sweeps, and a synthetic-corruption recovery analysis. The paper reports that RHEA consistently outperforms the strongest baseline, with larger margins as attribute quality deteriorates, and that the learned reliability field detects injected corruption with AUROC above 0.95.

Significance. If the mechanism worked as described, this would be a meaningful contribution: a fully unsupervised, node-specific reliability signal derived from graph neighborhoods, propagated through reconstruction, fusion, and optimal-transport assignment, with strong and internally consistent benchmark evidence. The empirical work is a genuine strength: RHEA wins in all 20 main perturbation settings, the ablation in Table 3 is informative, and the synthetic-corruption AUROC analysis is a useful sanity check. The concern is that the load-bearing mechanism for the corruption regime is not fully supported as written: Eq. (4) admits corrupted observed neighbors as donors, and the reconstruction-confidence row marginal in Eq. (7) does not change a node's own argmax assignment. These issues are fixable but require either a modification of the method, a corrected description, or dedicated experiments that directly test donor-contamination robustness. The central empirical claim may still hold, but the paper currently overstates the mechanism that is claimed to deliver it.

major comments (3)
  1. [§4.2, Eq. (4)] The donor set N+(i) = {j in N(i) : o_j = 1} includes corrupted neighbors. Under the paper's own corruption protocol (Appendix F.1), corruption replaces a feature vector with a random Gaussian vector but does not set o_j = 0; corrupted nodes remain observed. Therefore the sentence immediately after Eq. (4), claiming that restricting message passing to observed neighbors 'prevents missing or corrupted representations from propagating through the graph,' is false for the corruption regime. This is load-bearing because reconstruction is the main repair mechanism under corruption (Table 3). Please either restrict donors by estimated reliability (e.g., require rho_j,img above a threshold) or, if the current design is retained, add experiments that vary the fraction of corrupted donors and report reconstruction error or downstream NMI stratified by the number of corrupted neighbors.
  2. [§4.3, Eq. (7) and row marginal r_i ∝ c_i] Scaling the row marginal by reconstruction confidence does not, by itself, change the hard assignment of the reconstructed node. In the entropic OT solution P* = diag(u) K diag(v), a uniform scaling of row i's marginal is absorbed by u_i and multiplies every entry P*_ik by the same factor; hence argmax_k P*_ik is invariant. The confidence thus only reduces the node's influence on the prototypes, not the contamination in its own assignment. This undercuts the claim that confidence allows 'uncertain reconstructions to contribute proportionally' to the clustering objective. Please either integrate confidence into the reconstruction/fusion step itself (e.g., confidence-weighted interpolation between the corrupted embedding and the neighbor average), or provide an analysis showing how the row marginal changes the prototype updates and thereby the final assignments.
  3. [§5.6 and Fig. 3] The synthetic-corruption AUROC analysis demonstrates that the reliability field detects injected corruption, but it does not test whether the reconstruction step is robust when donors are themselves corrupted. Fig. 3b stratifies detection AUROC by clean vs. corrupt neighbors, not reconstruction error or downstream clustering quality. Since the corruption protocol keeps corrupted nodes in N+(i), the central robustness claim requires evidence that reconstruction still helps nodes whose neighborhoods contain many corrupted donors. Please add an experiment conditioning on the fraction of corrupted neighbors among the donors, reporting NMI or reconstruction error for that stratum, and compare RHEA's reconstruction against no reconstruction and against a reliability-gated donor set.
minor comments (4)
  1. [§4.3, Eq. (7)] The column marginal c is said to be 'estimated from the graph topology to reflect the empirical cluster-size prior,' but no formula is given. Please specify how this marginal is computed; otherwise the transport polytope is not fully defined.
  2. [§4.2/§4.3] Notation conflict: c_i is used for reconstruction confidence in Eq. (4) and c for the column marginal in Eq. (7). Rename one of them to avoid ambiguity.
  3. [Appendix C] The text claims that 'performance gains decrease smoothly as homophily decreases,' but no controlled experiment or data table is provided; the claim is inferred from four dataset-level points. A homophily-perturbation experiment would make this statement concrete.
  4. [§5.6] The statement 'rising with the corruption severity η (Fig. 3a)' is consistent with the figure, but the y-axis scale (0.95–0.99) makes the trend look small; consider reporting the actual AUROC values in the text or a table.

Circularity Check

1 steps flagged · score 6.0 of 10

Corruption-recovery validation is self-definitional; the core clustering benchmark is independent.

  1. self definitional [Sec. 5.6 (RQ5), Eqs. (1)-(2), and Appendix F.1 corruption model]
    "The detector (1−ρ i,img) separates injected-corrupt from clean nodes at AUROC0.958–0.982 across all four benchmarks, rising with the corruption severityη(Fig. 3a); ... Corruption with ratioηselects⌊ηN⌋nodes uniformly at random and replaces each selected node’s feature vector with a random Gaussian vector rescaled to theℓ2 norm of the original feature, so the content of the modality is destroyed while its scale is preserved."

    Eq. (2) defines ρ_i,m as a softmax over the neighbor-agreement scores of Eq. (1), i.e., reliability is by definition the cosine similarity between a node's embedding and its neighbors' embeddings. The synthetic corruption procedure in F.1 replaces the target node's features with an i.i.d. Gaussian vector, which under homophily removes exactly the neighborhood agreement that Eq. (1) measures. Thus the 'detector' 1−ρ is the same statistic that the corruption is constructed to drive down; recovering injected corruption at AUROC ≈ 0.96 is entailed by the definitions and cannot independently validate the reliability model. The downstream clustering claims are tested on external labels and do not inherit this circularity.

full rationale

The paper's central clustering result is benchmarked against external labeled data and does not depend on self-citations or imported uniqueness theorems. The main circular content is the synthetic-corruption recoverability experiment (Sec. 5.6, Fig. 3): the claimed AUROC validation reduces to the definition of the reliability field. Because Eq. (1)-(2) define reliability as neighborhood agreement, and the corruption protocol destroys neighborhood agreement by replacing features with random Gaussian vectors, detecting such corruption is a tautological consequence of the estimator's definition rather than an independent empirical discovery. I also noted but did not count as circularity a separate correctness issue: Eq. (4) restricts donors to o_j = 1, but corrupted nodes still have o_j = 1, so the claim that this 'prevents corrupted representations from propagating' is not supported; this is a flaw in the method's robustness story, not a circular derivation. Overall, the paper is only partially circular: the corruption-recovery claim is self-definitional, but the main clustering gains on real benchmarks remain independent evidence.

Assumptions & free parameters 13 free parameters · 7 assumptions · 2 invented entities

The central claim rests less on formal derivation than on an empirical assumption set. Free parameters are mostly hyperparameters shared across all datasets; they are not fit to labels, but their values (gamma=0.45, beta=3, lambda weights) are chosen by hand and affect which nodes get reconstructed. The key axioms are homophily and donor cleanliness; the key invented quantities are the reliability field and reconstruction confidence, both internal to the method and only validated in-paper.

free parameters (13)
  • gamma (reliability gate threshold) = 0.45
    Decides which nodes enter the reconstruction set in Eq. (3); sensitivity shown in Table 4.
  • tau (reliability temperature) = 1.0
    Converts neighbor agreement into per-node softmax reliability in Eq. (2).
  • epsilon (entropic OT strength) = 0.05
    Controls smoothness of the transport plan in Eq. (7); a 20x sweep changes NMI by at most 0.020.
  • Sinkhorn iterations = 20
    Number of iterations used to solve Eq. (7); sweep 5 to 20 changes NMI by at most 0.006.
  • tau_c (contrastive temperature) = 0.5
    Temperature in the InfoNCE objective Eq. (10).
  • tau_a (assignment temperature) = 1.0
    Temperature for soft cluster assignment q in Eq. (9).
  • T_s (sharpening temperature) = 0.5
    Sharpens the propagated neighbor-consensus target in Eq. (9).
  • beta (confidence decay) = 3.0
    Controls how strongly donor variance lowers reconstruction confidence c_i in Sec. 4.2.
  • L (consensus propagation depth) = 2
    Number of hops for assignment propagation in Eq. (9).
  • lambda_ncrc = 1.0
    Loss weight for neighbor-consensus assignment distillation in Eq. (12).
  • lambda_nbr = 0.5
    Loss weight for random-walk contrastive loss in Eq. (12).
  • lambda_mod = 0.1
    Loss weight for cross-modal alignment loss in Eq. (12).
  • m0 (repulsion margin) = 0.1
    Margin in inter-node repulsion term Eq. (13).
assumptions (7)
  • domain assumption Attribute homophily: neighboring nodes have similar semantic attributes in each modality.
    Sec. 1 and Sec. 4.1: reliability is measured as agreement with neighbors; if homophily fails, low agreement does not imply corruption.
  • domain assumption Observed neighbors are clean enough to serve as reconstruction donors.
    Eq. (4) restricts donors to o_j=1, not to reliable nodes; corruption among observed neighbors can propagate into the reconstruction.
  • domain assumption Frozen pretrained encoders (RoBERTa, CLIP-ViT-L/14) provide features that preserve semantic similarity under graph homophily.
    Sec. 5.1 and App. F.2: all comparisons rely on these encoders, with no explicit encoder noise model.
  • standard math Entropic optimal transport with Sinkhorn iterations converges to the regularized plan in Eq. (7).
    Standard result from Cuturi 2013, used without proof.
  • domain assumption Random-walk co-occurrence pairs approximate cluster co-membership for the InfoNCE objective.
    Eq. (10) assumes walk co-occurrence is a positive signal for clustering; not proven for these graphs.
  • ad hoc to paper The reconstruction confidence formula c_i = exp(-beta v_i) * w_i/(w_i+1) captures the quality of a reconstructed modality.
    Sec. 4.2: no theoretical justification; the product form and decay coefficient are chosen by hand.
  • ad hoc to paper The graph-topology-based column marginal in Eq. (7) reflects the empirical cluster-size prior.
    Sec. 4.3: the construction is not specified, so the claim that marginals are reliability-aware is under-specified.
invented entities (2)
  • Node-specific modality reliability field rho_i,m
    purpose: Weights each modality per node during fusion and gates reconstruction.
    Internal construct estimated from neighbor agreement; validated only on synthetic corruption within the paper.
  • Reconstruction confidence c_i
    purpose: Downweights unreliable reconstructions through the OT row marginal.
    Computed from donor count and variance; no external benchmark independent of the method.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RHEA: Reliability-Harmonized Reconstruction and Assignment for Robust Multimodal-Attributed Graph Clustering." pith.science (2026). https://pith.science/paper/LQH7P6QO

@misc{pith2026260800621,
  author       = {Pith},
  title        = {Pith review of: RHEA: Reliability-Harmonized Reconstruction and Assignment for Robust Multimodal-Attributed Graph Clustering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LQH7P6QO}},
  note         = {Machine review of arXiv:2608.00621}
}
read the original abstract

Multimodal-attributed graphs (MAGs), whose nodes carry heterogeneous attributes such as text and images over a relational structure, have become a fundamental substrate for label-free entity grouping tasks, including community discovery and product segmentation. Existing MAG clustering methods effectively integrate complementary modalities when attributes are clean and complete, but degrade substantially under noisy or missing attributes because they implicitly assume equal modality reliability across all nodes. In practice, modality reliability is inherently node-specific: images may be corrupted or absent, while textual descriptions are incomplete or noisy. We argue that, under attribute homophily, graph neighborhoods naturally provide supervision-free evidence for estimating node-specific modality reliability. Based on this insight, we propose RHEA, a reliability-aware framework for MAG clustering that estimates node-specific modality reliability from neighborhood consensus and propagates this signal throughout the clustering pipeline. RHEA reconstructs unreliable or missing modalities from graph neighborhoods, adaptively weights modalities during reliability-aware fusion, and performs topology-aware optimal transport clustering with reliability-aware transport assignment and neighbor-consensus assignment distillation. Furthermore, the confidence of reconstructed representations is incorporated into the clustering objective, allowing uncertain reconstructions to contribute proportionally during optimization. Experiments on four MAG benchmarks under five attribute conditions show that RHEA consistently outperforms the strongest baseline, with NMI gains increasing as attribute quality deteriorates.

Figures

Figures reproduced from arXiv: 2608.00621 by the authors.

Figure 1
Figure 1. Overview of RHEA. RHEA consists of four stages: (1) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. NMI vs. perturbation severity on RedditS. Each panel varies one perturbation regime. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Synthetic-corruption recoverability of the reliabil [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: t-SNE of RHEA fused embeddings on RedditS. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Reconstruction operator ablation. NMI under [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Reconstruction confidence vs. corruption (Red [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Stability of the reliability field. Per-step variation [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 46 canonical work pages

  1. [1]

    arXiv preprint arXiv:2511.20030 , year=

    Cross-Contrastive Clustering for Multimodal Attributed Graphs with Dual Graph Filtering , author=. arXiv preprint arXiv:2511.20030 , year=

    arXiv

  2. [2]

    Proceedings of the 33rd ACM International Conference on Multimedia (ACM MM) , year=

    Disentangling Homophily and Heterophily in Multimodal Graph Clustering , author=. Proceedings of the 33rd ACM International Conference on Multimedia (ACM MM) , year=

    work page

  3. [3]

    IEEE Transactions on Neural Networks and Learning Systems , year=

    Dual Information Enhanced Multi-view Attributed Graph Clustering , author=. IEEE Transactions on Neural Networks and Learning Systems , year=

    work page

  4. [4]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Multi-view Contrastive Graph Clustering , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

    work page

  5. [5]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    MoEGCL: Mixture of Ego-Graphs Contrastive Representation Learning for Multi-View Clustering , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

    work page

  6. [6]

    Pattern Analysis and Applications , year=

    Tri-Learn Graph Fusion Network for Attributed Graph Clustering , author=. Pattern Analysis and Applications , year=

    work page

  7. [7]

    IEEE Transactions on Circuits and Systems for Video Technology , year=

    A Clustering-guided Contrastive Fusion for Multi-view Representation Learning , author=. IEEE Transactions on Circuits and Systems for Video Technology , year=

    work page

  8. [8]

    Neurocomputing , year=

    Attributed Graph Clustering with Multi-Scale Weight-Based Pairwise Coarsening and Contrastive Learning , author=. Neurocomputing , year=

    work page

Show all 57 references
  1. [9]

    Proceedings of The Web Conference (WWW) , year=

    Structural Deep Clustering Network , author=. Proceedings of The Web Conference (WWW) , year=

  2. [10]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Deep Fusion Clustering Network , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  3. [11]

    Proceedings of the 29th ACM International Conference on Multimedia (ACM MM) , year=

    Attention-driven Graph Clustering Network , author=. Proceedings of the 29th ACM International Conference on Multimedia (ACM MM) , year=

  4. [12]

    Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI) , year=

    Attributed Graph Clustering: A Deep Attentional Embedding Approach , author=. Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI) , year=

  5. [13]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Deep Graph Clustering via Dual Correlation Reduction , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  6. [14]

    Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) , year=

    Adaptive Graph Encoder for Attributed Graph Embedding , author=. Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) , year=

  7. [15]

    Proceedings of the 40th International Conference on Machine Learning (ICML) , year=

    Dink-Net: Neural Clustering on Large Graphs , author=. Proceedings of the 40th International Conference on Machine Learning (ICML) , year=

  8. [16]

    Proceedings of the ACM Web Conference (WWW) , year=

    Towards Unsupervised Deep Graph Structure Learning , author=. Proceedings of the ACM Web Conference (WWW) , year=

  9. [17]

    Proceedings of the 37th International Conference on Machine Learning (ICML) , year=

    Contrastive Multi-View Representation Learning on Graphs , author=. Proceedings of the 37th International Conference on Machine Learning (ICML) , year=

  10. [18]

    and Jain, Prateek , booktitle=

    Devvrit, Fnu and Sinha, Aditya and Dhillon, Inderjit S. and Jain, Prateek , booktitle=

  11. [19]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Hard Sample Aware Network for Contrastive Deep Graph Clustering , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  12. [20]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Cluster-guided Contrastive Graph Clustering Network , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  13. [21]

    IEEE Transactions on Knowledge and Data Engineering , year=

    Reliable Node Similarity Matrix Guided Contrastive Graph Clustering , author=. IEEE Transactions on Knowledge and Data Engineering , year=

  14. [22]

    Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) , year=

    Revisiting Modularity Maximization for Graph Clustering: A Contrastive Learning Perspective , author=. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) , year=

  15. [23]

    Journal of Machine Learning Research , volume=

    Graph Clustering with Graph Neural Networks , author=. Journal of Machine Learning Research , volume=

  16. [24]

    Proceedings of the Web Conference (WWW) , year=

    Graph Contrastive Learning with Adaptive Augmentation , author=. Proceedings of the Web Conference (WWW) , year=

  17. [25]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Graph Contrastive Learning with Augmentations , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  18. [26]

    Proceedings of the 39th International Conference on Machine Learning (ICML) , year=

    ProGCL: Rethinking Hard Negative Mining in Graph Contrastive Learning , author=. Proceedings of the 39th International Conference on Machine Learning (ICML) , year=

  19. [27]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    COMPLETER: Incomplete Multi-view Clustering via Contrastive Prediction , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  20. [28]

    IEEE Transactions on Image Processing , year=

    Adaptive Feature Projection with Distribution Alignment for Deep Incomplete Multi-view Clustering , author=. IEEE Transactions on Image Processing , year=

  21. [29]

    Proceedings of the 42nd International Conference on Machine Learning (ICML) , year=

    PROTOCOL: Partial Optimal Transport-enhanced Contrastive Learning for Imbalanced Multi-view Clustering , author=. Proceedings of the 42nd International Conference on Machine Learning (ICML) , year=

  22. [30]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Trusted Multi-View Classification with Dynamic Evidential Fusion , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  23. [31]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Differentiable Information Bottleneck for Deterministic Multi-view Clustering , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  24. [32]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Sinkhorn Distances: Lightspeed Computation of Optimal Transport , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  25. [33]

    Findings of the Association for Computational Linguistics: EMNLP , year=

    Multimodal Contrastive Learning via Uni-Modal Coding and Cross-Modal Prediction for Multimodal Sentiment Analysis , author=. Findings of the Association for Computational Linguistics: EMNLP , year=

  26. [34]

    International Conference on Learning Representations (ICLR) , year=

    Semi-Supervised Classification with Graph Convolutional Networks , author=. International Conference on Learning Representations (ICLR) , year=

  27. [35]

    International Conference on Learning Representations (ICLR) , year=

    Graph Attention Networks , author=. International Conference on Learning Representations (ICLR) , year=

  28. [36]

    Proceedings of the National Academy of Sciences , volume=

    Modularity and Community Structure in Networks , author=. Proceedings of the National Academy of Sciences , volume=

  29. [37]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Attention Is All You Need , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  30. [38]

    Naval Research Logistics Quarterly , volume=

    The Hungarian Method for the Assignment Problem , author=. Naval Research Logistics Quarterly , volume=

  31. [39]

    Journal of Machine Learning Research , volume=

    Cluster Ensembles -- A Knowledge Reuse Framework for Combining Multiple Partitions , author=. Journal of Machine Learning Research , volume=

  32. [40]

    Journal of Classification , volume=

    Comparing Partitions , author=. Journal of Classification , volume=

  33. [41]

    Annual Review of Sociology , volume=

    Birds of a Feather: Homophily in Social Networks , author=. Annual Review of Sociology , volume=

  34. [42]

    Proceedings of the 38th International Conference on Machine Learning (ICML) , year=

    Learning Transferable Visual Models From Natural Language Supervision , author=. Proceedings of the 38th International Conference on Machine Learning (ICML) , year=

  35. [43]

    Foundations and Trends in Machine Learning , volume=

    Computational Optimal Transport , author=. Foundations and Trends in Machine Learning , volume=

  36. [44]

    arXiv preprint arXiv:2410.09132 , year =

    When Graph meets Multimodal: Benchmarking and Meditating on Multimodal Attributed Graphs Learning , author =. arXiv preprint arXiv:2410.09132 , year =

    arXiv

  37. [45]

    arXiv preprint arXiv:2602.05576 , year =

    OpenMAG: A Comprehensive Benchmark for Multimodal-Attributed Graph , author =. arXiv preprint arXiv:2602.05576 , year =

  38. [46]

    Li, Xunkai and Wu, Zhengyu and Chen, Zekai and Sun, Henan and Su, Daohan and Zeng, Guang and Qin, Hongchao and Li, Rong-Hua and Wang, Guoren , journal =

  39. [47]

    Neural Networks , volume =

    Incomplete Graph Learning: A Comprehensive Survey , author =. Neural Networks , volume =. 2025 , publisher =

    2025

  40. [48]

    Transactions on Machine Learning Research (TMLR) , year =

    Deep Multimodal Learning with Missing Modality: A Survey , author =. Transactions on Machine Learning Research (TMLR) , year =

  41. [49]

    arXiv preprint arXiv:2602.17354 , year =

    Training-free Graph-based Imputation of Missing Modalities in Multimodal Recommendation , author =. arXiv preprint arXiv:2602.17354 , year =

  42. [50]

    Engineering Applications of Artificial Intelligence , year =

    Training-free Adjustable Polynomial Graph Filtering for Ultra-fast Multimodal Recommendation , author =. Engineering Applications of Artificial Intelligence , year =

  43. [51]

    Proceedings of the 32nd ACM International Conference on Multimedia (MM) , year =

    DiffMM: Multi-Modal Diffusion Model for Recommendation , author =. Proceedings of the 32nd ACM International Conference on Multimedia (MM) , year =

  44. [52]

    Proceedings of the 31st ACM International Conference on Multimedia (MM) , year =

    A Tale of Two Graphs: Freezing and Denoising Graph Structures for Multimodal Recommendation , author =. Proceedings of the 31st ACM International Conference on Multimedia (MM) , year =

  45. [53]

    arXiv preprint arXiv:1901.10610 , year=

    Robust Deep Multi-Modal Sensor Fusion using Fusion Weight Regularization and Target Learning , author=. arXiv preprint arXiv:1901.10610 , year=

    1901 arXiv

  46. [54]

    Complex & Intelligent Systems , year=

    Modality-uncertainty-aware knowledge distillation framework for multimodal sentiment analysis , author=. Complex & Intelligent Systems , year=

  47. [55]

    2025 , doi=

    Chen, Junyu and Shi, Long and Chen, Badong , journal=. 2025 , doi=

    2025

  48. [56]

    Advances in Neural Information Processing Systems, NeurIPS, Datasets and Benchmarks Track, NeurIPS DB Track , year=

    RedCaps: Web-curated image-text data created by the people, for the people , author=. Advances in Neural Information Processing Systems, NeurIPS, Datasets and Benchmarks Track, NeurIPS DB Track , year=

  49. [57]

    McAuley , title =

    Jianmo Ni and Jiacheng Li and Julian J. McAuley , title =

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.